解决bug

README.md

@@ -10,3 +10,36 @@ pinned: false
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+
+## ASR LLM Agent Upgrade
+
+This version adds an LLM-based diagnosis layer on top of alignment/event statistics:
+
+- `analysis/llm_analyzer.py`: sends representative ASR error cases + aggregate stats to an LLM
+- `pipeline/run_analysis.py`: optionally runs LLM diagnosis when `OPENAI_API_KEY` is set
+- `scripts/run_diagnostic.py`: regenerate `llm_diagnosis.json` and `diagnostic_report.md`
+- `report.md`: now includes LLM semantic findings and priority actions
+
+### What the LLM adds
+
+Compared with rule-only classification, the LLM layer can:
+
+- separate surface-form differences from true semantic distortions
+- identify meaning-preserving paraphrases vs business-critical errors
+- infer likely causes from representative cases
+- propose prioritized, actionable improvement suggestions
+
+### Quick start
+
+```bash
+export OPENAI_API_KEY=your_key
+python pipeline/run_all.py   --manifest data/manifest.jsonl   --model_name openai/whisper-small   --llm_model gpt-4.1-mini
+```
+
+Or rerun diagnosis only for an existing run:
+
+```bash
+export OPENAI_API_KEY=your_key
+python scripts/run_diagnostic.py --run_id <run_id> --model gpt-4.1-mini
+```

