Use user's OpenVoice V2 elderly voice conversion (Gradio SDK)

Dockerfile

@@ -1,21 +0,0 @@
-FROM python:3.11-slim-bookworm
-
-ENV DEBIAN_FRONTEND=noninteractive
-ENV PIP_NO_CACHE_DIR=1
-ENV PYTHONUNBUFFERED=1
-
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    ffmpeg libsndfile1 libportaudio2 libasound2 \
-    && rm -rf /var/lib/apt/lists/*
-
-WORKDIR /app
-
-# Ultra-minimal deps: only what's needed for pitch/formant shifting
-RUN pip install --no-cache-dir --timeout=120 \
-    numpy scipy librosa soundfile praat-parselmouth gradio fastapi uvicorn pydantic httpx anyio starlette websockets
-
-COPY app_simple.py app.py
-COPY inference_simple.py inference.py
-
-EXPOSE 7860
-CMD ["python", "app.py"]

README.md

@@ -1,104 +1,25 @@
 ---
-title: Blender Headless API
-emoji: 🦖
-colorFrom: blue
-colorTo: purple
-sdk: docker
-app_port: 7860
+title: 老年声音转换
+emoji: 🎙️
+colorFrom: gray
+colorTo: yellow
+sdk: gradio
+sdk_version: 4.44.0
+app_file: app.py
 pinned: false
 ---
 
-# Blender Headless API
+# 老年声音转换
 
-在 Hugging Face Space（CPU basic）上跑一个常驻的 FastAPI 服务，
-每次请求用 `blender --background --python worker.py` 无头执行一组
-白名单 3D 操作（建模 / 材质 / 灯光 / 相机 / 动画 / 渲染 / 导出），
-供 OpenClaw Agent 通过 HTTP 调用。
+上传一段录音 + 一段几秒钟的老年人声音样本,基于 OpenVoice V2 的零样本音色转换,
+把录音的音色替换成老年人音色(保留原始语调、节奏、内容)。
 
-## 目录结构
+首次启动会自动从 Hugging Face Hub 下载 `myshell-ai/OpenVoiceV2` 的转换器权重
+(约几百 MB),之后会缓存在 Space 里。CPU basic 上单次转换约 10~30 秒。
 
-```
-.
-├── Dockerfile          # 安装 Blender + 启动 FastAPI
-├── requirements.txt
-├── app.py              # FastAPI 服务，接收请求，拉起 blender 子进程
-├── worker.py           # 运行在 Blender 内部（bpy）的操作分发器
-└── example_request.json
-```
+## 部署步骤
 