analysis/llm_analyzer.py

@@ -0,0 +1,317 @@
+from __future__ import annotations
+
+import json
+import os
+from typing import Any, Dict, List, Optional
+
+import pandas as pd
+from openai import OpenAI
+
+
+SEMANTIC_SCHEMA_EXAMPLE = {
+    "utt_id": "string",
+    "semantic_judgement": "语义基本等价|轻微偏差|明显偏差|严重失真",
+    "severity": "high|medium|low",
+    "semantic_error_types": ["string"],
+    "business_impact": "high|medium|low",
+    "reason": "string",
+    "improvement_suggestions": ["string"],
+    "confidence": 0.0,
+}
+
+
+def _safe_float(v: Any) -> Optional[float]:
+    try:
+        if v is None:
+            return None
+        return float(v)
+    except Exception:
+        return None
+
+
+def build_case_pack(df_align: pd.DataFrame, df_events: pd.DataFrame, max_cases: int = 24) -> List[Dict[str, Any]]:
+    if df_align is None or len(df_align) == 0:
+        return []
+
+    align = df_align.copy()
+    if "cer" not in align.columns:
+        align["cer"] = None
+    if "wer" not in align.columns:
+        align["wer"] = None
+
+    sort_cols = [c for c in ["cer", "wer"] if c in align.columns]
+    if sort_cols:
+        align = align.sort_values(sort_cols, ascending=False, na_position="last")
+
+    cases: List[Dict[str, Any]] = []
+    seen = set()
+
+    def _event_summary(utt_id: str) -> Dict[str, Any]:
+        if df_events is None or len(df_events) == 0 or "utt_id" not in df_events.columns:
+            return {"error_classes": {}, "ops": {}, "examples": []}
+        x = df_events[df_events["utt_id"] == utt_id].copy()
+        if len(x) == 0:
+            return {"error_classes": {}, "ops": {}, "examples": []}
+        examples = []
+        for _, row in x.head(10).iterrows():
+            examples.append({
+                "level": row.get("level"),
+                "op_type": row.get("op_type"),
+                "ref": row.get("ref"),
+                "hyp": row.get("hyp"),
+                "error_class": row.get("error_class"),
+            })
+        return {
+            "error_classes": {str(k): int(v) for k, v in x["error_class"].value_counts().head(10).to_dict().items()} if "error_class" in x.columns else {},
+            "ops": {str(k): int(v) for k, v in x["op_type"].value_counts().to_dict().items()} if "op_type" in x.columns else {},
+            "examples": examples,
+        }
+
+    for _, row in align.head(max_cases).iterrows():
+        utt_id = str(row.get("utt_id"))
+        if utt_id in seen:
+            continue
+        seen.add(utt_id)
+        meta = {k: row.get(k) for k in ["device", "domain", "accent", "speaker"] if k in row.index and pd.notna(row.get(k))}
+        cases.append({
+            "utt_id": utt_id,
+            "ref_text": row.get("ref_text"),
+            "hyp_text": row.get("hyp_text"),
+            "norm_ref": row.get("norm_ref"),
+            "norm_hyp": row.get("norm_hyp"),
+            "wer": _safe_float(row.get("wer")),
+            "cer": _safe_float(row.get("cer")),
+            "meta": meta,
+            "event_summary": _event_summary(utt_id),
+        })
+        if len(cases) >= max_cases:
+            break
+
+    return cases
+
+
+def build_global_stats(df_align: pd.DataFrame, df_events: pd.DataFrame, summary: Dict[str, Any]) -> Dict[str, Any]:
+    stats: Dict[str, Any] = {
+        "summary": summary,
+        "num_utterances": int(len(df_align)) if df_align is not None else 0,
+        "num_events": int(len(df_events)) if df_events is not None else 0,
+    }
+    if df_events is not None and len(df_events) > 0:
+        if "op_type" in df_events.columns:
+            stats["op_counts"] = {str(k): int(v) for k, v in df_events["op_type"].value_counts().to_dict().items()}
+        if "error_class" in df_events.columns:
+            stats["error_class_counts"] = {str(k): int(v) for k, v in df_events["error_class"].value_counts().head(20).to_dict().items()}
+
+    slice_stats = {}
+    if df_align is not None and len(df_align) > 0 and "cer" in df_align.columns:
+        for key in ["device", "domain", "accent", "speaker"]:
+            if key in df_align.columns and df_align[key].notna().any():
+                g = df_align.groupby(key)["cer"].mean().dropna().sort_values(ascending=False).head(10)
+                if len(g) > 0:
+                    slice_stats[key] = [{"key": str(k), "cer": float(v)} for k, v in g.items()]
+    if slice_stats:
+        stats["slice_stats"] = slice_stats
+    return stats
+
+
+SYSTEM_PROMPT = """你是资深 ASR 诊断专家，同时具备语音识别、语言学和业务语义分析能力。
+你的任务不是只做 S/I/D 统计，而是识别：
+1. 结构性错误（替换、删除、插入、数字、英文、专名等）
+2. 语义层错误（是否改变原意、是否造成业务理解偏差、是否只是表面字词不同但语义基本等价）
+3. 可能成因（口音、同音混淆、领域词缺失、分段/VAD、噪声、数字口语化、语言模型偏置等）
+4. 可执行的改进建议
+
+必须严格基于输入证据，不要编造音频层信息。若证据不足，明确写“不确定”���
+请输出严格 JSON。"""
+
+
+def _extract_json(text: str) -> Dict[str, Any]:
+    text = text.strip()
+    try:
+        return json.loads(text)
+    except Exception:
+        pass
+    start = text.find("{")
+    end = text.rfind("}")
+    if start >= 0 and end > start:
+        return json.loads(text[start:end + 1])
+    raise ValueError("LLM output is not valid JSON")
+
+
+def analyze_with_llm(
+    df_align: pd.DataFrame,
+    df_events: pd.DataFrame,
+    summary: Dict[str, Any],
+    model: str = "gpt-4.1-mini",
+    client: Optional[OpenAI] = None,
+    max_cases: int = 24,
+) -> Dict[str, Any]:
+    client = client or OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
+
+    case_pack = build_case_pack(df_align, df_events, max_cases=max_cases)
+    global_stats = build_global_stats(df_align, df_events, summary)
+
+    user_prompt = {
+        "task": "请对 ASR 结果做结构+语义联合诊断，并给出改进建议。",
+        "instructions": {
+            "output_schema": {
+                "executive_summary": "string",
+                "major_patterns": [
+                    {
+                        "title": "string",
+                        "priority": 1,
+                        "phenomenon": "string",
+                        "evidence": ["string"],
+                        "semantic_impact": "high|medium|low",
+                        "likely_causes": ["string"],
+                        "recommendations": ["string"],
+                        "confidence": 0.0,
+                    }
+                ],
+                "case_findings": [
+                    {
+                        "utt_id": "string",
+                        "semantic_judgement": "语义基本等价|轻微偏差|明显偏差|严重失真",
+                        "severity": "high|medium|low",
+                        "reason": "string",
+                        "semantic_error_types": ["string"],
+                        "suggestions": ["string"],
+                    }
+                ],
+                "priority_actions": ["string"],
+                "uncertainties": ["string"],
+            },
+            "requirements": [
+                "不要复述所有 case，只保留最有代表性的 findings。",
+                "要区分字符差异和真正改变语义的错误。",
+                "如果 ref/hyp 只是同义改写或口语差异，应该指出语义影响较低。",
+                "建议必须可执行，优先包含数据、解码、后处理、术语表、prompt/context、评测集扩展等方向。",
+            ],
+        },
+        "global_stats": global_stats,
+        "cases": case_pack,
+    }
+
+    resp = client.chat.completions.create(
+        model=model,
+        temperature=0.2,
+        messages=[
+            {"role": "system", "content": SYSTEM_PROMPT},
+            {"role": "user", "content": json.dumps(user_prompt, ensure_ascii=False)},
+        ],
+        response_format={"type": "json_object"},
+    )
+    raw = resp.choices[0].message.content or "{}"
+    parsed = _extract_json(raw)
+    parsed["model"] = model
+    parsed["num_cases_sent"] = len(case_pack)
+    parsed["global_stats"] = global_stats
+    return parsed
+
+
+def _build_semantic_rows(df_align: pd.DataFrame, df_events: pd.DataFrame) -> List[Dict[str, Any]]:
+    rows: List[Dict[str, Any]] = []
+    event_map: Dict[str, List[Dict[str, Any]]] = {}
+    if df_events is not None and len(df_events) > 0 and "utt_id" in df_events.columns:
+        for utt_id, sub in df_events.groupby("utt_id"):
+            examples = []
+            for _, row in sub.head(8).iterrows():
+                examples.append({
+                    "op_type": row.get("op_type"),
+                    "ref": row.get("ref"),
+                    "hyp": row.get("hyp"),
+                    "error_class": row.get("error_class"),
+                    "level": row.get("level"),
+                })
+            event_map[str(utt_id)] = examples
+
+    for _, row in df_align.iterrows():
+        utt_id = str(row.get("utt_id"))
+        meta = {k: row.get(k) for k in ["device", "domain", "accent", "speaker"] if k in row.index and pd.notna(row.get(k))}
+        rows.append({
+            "utt_id": utt_id,
+            "ref_text": row.get("ref_text"),
+            "hyp_text": row.get("hyp_text"),
+            "norm_ref": row.get("norm_ref"),
+            "norm_hyp": row.get("norm_hyp"),
+            "wer": _safe_float(row.get("wer")),
+            "cer": _safe_float(row.get("cer")),
+            "meta": meta,
+            "events": event_map.get(utt_id, []),
+        })
+    return rows
+
+
+def _normalize_semantic_item(item: Dict[str, Any], fallback: Dict[str, Any]) -> Dict[str, Any]:
+    semantic_types = item.get("semantic_error_types") or []
+    suggestions = item.get("improvement_suggestions") or item.get("suggestions") or []
+    return {
+        "utt_id": str(item.get("utt_id") or fallback.get("utt_id")),
+        "semantic_judgement": str(item.get("semantic_judgement") or "不确定"),
+        "severity": str(item.get("severity") or "low"),
+        "semantic_error_types": semantic_types if isinstance(semantic_types, list) else [str(semantic_types)],
+        "business_impact": str(item.get("business_impact") or "low"),
+        "reason": str(item.get("reason") or ""),
+        "improvement_suggestions": suggestions if isinstance(suggestions, list) else [str(suggestions)],
+        "confidence": _safe_float(item.get("confidence")) if item.get("confidence") is not None else None,
+        "ref_text": fallback.get("ref_text"),
+        "hyp_text": fallback.get("hyp_text"),
+        "wer": fallback.get("wer"),
+        "cer": fallback.get("cer"),
+        "device": (fallback.get("meta") or {}).get("device"),
+        "domain": (fallback.get("meta") or {}).get("domain"),
+        "accent": (fallback.get("meta") or {}).get("accent"),
+        "speaker": (fallback.get("meta") or {}).get("speaker"),
+    }
+
+
+def analyze_semantic_per_utterance(
+    df_align: pd.DataFrame,
+    df_events: pd.DataFrame,
+    model: str = "gpt-4.1-mini",
+    client: Optional[OpenAI] = None,
+    batch_size: int = 12,
+) -> pd.DataFrame:
+    client = client or OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
+    rows = _build_semantic_rows(df_align, df_events)
+    if not rows:
+        return pd.DataFrame()
+
+    results: List[Dict[str, Any]] = []
+    for start in range(0, len(rows), batch_size):
+        batch = rows[start:start + batch_size]
+        payload = {
+            "task": "逐条 utterance 做语义级错误判断。",
+            "requirements": [
+                "逐条判断 ref_text 与 hyp_text 的语义偏差程度。",
+                "不要因为表面字不同就判严重错误；如果基本不改变含义，应标注为语义基本等价或轻微偏差。",
+                "结合 events 判断数字、时间、专名、否定、实体、动作关系等关键语义是否出错。",
+                "输出必须覆盖 batch 中每个 utt_id，且仅输出 JSON 对象。",
+            ],
+            "output_schema": {"items": [SEMANTIC_SCHEMA_EXAMPLE]},
+            "items": batch,
+        }
+        resp = client.chat.completions.create(
+            model=model,
+            temperature=0,
+            messages=[
+                {"role": "system", "content": SYSTEM_PROMPT},
+                {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
+            ],
+            response_format={"type": "json_object"},
+        )
+        raw = resp.choices[0].message.content or "{}"
+        parsed = _extract_json(raw)
+        items = parsed.get("items") or []
+        by_id = {str(x.get("utt_id")): x for x in items if isinstance(x, dict)}
+        for fallback in batch:
+            item = by_id.get(str(fallback.get("utt_id")), {"utt_id": fallback.get("utt_id"), "semantic_judgement": "不确定", "reason": "LLM 未返回该条结果。"})
+            results.append(_normalize_semantic_item(item, fallback))
+
+    df = pd.DataFrame(results)
+    if len(df) > 0 and "semantic_error_types" in df.columns:
+        df["semantic_error_types_str"] = df["semantic_error_types"].apply(lambda xs: " | ".join(xs) if isinstance(xs, list) else str(xs))
+    if len(df) > 0 and "improvement_suggestions" in df.columns:
+        df["improvement_suggestions_str"] = df["improvement_suggestions"].apply(lambda xs: " | ".join(xs) if isinstance(xs, list) else str(xs))
+    df["llm_model"] = model
+    return df

data/manifest_hf.jsonl

@@ -0,0 +1,50 @@
+{"utt_id": "fsicoli_common_voice_22_0_00000", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00000.wav", "ref_text": "巴顿是位于美国加利福尼亚州阿马多尔县的一个非建制地区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00001", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00001.wav", "ref_text": "恩骑尉，是中国清朝时的爵名。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00002", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00002.wav", "ref_text": "仙台盐釜港是位于日本宫城县、内的海港，由宫城县政府负责港务营运。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00003", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00003.wav", "ref_text": "利马的阳台是西班牙殖民时期和共和国时期建造的文化遗产。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00004", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00004.wav", "ref_text": "成山，字屏临，号进斋，满洲正蓝旗人。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00005", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00005.wav", "ref_text": "嘉靖十一年任福建龙溪县知县。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00006", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00006.wav", "ref_text": "科莫巴比是位于美国亚利桑那州皮马县的一个人口普查指定地区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00007", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00007.wav", "ref_text": "历史上明永乐皇帝、清干隆皇帝等曾经多次到访，并留下牌匾和诗句。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00008", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00008.wav", "ref_text": "小花仙动画角色列表记录了所有在中国大陆动画《小花仙》系列中出场角色的详细介绍。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00009", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00009.wav", "ref_text": "妳不要再去那里了", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00010", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00010.wav", "ref_text": "银座松竹广场是位于日本东京都中央区筑地一丁目的摩天大楼。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00011", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00011.wav", "ref_text": "儿童权利监察使办公室设于华沙。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00012", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00012.wav", "ref_text": "灰阶音乐是位于香港的一家独立唱片厂牌和音乐出版公司。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00013", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00013.wav", "ref_text": "梁士济，字遂良，广东广州府南海县人，明朝、南明政治人物。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00014", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00014.wav", "ref_text": "姜涛曾就读轩尼诗道官立下午小学、邓肇坚维多利亚官立中学和青年学院。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00015", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00015.wav", "ref_text": "上海江南长兴重工有限责任公司简称长兴重工，厂区位于上海长兴岛船舶制造基地。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00016", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00016.wav", "ref_text": "卢启贤是香港的亿万富翁企业家和慈善家。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00017", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00017.wav", "ref_text": "在这类故事的早期版本里，女人的猪脸外观是由巫术导致的。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00018", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00018.wav", "ref_text": "事件起因据信是天然气爆炸。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00019", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00019.wav", "ref_text": "在工作了九年后，伯爵不幸去世。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00020", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00020.wav", "ref_text": "整个系统称为键接合。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00021", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00021.wav", "ref_text": "大和和纪，日本漫画家，代表作有《窈窕淑女》、《源氏物语》等。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00022", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00022.wav", "ref_text": "事后三天，赵宇被福州市公安局晋安分局以涉嫌故意伤害罪刑事拘留。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00023", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00023.wav", "ref_text": "由春岗互通向萝岗方向排列", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00024", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00024.wav", "ref_text": "弘光帝即位，让刘文照袭封新乐伯，南京沦陷后寄居在高邮，开辟农田种菜直到去世。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "bcb4464171113dd9b51f371c3eecea06771fde83e7e3239ad0516469c6dcdf80170d26c7d1b1ef2476c45b51bfb4ee5549f07d7002bcfcec9b371a30c873b92d", "gender": "male_masculine", "accent": "", "age": "twenties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00025", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00025.wav", "ref_text": "露露夫人终究与三姐弟达成了协议。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00026", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00026.wav", "ref_text": "武定州，中国唐朝时设置的州。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00027", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00027.wav", "ref_text": "宝陀寺，可以指", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00028", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00028.wav", "ref_text": "去札幌啤酒博物馆", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00029", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00029.wav", "ref_text": "洛莱塔是位于美国加利福尼亚州洪堡县的一个人口普查指定地区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00030", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00030.wav", "ref_text": "许州人。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00031", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00031.wav", "ref_text": "班纳镇区为美国堪萨斯州杰克逊县辖下的镇区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00032", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00032.wav", "ref_text": "范家庄遗址，位于山东省潍坊市坊子区坊城街道。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00033", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00033.wav", "ref_text": "郭新立，河北安国人，出生于北京，中国教育人物，现任山东大学党委书记。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "99a4cee094a7058f27e615982d793da9039f8916c4cb0934eafecb601214cb89657ddee22f688a38782a72f5b6622a323ed6dca74f6663430f8cb3c0804563ea", "gender": "male_masculine", "accent": "出生地：31 上海市", "age": "teens", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00034", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00034.wav", "ref_text": "龟山风景区管理处是下辖的一个类似乡级单位。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00035", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00035.wav", "ref_text": "后来他随着李成栋反正，历任光禄卿、户部右侍郎，兵部左侍郎，永历二年晋兵部尚书。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00036", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00036.wav", "ref_text": "同年加入中国人民解放军。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00037", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00037.wav", "ref_text": "由马可、许亚军领衔主演，并由岳红、柯蓝、王策、孙爽联合主演。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00038", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00038.wav", "ref_text": "生于崎玉县川越市，女子美术大学肄业。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00039", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00039.wav", "ref_text": "旧福布斯敦是位于美国加利福尼亚州比尤特县的一个非建制地区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00040", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00040.wav", "ref_text": "大厅供穆斯林祈祷，这也是他们见面以结束禁食的地方。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00041", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00041.wav", "ref_text": "我们就没办法改善", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00042", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00042.wav", "ref_text": "四号镇区是位于美国阿肯色州本顿县的一个镇区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00043", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00043.wav", "ref_text": "格梅林后来出版了若干本关于化学、制药科学、矿物学和植物学的教科书。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "22950d9b987d2554c0d7130808cc60fcb5255d92bb579ad138f4da5e2d5fc52b02d4639e4fe708ef5b820a04812fd3f530e3ea93abfac3e55c8dc2ad22696403", "gender": "", "accent": "", "age": "", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00044", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00044.wav", "ref_text": "同年获选澳门十大杰出运动员，是首位获奖的篮球员。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "3c71635420e0de3a0272e28a63d340dbaaeb5d99e246668955f38c25279dfdbbd8eec8cc8663601fe11d6cfd81a45f9a2e8a5d55379220fe71d24a00bee0effb", "gender": "male_masculine", "accent": "出生地：42 湖北省", "age": "thirties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00045", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00045.wav", "ref_text": "阿尔德斯普林斯是位于美国加利福尼亚州弗雷斯诺县的一个非建制地区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "3c71635420e0de3a0272e28a63d340dbaaeb5d99e246668955f38c25279dfdbbd8eec8cc8663601fe11d6cfd81a45f9a2e8a5d55379220fe71d24a00bee0effb", "gender": "male_masculine", "accent": "出生地：42 湖北省", "age": "thirties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00046", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00046.wav", "ref_text": "巴特勒是位于美国亚利桑那州莫哈维县的一个非建制地区。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "3c71635420e0de3a0272e28a63d340dbaaeb5d99e246668955f38c25279dfdbbd8eec8cc8663601fe11d6cfd81a45f9a2e8a5d55379220fe71d24a00bee0effb", "gender": "male_masculine", "accent": "出生地：42 湖北省", "age": "thirties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00047", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00047.wav", "ref_text": "最后放弃", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "3c71635420e0de3a0272e28a63d340dbaaeb5d99e246668955f38c25279dfdbbd8eec8cc8663601fe11d6cfd81a45f9a2e8a5d55379220fe71d24a00bee0effb", "gender": "male_masculine", "accent": "出生地：42 湖北省", "age": "thirties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00048", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00048.wav", "ref_text": "薄刀峰林场，是下辖的一个类似乡级单位。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "3c71635420e0de3a0272e28a63d340dbaaeb5d99e246668955f38c25279dfdbbd8eec8cc8663601fe11d6cfd81a45f9a2e8a5d55379220fe71d24a00bee0effb", "gender": "male_masculine", "accent": "出生地：42 湖北省", "age": "thirties", "locale": "zh-CN"}}
+{"utt_id": "fsicoli_common_voice_22_0_00049", "audio_uri": "C:\\Users\\hp\\Desktop\\ASR Agent\\ASR_AGENT_\\data\\hf_audio\\fsicoli_common_voice_22_0_00049.wav", "ref_text": "该季他第一次出赛是在九局上担任普林斯·菲尔德的代跑。", "meta": {"dataset_id": "fsicoli/common_voice_22_0", "dataset_config": "zh-CN", "split": "validation", "text_field": "sentence", "sample_rate": 48000, "client_id": "dfacf81ef98f2b80ebf3a932d8c926f7fa65ffaa8dfc35edefc1344d0e4096cc52dd6cd86f2b29d9ae8dc8bf25d4ac3e0fd6133ed370de7f4e6df6d89193c9b4", "gender": "male_masculine", "accent": "出生地：35 福建省", "age": "twenties", "locale": "zh-CN"}}