-## API
-
-### `GET /health`
-检查服务和 Blender 二进制是否就绪。
-
-### `POST /execute`
-```json
-{
-  "scene_b64": null,
-  "ops": [
-    {"op": "create_object", "type": "sphere", "name": "Ball", "location": [0,0,1]},
-    {"op": "add_material", "target": "Ball", "color": [0.8, 0.1, 0.1, 1.0], "roughness": 0.3},
-    {"op": "add_light", "type": "SUN", "location": [4,-4,6], "energy": 3},
-    {"op": "set_camera", "location": [6,-6,4], "look_at": [0,0,1]},
-    {"op": "set_render_settings", "resolution": [960, 540], "samples": 32},
-    {"op": "render_image", "filename": "ball.png"},
-    {"op": "export_glb", "filename": "ball.glb"}
-  ]
-}
-```
-
-返回：
-```json
-{
-  "job_id": "a1b2c3d4e5f6",
-  "ops_result": {"ok": true, "results": [...], "errors": []},
-  "files": ["ball.png", "ball.glb"],
-  "download_base": "/files/a1b2c3d4e5f6/"
-}
-```
-
-### `GET /files/{job_id}/{filename}`
-下载生成的文件（渲染图 / glb / fbx / blend）。
-
-### `DELETE /jobs/{job_id}`
-清理某次任务的临时文件。
-
-## 支持的 op（在 `worker.py` 里维护白名单）
-
-| op | 说明 |
-|---|---|
-| `create_object` | 创建基础几何体（sphere/cube/cylinder/cone/plane/monkey/torus/empty） |
-| `move_object` | 绝对位置或相对位移 |
-| `set_transform` | 设置位置/旋转(度)/缩放 |
-| `add_material` | 创建/应用材质（颜色、金属度、粗糙度） |
-| `set_camera` | 创建相机，支持 look_at 自动朝向 |
-| `add_light` | 创建灯光（POINT/SUN/AREA/SPOT） |
-| `add_animation` | 按关键帧设置位置/旋转/缩放 |
-| `set_render_settings` | 分辨率/引擎/采样数/帧率/起止帧 |
-| `render_image` | 渲染单帧 PNG |
-| `render_animation` | 渲染动画序列帧 |
-| `import_model` | 导入 glb/gltf/fbx/obj |
-| `export_glb` / `export_fbx` | 导出模型 |
-| `delete_object` | 删除物体 |
-| `save_blend` | 保存 .blend 文件 |
-
-不支持任意 Python 执行——Agent 只能组合调用这些 op，避免生成失控脚本
-（死循环、超大渲染、越权文件访问）。
-
-## 部署前需要改的地方
-
-1. `Dockerfile` 里的 `BLENDER_VERSION`，默认是 `4.2.9`（4.2 LTS 最新补丁版）。
-   如需其他版本，去 https://download.blender.org/release/ 确认对应文件名。
-2. `JOB_TIMEOUT_SECONDS`（Dockerfile 里的环境变量），默认 180 秒，
-   渲染复杂场景时按需调大——但 HF Space CPU basic 的算力有限，
-   建议先用低分辨率/低采样数验证流程。
-3. 认证：这个 MVP 没有做鉴权，Space 一旦公开任何人都能调用你的 `/execute`。
-   建议在 `app.py` 加一个简单的 API Key header 校验（`X-API-Key`），
-   OpenClaw 那边配置对应的 key。
-
-## OpenClaw 集成建议
-
-按你之前的设计，把它包成一个 skill：
-`blender_execute` = bash wrapper 拼装 ops.json → POST 到这个 Space 的
-`/execute` → 轮询/直接拿到 `files` 列表 → 用 `GET /files/...` 下载渲染结果
-或 glb，交给你现有的 video-gen / 3d-video pipeline 继续处理。
+1. 在 Hugging Face 新建一个 Space:SDK 选 **Gradio**,Hardware 选 **CPU basic**
+2. 把本文件夹里的 `app.py`、`requirements.txt`、`packages.txt`、`README.md`
+   四个文件都传到 Space 仓库根目录(可以直接在网页上传,或 `git push`)
+3. 等 Space 自动 build 完成即可使用

app.py

@@ -1,217 +1,79 @@
-#!/usr/bin/env python3
-"""Gradio Web UI for Elderly Voice Conversion (RVC-based)."""
 import os
 import tempfile
-from pathlib import Path
+
 import gradio as gr
-import torch
-import numpy as np
-import soundfile as sf
-import librosa
+from huggingface_hub import snapshot_download
 
-from inference import convert_voice, vc, MODEL_PATH
+from openvoice import se_extractor
+from openvoice.api import ToneColorConverter
 
-# Check GPU availability
-device = "cuda" if torch.cuda.is_available() else "cpu"
-has_gpu = torch.cuda.is_available()
+DEVICE = "cpu"
 
-# UI Configuration
-TITLE = "🎭 Elderly Voice Converter"
-DESCRIPTION = """
-基于 RVC (Retrieval-based Voice Conversion) 的声音转换工具，将输入音频转换为老年人声音特征。
+# Downloads OpenVoice V2's tone-color converter checkpoint from the HF Hub
+# the first time the Space boots. Cached afterwards in the Space's storage.
+CKPT_DIR = snapshot_download(repo_id="myshell-ai/OpenVoiceV2")
+CONVERTER_DIR = os.path.join(CKPT_DIR, "converter")
 
-**使用方法：**
-1. 上传音频文件（支持 wav, mp3, flac 等格式）
-2. 调整参数（音调偏移建议 -2 到 -5 半音，让声音更沉稳）
-3. 点击转换，等待处理完成
-4. 下载或试听结果
+tone_color_converter = ToneColorConverter(
+    os.path.join(CONVERTER_DIR, "config.json"), device=DEVICE
+)
+tone_color_converter.load_ckpt(os.path.join(CONVERTER_DIR, "checkpoint.pth"))
 
-**参数说明：**
-- **音调偏移**：负值降低音调（更沉稳），正值升高。老年声建议 -2 ~ -5
-- **F0 方法**：rmvpe 效果最好，crepe 需 GPU，harvest/pm 速度较快
-- **Index 率**：控制模型特征强度，0.5-0.8 通常效果较好
-- **保护无声辅音**：避免气声过重，建议 0.33
-"""
 