pipeline/run_all.py

@@ -10,6 +10,8 @@ def main():
     ap.add_argument("--model_name", default="small")
     ap.add_argument("--device", default="cpu")
     ap.add_argument("--compute_type", default="int8")
+    ap.add_argument("--llm_model", default="gpt-4.1-mini")
+    ap.add_argument("--disable_llm", action="store_true")
     args = ap.parse_args()
 
     run_id = run_asr(
@@ -18,7 +20,7 @@ def main():
         device=args.device,
         compute_type=args.compute_type,
     )
-    run_analysis(run_id)
+    run_analysis(run_id, llm_enabled=not args.disable_llm, llm_model=args.llm_model)
     print(f"Done. Run: runs/{run_id}")
 
 

pipeline/run_analysis.py

@@ -1,29 +1,42 @@
 from __future__ import annotations
 import json
+import os
 from pathlib import Path
-from typing import List, Dict
+from typing import List, Dict, Optional
 
 import pandas as pd
+from openai import OpenAI
 
 from core.io import read_jsonl, write_jsonl
 from analysis.align import align_one
 from analysis.events import extract_events
 from analysis.aggregate import aggregate_summary
+from analysis.llm_analyzer import analyze_with_llm, analyze_semantic_per_utterance
 from report.generate import write_report
+from report.diagnostic_report import generate_report_with_openai
 
 
-def run_analysis(run_id: str, out_root: str = "runs") -> None:
+def run_analysis(
+    run_id: str,
+    out_root: str = "runs",
+    llm_enabled: bool = True,
+    llm_model: str = "gpt-4.1-mini",
+    write_diagnostic_report: bool = True,
+) -> None:
     run_dir = Path(out_root) / run_id
     run_meta = json.loads((run_dir / "run_meta.json").read_text(encoding="utf-8"))
 
     asr_path = run_dir / "asr_outputs.jsonl"
     aligned_path = run_dir / "aligned.jsonl"
     events_path = run_dir / "events.parquet"
+    semantic_path = run_dir / "semantic_findings.parquet"
+    semantic_jsonl_path = run_dir / "semantic_findings.jsonl"
 
     aligned_records: List[Dict] = []
     events_records: List[Dict] = []
 
-    for r in read_jsonl(asr_path):
+    asr_rows = list(read_jsonl(asr_path))
+    for r in asr_rows:
         utt_id = r["utt_id"]
         ref = r.get("ref_text")
         hyp = r.get("hyp_text", "")
@@ -44,17 +57,46 @@ def run_analysis(run_id: str, out_root: str = "runs") -> None:
         df_events = pd.DataFrame(events_records)
         df_events.to_parquet(events_path, index=False)
     else:
-        df_events = pd.DataFrame(columns=["op_type", "error_class"])
-
-    # merge meta slice fields into align if they exist
-    # df_align currently doesn't have meta; add a few common keys if present
-    # We'll reconstruct from asr_outputs for slicing.
-    # (Simple approach) reload and join on utt_id:
-    meta_map = {}
-    for r in read_jsonl(asr_path):
-        meta_map[r["utt_id"]] = r.get("meta", {}) or {}
+        df_events = pd.DataFrame(columns=["utt_id", "op_type", "error_class", "ref", "hyp", "level"])
+
+    meta_map = {r["utt_id"]: r.get("meta", {}) or {} for r in asr_rows}
     for key in ["device", "domain", "accent", "speaker"]:
         df_align[key] = df_align["utt_id"].map(lambda u: meta_map.get(u, {}).get(key))
 
     summary = aggregate_summary(df_events, df_align)
-    write_report(run_dir, run_meta, summary)
+
+    llm_diagnosis: Optional[Dict] = None
+    semantic_df = pd.DataFrame()
+    if llm_enabled and os.getenv("OPENAI_API_KEY") and len(df_align) > 0:
+        client = OpenAI()
+        llm_diagnosis = analyze_with_llm(
+            df_align=df_align,
+            df_events=df_events,
+            summary=summary,
+            model=llm_model,
+            client=client,
+        )
+        semantic_df = analyze_semantic_per_utterance(
+            df_align=df_align,
+            df_events=df_events,
+            model=llm_model,
+            client=client,
+        )
+        if len(semantic_df) > 0:
+            semantic_df.to_parquet(semantic_path, index=False)
+            write_jsonl(semantic_jsonl_path, semantic_df.to_dict(orient="records"))
+        if write_diagnostic_report:
+            report = generate_report_with_openai(llm_diagnosis, summary, client, model=llm_model)
+            (run_dir / "diagnostic_report.md").write_text(report, encoding="utf-8")
+    elif write_diagnostic_report and not (run_dir / "diagnostic_report.md").exists():
+        (run_dir / "diagnostic_report.md").write_text(
+            "LLM diagnostic report was skipped because OPENAI_API_KEY is not set.\n"
+            "You can still inspect summary.json and report.md, or rerun with an API key.",
+            encoding="utf-8",
+        )
+
+    write_report(run_dir, run_meta, summary, llm_diagnosis=llm_diagnosis)
+
+
+if __name__ == "__main__":
+    raise SystemExit("Use pipeline/run_all.py or import run_analysis()")

report/diagnostic_report.py

@@ -6,39 +6,40 @@ from typing import Dict, Any
 
 SYSTEM_PROMPT = """You are an ASR diagnostics expert.
 Write a concise but evidence-based ASR error analysis report in Chinese.
-Do not invent evidence. Only use the provided structured statistics.
+Do not invent evidence. Use both structured metrics and LLM semantic diagnosis.
 Focus on:
 1. major error patterns
-2. likely root causes
-3. confidence and uncertainty
-4. actionable next steps
+2. semantic impact of errors
+3. likely root causes
+4. confidence and uncertainty
+5. actionable next steps
 """
 
 
-def build_prompt(root_cause: Dict[str, Any], summary: Dict[str, Any]) -> str:
+def build_prompt(llm_diagnosis: Dict[str, Any], summary: Dict[str, Any]) -> str:
     return f"""
 请基于下面的结构化分析结果，生成一份中文 ASR 错误诊断报告。
 
 要求：
 - 先写总体结论
 - 再写主要错误原因（按优先级排序）
-- 每个原因要包含：现象、证据、可能原因、改进建议
+- 每个原因要包含：现象、证据、语义影响、可能原因、改进建议
 - 最后给出一个优先级排序的行动清单
 - 如果证据不足，要明确说“不确定”
 
 【summary.json】
 {json.dumps(summary, ensure_ascii=False, indent=2)}
 
-【root_cause.json】
-{json.dumps(root_cause, ensure_ascii=False, indent=2)}
+【llm_diagnosis.json】
+{json.dumps(llm_diagnosis, ensure_ascii=False, indent=2)}
 """
 
 
-def generate_report_with_openai(root_cause: Dict[str, Any], summary: Dict[str, Any], client) -> str:
-    prompt = build_prompt(root_cause, summary)
+def generate_report_with_openai(llm_diagnosis: Dict[str, Any], summary: Dict[str, Any], client, model: str = "gpt-4.1-mini") -> str:
+    prompt = build_prompt(llm_diagnosis, summary)
 
     resp = client.chat.completions.create(
-        model="gpt-4.1-mini",
+        model=model,
         messages=[
             {"role": "system", "content": SYSTEM_PROMPT},
             {"role": "user", "content": prompt},

report/generate.py

@@ -1,13 +1,15 @@
 from __future__ import annotations
 import json
 from pathlib import Path
-from typing import Dict
+from typing import Dict, Optional
 from .templates import render_markdown
 
 
-def write_report(run_dir: Path, run_meta: Dict, summary: Dict) -> None:
+def write_report(run_dir: Path, run_meta: Dict, summary: Dict, llm_diagnosis: Optional[Dict] = None) -> None:
     run_dir.mkdir(parents=True, exist_ok=True)
 
     (run_dir / "summary.json").write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")
-    md = render_markdown(run_meta, summary)
+    if llm_diagnosis is not None:
+        (run_dir / "llm_diagnosis.json").write_text(json.dumps(llm_diagnosis, ensure_ascii=False, indent=2), encoding="utf-8")
+    md = render_markdown(run_meta, summary, llm_diagnosis=llm_diagnosis)
     (run_dir / "report.md").write_text(md, encoding="utf-8")

report/templates.py

@@ -1,10 +1,10 @@
 from __future__ import annotations
-from typing import Dict
+from typing import Dict, Optional
 
 
-def render_markdown(run_meta: Dict, summary: Dict) -> str:
+def render_markdown(run_meta: Dict, summary: Dict, llm_diagnosis: Optional[Dict] = None) -> str:
     lines = []
-    lines.append(f"# ASR Error Analysis Report\n")
+    lines.append("# ASR Error Analysis Report\n")
     lines.append(f"**Run ID:** {run_meta.get('run_id')}")
     lines.append(f"**Model:** {run_meta.get('model_info')}\n")
 
@@ -25,7 +25,6 @@ def render_markdown(run_meta: Dict, summary: Dict) -> str:
     for k, v in (summary.get("top_confusions", {}) or {}).items():
         lines.append(f"- {k}: {v}")
 
-    # Slice
     for key in ["device", "domain", "accent", "speaker"]:
         k2 = f"worst_{key}_by_cer"
         if k2 in summary:
@@ -33,9 +32,48 @@ def render_markdown(run_meta: Dict, summary: Dict) -> str:
             for item in summary[k2]:
                 lines.append(f"- {item['key']}: {item['cer']:.4f}")
 
-    lines.append("\n## Recommendations (auto-generated starter)")
-    lines.append("- 优先检查 CER/WER 在特定子集（device/domain/accent）是否显著升高，针对性补数据或做增强。")
-    lines.append("- 如果 top_confusions 集中在数字/时间类，可加入数字规范化与专门的后处理规则。")
-    lines.append("- 如果 mixed_language 占比高，考虑加入英文热词/专名词表或做 LM/解码侧增强。")
+    if llm_diagnosis:
+        lines.append("\n## LLM Executive Summary")
+        if llm_diagnosis.get("executive_summary"):
+            lines.append(llm_diagnosis["executive_summary"])
+
+        patterns = llm_diagnosis.get("major_patterns") or []
+        if patterns:
+            lines.append("\n## LLM Major Patterns")
+            for idx, item in enumerate(patterns, 1):
+                lines.append(f"### {idx}. {item.get('title', 'Untitled Pattern')}")
+                if item.get("phenomenon"):
+                    lines.append(f"- 现象: {item['phenomenon']}")
+                if item.get("semantic_impact"):
+                    lines.append(f"- 语义影响: {item['semantic_impact']}")
+                if item.get("confidence") is not None:
+                    lines.append(f"- 置信度: {item['confidence']}")
+                for e in item.get("evidence", []) or []:
+                    lines.append(f"- 证据: {e}")
+                for c in item.get("likely_causes", []) or []:
+                    lines.append(f"- 可能原因: {c}")
+                for r in item.get("recommendations", []) or []:
+                    lines.append(f"- 建议: {r}")
+
+        findings = llm_diagnosis.get("case_findings") or []
+        if findings:
+            lines.append("\n## Representative Semantic Findings")
+            for item in findings[:10]:
+                lines.append(f"- {item.get('utt_id')}: {item.get('semantic_judgement')}；原因：{item.get('reason')}")
+
+        if llm_diagnosis.get("priority_actions"):
+            lines.append("\n## Priority Actions")
+            for x in llm_diagnosis["priority_actions"]:
+                lines.append(f"- {x}")
+
+        if llm_diagnosis.get("uncertainties"):
+            lines.append("\n## Uncertainties")
+            for x in llm_diagnosis["uncertainties"]:
+                lines.append(f"- {x}")
+    else:
+        lines.append("\n## Recommendations (auto-generated starter)")
+        lines.append("- 优先检查 CER/WER 在特定子集（device/domain/accent）是否显著升高，针对性补数据或做增强。")
+        lines.append("- 如果 top_confusions 集中在数字/时间类，可加入数字规范化与专门的后处理规则。")
+        lines.append("- 如果 mixed_language 占比高，考虑加入英文热词/专名词表或做 LM/解码侧增强。")
 
     return "\n".join(lines)

scripts/run_diagnostic.py

@@ -6,7 +6,7 @@ from pathlib import Path
 import pandas as pd
 from openai import OpenAI
 
-from analysis.root_cause import infer_root_causes
+from analysis.llm_analyzer import analyze_with_llm
 from report.diagnostic_report import generate_report_with_openai
 
 
@@ -20,21 +20,21 @@ def load_jsonl(path: Path):
     return rows
 
 
-def main(run_id: str, runs_dir: str = "runs"):
+def main(run_id: str, runs_dir: str = "runs", model: str = "gpt-4.1-mini"):
     run_dir = Path(runs_dir) / run_id
 
     df_align = pd.DataFrame(load_jsonl(run_dir / "aligned.jsonl"))
     df_events = pd.read_parquet(run_dir / "events.parquet") if (run_dir / "events.parquet").exists() else pd.DataFrame()
     summary = json.loads((run_dir / "summary.json").read_text(encoding="utf-8")) if (run_dir / "summary.json").exists() else {}
 
-    root_cause = infer_root_causes(df_events, df_align)
-    (run_dir / "root_cause.json").write_text(
-        json.dumps(root_cause, ensure_ascii=False, indent=2),
+    client = OpenAI()
+    llm_diagnosis = analyze_with_llm(df_align, df_events, summary, model=model, client=client)
+    (run_dir / "llm_diagnosis.json").write_text(
+        json.dumps(llm_diagnosis, ensure_ascii=False, indent=2),
         encoding="utf-8"
     )
 
-    client = OpenAI()
-    report = generate_report_with_openai(root_cause, summary, client)
+    report = generate_report_with_openai(llm_diagnosis, summary, client, model=model)
     (run_dir / "diagnostic_report.md").write_text(report, encoding="utf-8")
 
     print(f"Diagnostic report written to: {run_dir / 'diagnostic_report.md'}")
@@ -45,5 +45,6 @@ if __name__ == "__main__":
     ap = argparse.ArgumentParser()
     ap.add_argument("--run_id", required=True)
     ap.add_argument("--runs_dir", default="runs")
+    ap.add_argument("--model", default="gpt-4.1-mini")
     args = ap.parse_args()
-    main(args.run_id, args.runs_dir)
+    main(args.run_id, args.runs_dir, args.model)

ui/app.py

@@ -5,11 +5,15 @@ import subprocess
 import sys
 from pathlib import Path
 
-import pandas as pd
 import gradio as gr
+import pandas as pd
 
 RUNS_DIR = Path("runs")
 
+SEMANTIC_JUDGEMENTS = ["ALL", "语义基本等价", "轻微偏差", "明显偏差", "严重失真", "不确定"]
+SEVERITIES = ["ALL", "high", "medium", "low"]
+BUSINESS_IMPACTS = ["ALL", "high", "medium", "low"]
+
 
 def list_runs():
     if not RUNS_DIR.exists():
@@ -20,107 +24,254 @@ def list_runs():
     )
 
 
-def load_run(run_id: str):
-    run_dir = RUNS_DIR / run_id
-
-    meta = json.loads((run_dir / "run_meta.json").read_text(encoding="utf-8"))
+def _read_json(path: Path, default):
+    return json.loads(path.read_text(encoding="utf-8")) if path.exists() else default
+
+
+def _read_jsonl(path: Path):
+    rows = []
+    if not path.exists():
+        return rows
+    with path.open("r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                rows.append(json.loads(line))
+    return rows
+
+
+def _normalize_semantic_cell(xs):
+    def _clean_seq(seq):
+        out = []
+        for x in seq:
+            if x is None:
+                continue
+            try:
+                na = pd.isna(x)
+                if isinstance(na, bool) and na:
+                    continue
+            except Exception:
+                pass
+            s = str(x).strip()
+            if s:
+                out.append(s)
+        return out
+
+    if xs is None:
+        return []
+    if isinstance(xs, (list, tuple, set)):
+        return _clean_seq(xs)
+    if hasattr(xs, "tolist") and not isinstance(xs, (str, bytes, dict)):
+        try:
+            arr = xs.tolist()
+            if isinstance(arr, (list, tuple, set)):
+                return _clean_seq(arr)
+            if arr is None:
+                return []
+            xs = arr
+        except Exception:
+            pass
+    try:
+        na = pd.isna(xs)
+        if isinstance(na, bool) and na:
+            return []
+    except Exception:
+        pass
+    s = str(xs).strip()
+    return [s] if s else []
+
+
+def _normalize_semantic_df(df: pd.DataFrame) -> pd.DataFrame:
+    if df is None or len(df) == 0:
+        return pd.DataFrame()
+    out = df.copy()
+    for col in ["semantic_error_types", "improvement_suggestions"]:
+        if col in out.columns:
+            out[col] = out[col].apply(_normalize_semantic_cell)
+    if "semantic_error_types" in out.columns and "semantic_error_types_str" not in out.columns:
+        out["semantic_error_types_str"] = out["semantic_error_types"].apply(lambda xs: " | ".join(xs))
+    if "improvement_suggestions" in out.columns and "improvement_suggestions_str" not in out.columns:
+        out["improvement_suggestions_str"] = out["improvement_suggestions"].apply(lambda xs: " | ".join(xs))
+    return out
 
-    summary = (
-        json.loads((run_dir / "summary.json").read_text(encoding="utf-8"))
-        if (run_dir / "summary.json").exists()
-        else {}
-    )
 
-    aligned_path = run_dir / "aligned.jsonl"
-    if aligned_path.exists():
-        rows = []
-        with aligned_path.open("r", encoding="utf-8") as f:
-            for line in f:
-                line = line.strip()
-                if line:
-                    rows.append(json.loads(line))
-        df_align = pd.DataFrame(rows)
+def load_run(run_id: str):
+    run_dir = RUNS_DIR / run_id
+    meta = _read_json(run_dir / "run_meta.json", {})
+    summary = _read_json(run_dir / "summary.json", {})
+    df_align = pd.DataFrame(_read_jsonl(run_dir / "aligned.jsonl"))
+    df_events = pd.read_parquet(run_dir / "events.parquet") if (run_dir / "events.parquet").exists() else pd.DataFrame()
+    if (run_dir / "semantic_findings.parquet").exists():
+        df_semantic = _normalize_semantic_df(pd.read_parquet(run_dir / "semantic_findings.parquet"))
     else:
-        df_align = pd.DataFrame()
-
-    events_path = run_dir / "events.parquet"
-    df_events = pd.read_parquet(events_path) if events_path.exists() else pd.DataFrame()
-
-    diagnostic_path = run_dir / "diagnostic_report.md"
-    diagnostic_text = (
-        diagnostic_path.read_text(encoding="utf-8")
-        if diagnostic_path.exists()
-        else "No diagnostic report yet."
-    )
-
-    return meta, summary, df_align, df_events, diagnostic_text
+        df_semantic = pd.DataFrame()
+    llm_diagnosis = _read_json(run_dir / "llm_diagnosis.json", {})
+    diagnostic_text = (run_dir / "diagnostic_report.md").read_text(encoding="utf-8") if (run_dir / "diagnostic_report.md").exists() else "No diagnostic report yet."
+    return meta, summary, df_align, df_events, df_semantic, llm_diagnosis, diagnostic_text
 
 
-def build_summary_md(meta, summary):
+def build_summary_md(meta, summary, df_semantic: pd.DataFrame | None = None):
     lines = []
     lines.append(f"### Run ID: `{meta.get('run_id')}`")
     lines.append(f"- Model: `{meta.get('model_info')}`")
-
     if "wer_mean" in summary and summary["wer_mean"] is not None:
         lines.append(f"- WER(mean): **{summary['wer_mean']:.4f}**")
-
     if "cer_mean" in summary and summary["cer_mean"] is not None:
         lines.append(f"- CER(mean): **{summary['cer_mean']:.4f}**")
-
     lines.append(f"- S/I/D: `{summary.get('sid_counts', {})}`")
-
     if "top_error_classes" in summary:
         lines.append(f"- Top error classes: `{summary.get('top_error_classes', {})}`")
+    if df_semantic is not None and len(df_semantic) > 0:
+        sem_counts = df_semantic["semantic_judgement"].fillna("不确定").value_counts().to_dict() if "semantic_judgement" in df_semantic.columns else {}
+        high_impact = int((df_semantic.get("business_impact", pd.Series(dtype=str)) == "high").sum()) if "business_impact" in df_semantic.columns else 0
+        lines.append(f"- Semantic judgements: `{sem_counts}`")
+        lines.append(f"- High business impact utterances: **{high_impact}**")
+    return "\n".join(lines)
+
 
+def build_semantic_overview_md(df_semantic: pd.DataFrame, llm_diagnosis: dict):
+    if df_semantic is None or len(df_semantic) == 0:
+        return "### Semantic Overview\n暂无 per-utterance LLM 语义诊断结果。请先用配置了 `OPENAI_API_KEY` 的流程跑分析。"
+    lines = ["### Semantic Overview"]
+    if "semantic_judgement" in df_semantic.columns:
+        counts = df_semantic["semantic_judgement"].fillna("不确定").value_counts().to_dict()
+        lines.append(f"- 语义判断分布: `{counts}`")
+    if "business_impact" in df_semantic.columns:
+        impact = df_semantic["business_impact"].fillna("low").value_counts().to_dict()
+        lines.append(f"- 业务影响分布: `{impact}`")
+    if "semantic_error_types" in df_semantic.columns:
+        flat = []
+        for xs in df_semantic["semantic_error_types"].dropna().tolist():
+            flat.extend(xs if isinstance(xs, list) else [str(xs)])
+        if flat:
+            top_types = pd.Series(flat).value_counts().head(8).to_dict()
+            lines.append(f"- 高频语义错误类型: `{top_types}`")
+    if llm_diagnosis.get("priority_actions"):
+        lines.append("- 优先行动:")
+        for action in llm_diagnosis.get("priority_actions", [])[:5]:
+            lines.append(f"  - {action}")
     return "\n".join(lines)
 
 
-def on_select_run(run_id):
-    if not run_id:
-        return "", pd.DataFrame(), pd.DataFrame(), "No diagnostic report yet."
+def _head_align(df_align: pd.DataFrame) -> pd.DataFrame:
+    if len(df_align) == 0:
+        return pd.DataFrame()
+    cols = [c for c in ["utt_id", "wer", "cer", "ref_text", "hyp_text"] if c in df_align.columns]
+    return df_align[cols].head(50)
 
-    meta, summary, df_align, df_events, diagnostic_text = load_run(run_id)
-    md = build_summary_md(meta, summary)
 
-    align_view = (
-        df_align[["utt_id", "wer", "cer"]].head(50)
-        if len(df_align) and all(c in df_align.columns for c in ["utt_id", "wer", "cer"])
-        else pd.DataFrame()
-    )
+def _head_events(df_events: pd.DataFrame) -> pd.DataFrame:
+    if len(df_events) == 0:
+        return pd.DataFrame()
+    cols = [c for c in ["utt_id", "op_type", "ref", "hyp", "error_class", "level"] if c in df_events.columns]
+    return df_events[cols].head(100)
 
-    if len(df_events) and all(
-        c in df_events.columns for c in ["utt_id", "op_type", "ref", "hyp", "error_class", "level"]
-    ):
-        events_view = df_events[["utt_id", "op_type", "ref", "hyp", "error_class", "level"]].head(100)
-    else:
-        events_view = pd.DataFrame()
 
-    return md, align_view, events_view, diagnostic_text
+def _head_semantic(df_semantic: pd.DataFrame) -> pd.DataFrame:
+    if len(df_semantic) == 0:
+        return pd.DataFrame()
+    cols = [
+        "utt_id", "semantic_judgement", "severity", "business_impact", "wer", "cer",
+        "semantic_error_types_str", "reason", "ref_text", "hyp_text",
+    ]
+    cols = [c for c in cols if c in df_semantic.columns]
+    return df_semantic.sort_values([c for c in ["business_impact", "severity", "cer"] if c in df_semantic.columns], ascending=[True, True, False][:len([c for c in ["business_impact", "severity", "cer"] if c in df_semantic.columns])]).head(100)[cols]
+
+
+def on_select_run(run_id):
+    if not run_id:
+        empty = pd.DataFrame()
+        return "", empty, empty, empty, "", "No diagnostic report yet.", gr.update(choices=[]), gr.update(choices=[])
+
+    meta, summary, df_align, df_events, df_semantic, llm_diagnosis, diagnostic_text = load_run(run_id)
+    md = build_summary_md(meta, summary, df_semantic)
+    semantic_md = build_semantic_overview_md(df_semantic, llm_diagnosis)
+
+    type_choices = ["ALL"]
+    if len(df_semantic) > 0 and "semantic_error_types" in df_semantic.columns:
+        types = set()
+        for xs in df_semantic["semantic_error_types"].dropna().tolist():
+            if isinstance(xs, list):
+                types.update(str(x) for x in xs if x)
+            elif xs:
+                types.add(str(xs))
+        type_choices.extend(sorted(types))
+
+    domain_choices = ["ALL"]
+    if len(df_semantic) > 0 and "domain" in df_semantic.columns:
+        domain_choices.extend(sorted(str(x) for x in df_semantic["domain"].dropna().unique()))
+
+    return (
+        md,
+        _head_align(df_align),
+        _head_events(df_events),
+        _head_semantic(df_semantic),
+        semantic_md,
+        diagnostic_text,
+        gr.update(choices=type_choices, value="ALL"),
+        gr.update(choices=domain_choices, value="ALL"),
+    )
 
 
 def search_events(run_id, error_class, contains):
     if not run_id:
         return pd.DataFrame()
-
-    _, _, _, df_events, _ = load_run(run_id)
+    _, _, _, df_events, _, _, _ = load_run(run_id)
     if df_events is None or len(df_events) == 0:
         return pd.DataFrame()
-
     q = df_events.copy()
-
-    if error_class and error_class != "ALL":
+    if error_class and error_class != "ALL" and "error_class" in q.columns:
         q = q[q["error_class"] == error_class]
-
     if contains:
         contains = str(contains)
         q = q[
             q["ref"].astype(str).str.contains(contains, na=False)
             | q["hyp"].astype(str).str.contains(contains, na=False)
         ]
+    cols = [c for c in ["utt_id", "op_type", "ref", "hyp", "error_class", "level"] if c in q.columns]
+    return q[cols].head(200)
+
+
+def search_semantic(run_id, judgement, severity, business_impact, semantic_type, domain, contains, min_cer):
+    if not run_id:
+        return pd.DataFrame()
+    _, _, _, _, df_semantic, _, _ = load_run(run_id)
+    if df_semantic is None or len(df_semantic) == 0:
+        return pd.DataFrame()
 
-    cols = ["utt_id", "op_type", "ref", "hyp", "error_class", "level"]
+    q = df_semantic.copy()
+    if judgement and judgement != "ALL" and "semantic_judgement" in q.columns:
+        q = q[q["semantic_judgement"] == judgement]
+    if severity and severity != "ALL" and "severity" in q.columns:
+        q = q[q["severity"] == severity]
+    if business_impact and business_impact != "ALL" and "business_impact" in q.columns:
+        q = q[q["business_impact"] == business_impact]
+    if semantic_type and semantic_type != "ALL" and "semantic_error_types_str" in q.columns:
+        q = q[q["semantic_error_types_str"].astype(str).str.contains(str(semantic_type), na=False)]
+    if domain and domain != "ALL" and "domain" in q.columns:
+        q = q[q["domain"].astype(str) == str(domain)]
+    if contains:
+        contains = str(contains)
+        q = q[
+            q["ref_text"].astype(str).str.contains(contains, na=False)
+            | q["hyp_text"].astype(str).str.contains(contains, na=False)
+            | q.get("reason", pd.Series(dtype=str)).astype(str).str.contains(contains, na=False)
+            | q.get("semantic_error_types_str", pd.Series(dtype=str)).astype(str).str.contains(contains, na=False)
+        ]
+    if min_cer is not None and "cer" in q.columns:
+        q = q[q["cer"].fillna(0) >= float(min_cer)]
+
+    order_cols = [c for c in ["business_impact", "severity", "cer"] if c in q.columns]
+    if order_cols:
+        q = q.sort_values(order_cols, ascending=[True, True, False][:len(order_cols)])
+
+    cols = [
+        "utt_id", "semantic_judgement", "severity", "business_impact", "wer", "cer",
+        "semantic_error_types_str", "reason", "improvement_suggestions_str", "domain", "accent",
+        "ref_text", "hyp_text",
+    ]
     cols = [c for c in cols if c in q.columns]
-    return q[cols].head(200)
+    return q[cols].head(300)
 
 
 def run_hf_job(dataset_id, dataset_config, split, text_field, model_repo_id, language, num_samples):
@@ -134,72 +285,49 @@ def run_hf_job(dataset_id, dataset_config, split, text_field, model_repo_id, lan
         "--language", language.strip(),
         "--num", str(int(num_samples)),
     ]
-
     if dataset_config and dataset_config.strip():
         cmd += ["--dataset_config", dataset_config.strip()]
 
     p = subprocess.run(cmd, capture_output=True, text=True)
     out = (p.stdout or "") + ("\n" + (p.stderr or "") if p.stderr else "")
-
     if p.returncode != 0:
-        out += (
-            "\n\n[HINT] If you see 401/403 for Common Voice: "
-            "set HF_TOKEN in Space Settings → Secrets, and accept dataset terms on HF."
-        )
-        return (
-            out,
-            gr.update(),
-            "",
-            pd.DataFrame(),
-            pd.DataFrame(),
-            "No diagnostic report yet.",
-        )
+        out += "\n\n[HINT] If you see 401/403 for Common Voice: set HF_TOKEN in Space Settings → Secrets, and accept dataset terms on HF."
+        empty = pd.DataFrame()
+        return out, gr.update(), "", empty, empty, empty, "", "No diagnostic report yet.", gr.update(), gr.update()
 
     runs = list_runs()
     latest = runs[0] if runs else None
-
     if latest:
-        md, align_view, events_view, diagnostic_text = on_select_run(latest)
+        md, align_view, events_view, semantic_view, semantic_md, diagnostic_text, type_dd, domain_dd = on_select_run(latest)
     else:
-        md, align_view, events_view, diagnostic_text = "", pd.DataFrame(), pd.DataFrame(), "No diagnostic report yet."
+        md, align_view, events_view, semantic_view, semantic_md, diagnostic_text, type_dd, domain_dd = "", pd.DataFrame(), pd.DataFrame(), pd.DataFrame(), "", "No diagnostic report yet.", gr.update(), gr.update()
 
-    return (
-        out,
-        gr.update(choices=runs, value=latest),
-        md,
-        align_view,
-        events_view,
-        diagnostic_text,
-    )
+    return out, gr.update(choices=runs, value=latest), md, align_view, events_view, semantic_view, semantic_md, diagnostic_text, type_dd, domain_dd
 
 
 with gr.Blocks() as demo:
-    gr.Markdown("# ASR Error Analysis (Stage 1)")
+    gr.Markdown("# ASR LLM Agent UI")
 
     with gr.Accordion("Run from Hugging Face", open=True):
         gr.Markdown(
             "Fill in a dataset and a Whisper model, then click **Run**. "
-            "If the dataset is gated, set `HF_TOKEN` in Space **Settings → Secrets**."
+            "If the dataset is gated, set `HF_TOKEN` in Space **Settings → Secrets**. "
+            "For LLM semantic diagnostics, make sure `OPENAI_API_KEY` is available."
         )
-
         with gr.Row():
             dataset_id = gr.Textbox(label="HF dataset repo id", value="fsicoli/common_voice_22_0")
             dataset_config = gr.Textbox(label="Dataset config (optional)", value="zh-CN")
-
         with gr.Row():
             split = gr.Textbox(label="Split", value="validation")
             text_field = gr.Textbox(label="Transcript field", value="sentence")
             num_samples = gr.Number(label="Num samples", value=50, precision=0)
-
         with gr.Row():
             model_repo_id = gr.Textbox(label="HF model repo id", value="openai/whisper-small")
             language = gr.Textbox(label="Language", value="zh")
-
         run_btn = gr.Button("Run")
         logs = gr.Textbox(label="Logs", lines=16)
 
     gr.Markdown("## Browse Existing Runs")
-
     runs = list_runs()
     run_dd = gr.Dropdown(choices=runs, label="Select run", value=(runs[0] if runs else None))
     summary_md = gr.Markdown()
@@ -210,15 +338,7 @@ with gr.Blocks() as demo:
 
     with gr.Accordion("Search Error Events", open=False):
         error_cls = gr.Dropdown(
-            choices=[
-                "ALL",
-                "number_or_time",
-                "mixed_language",
-                "substitution",
-                "deletion",
-                "insertion",
-                "other",
-            ],
+            choices=["ALL", "number_or_time", "mixed_language", "substitution", "deletion", "insertion", "other"],
             value="ALL",
             label="error_class",
         )
@@ -226,30 +346,53 @@ with gr.Blocks() as demo:
         search_btn = gr.Button("Search")
         result_tbl = gr.Dataframe(label="Search results", interactive=False)
 
+    gr.Markdown("## Per-Utterance Semantic Diagnostics")
+    semantic_overview_md = gr.Markdown("暂无语义诊断结果。")
+    semantic_tbl = gr.Dataframe(label="Semantic findings (head)", interactive=False)
+
+    with gr.Accordion("Filter Semantic Errors", open=True):
+        with gr.Row():
+            semantic_judgement = gr.Dropdown(choices=SEMANTIC_JUDGEMENTS, value="ALL", label="semantic_judgement")
+            semantic_severity = gr.Dropdown(choices=SEVERITIES, value="ALL", label="severity")
+            semantic_business_impact = gr.Dropdown(choices=BUSINESS_IMPACTS, value="ALL", label="business_impact")
+        with gr.Row():
+            semantic_type = gr.Dropdown(choices=["ALL"], value="ALL", label="semantic_error_type")
+            semantic_domain = gr.Dropdown(choices=["ALL"], value="ALL", label="domain")
+            semantic_min_cer = gr.Number(label="min CER", value=0.0)
+        semantic_contains = gr.Textbox(label="contains (ref/hyp/reason/type)")
+        semantic_search_btn = gr.Button("Filter semantic findings")
+        semantic_result_tbl = gr.Dataframe(label="Filtered semantic findings", interactive=False)
+
     with gr.Accordion("Diagnostic Report", open=True):
         diagnostic_md = gr.Markdown("No diagnostic report yet.")
 
     if runs:
-        md0, a0, e0, d0 = on_select_run(runs[0])
+        md0, a0, e0, s0, so0, d0, type0, domain0 = on_select_run(runs[0])
         summary_md.value = md0
         align_tbl.value = a0
         events_tbl.value = e0
+        semantic_tbl.value = s0
+        semantic_overview_md.value = so0
         diagnostic_md.value = d0
+        semantic_type.choices = type0["choices"]
+        semantic_domain.choices = domain0["choices"]
 
     run_dd.change(
         on_select_run,
         inputs=[run_dd],
-        outputs=[summary_md, align_tbl, events_tbl, diagnostic_md],
+        outputs=[summary_md, align_tbl, events_tbl, semantic_tbl, semantic_overview_md, diagnostic_md, semantic_type, semantic_domain],
     )
 
-    search_btn.click(
-        search_events,
-        inputs=[run_dd, error_cls, contains],
-        outputs=[result_tbl],
+    search_btn.click(search_events, inputs=[run_dd, error_cls, contains], outputs=[result_tbl])
+
+    semantic_search_btn.click(
+        search_semantic,
+        inputs=[run_dd, semantic_judgement, semantic_severity, semantic_business_impact, semantic_type, semantic_domain, semantic_contains, semantic_min_cer],
+        outputs=[semantic_result_tbl],
     )
 
     run_btn.click(
         run_hf_job,
         inputs=[dataset_id, dataset_config, split, text_field, model_repo_id, language, num_samples],
-        outputs=[logs, run_dd, summary_md, align_tbl, events_tbl, diagnostic_md],
+        outputs=[logs, run_dd, summary_md, align_tbl, events_tbl, semantic_tbl, semantic_overview_md, diagnostic_md, semantic_type, semantic_domain],
     )