-# Example audio (we'll create a simple one if needed)
-EXAMPLES_DIR = Path("/app/examples")
-EXAMPLES_DIR.mkdir(exist_ok=True)
+def convert_voice(source_audio, elderly_reference_audio):
+    if source_audio is None or elderly_reference_audio is None:
+        raise gr.Error("请先上传「原始录音」和「老年人参考音频」两个文件")
 
+    with tempfile.TemporaryDirectory() as tmp:
+        # vad=True uses lightweight Silero VAD to find clean speech segments
+        # (no transcription needed), which is enough for tone-color extraction.
+        src_se, _ = se_extractor.get_se(
+            source_audio, tone_color_converter, target_dir=tmp, vad=True
+        )
+        tgt_se, _ = se_extractor.get_se(
+            elderly_reference_audio, tone_color_converter, target_dir=tmp, vad=True
+        )
 
-def process_audio(
-    audio_file,
-    pitch_shift,
-    f0_method,
-    index_rate,
-    filter_radius,
-    rms_mix_rate,
-    protect,
-):
-    """Process uploaded audio and return converted audio."""
-    if audio_file is None:
-        return None, "请先上传音频文件"
-    
-    if vc is None:
-        return None, "❌ 模型未加载，请检查模型文件是否存在"
-    
-    try:
-        # Create temp files
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_in:
-            # Load and save input audio at 16kHz
-            y, sr = librosa.load(audio_file, sr=16000, mono=True)
-            sf.write(tmp_in.name, y, 16000)
-            input_path = tmp_in.name
-        
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_out:
-            output_path = tmp_out.name
-        
-        # Convert
-        convert_voice(
-            input_audio_path=input_path,
-            output_audio_path=output_path,
-            f0_up_key=pitch_shift,
-            f0_method=f0_method,
-            index_rate=index_rate,
-            filter_radius=filter_radius,
-            rms_mix_rate=rms_mix_rate,
-            protect=protect,
+        output_path = os.path.join(tmp, "output.wav")
+        tone_color_converter.convert(
+            audio_src_path=source_audio,
+            src_se=src_se,
+            tgt_se=tgt_se,
+            output_path=output_path,
         )
-        
-        # Clean up input temp file
-        os.unlink(input_path)
-        
-        return output_path, "✅ 转换完成！"
-    
-    except Exception as e:
-        return None, f"❌ 转换失败: {str(e)}"
 
+        with open(output_path, "rb") as f:
+            data = f.read()
 
-def create_interface():
-    with gr.Blocks(title=TITLE, theme=gr.themes.Soft()) as demo:
-        gr.Markdown(f"# {TITLE}")
-        gr.Markdown(DESCRIPTION)
-        
-        with gr.Row():
-            with gr.Column(scale=1):
-                # Input
-                audio_input = gr.Audio(
-                    label="输入音频",
-                    type="filepath",
-                    sources=["upload", "microphone"],
-                )
-                
-                # Parameters
-                with gr.Group():
-                    gr.Markdown("### 转换参数")
-                    pitch_shift = gr.Slider(
-                        minimum=-12,
-                        maximum=12,
-                        value=-3,
-                        step=1,
-                        label="音调偏移 (半音)",
-                        info="负值=更低沉(老年声建议 -2~-5)"
-                    )
-                    f0_method = gr.Dropdown(
-                        choices=["rmvpe", "crepe", "crepe-tiny", "harvest", "pm", "dio"],
-                        value="rmvpe",
-                        label="F0 提取方法",
-                        info="rmvpe 推荐；crepe 需 GPU"
-                    )
-                    index_rate = gr.Slider(
-                        minimum=0.0,
-                        maximum=1.0,
-                        value=0.75,
-                        step=0.05,
-                        label="Index 率",
-                        info="模型特征强度，0.5-0.8 推荐"
-                    )
-                    filter_radius = gr.Slider(
-                        minimum=0,
-                        maximum=7,
-                        value=3,
-                        step=1,
-                        label="滤波半径",
-                        info="音调平滑，0=关闭"
-                    )
-                    rms_mix_rate = gr.Slider(
-                        minimum=0.0,
-                        maximum=1.0,
-                        value=0.25,
-                        step=0.05,
-                        label="RMS 混合率",
-                        info="音量包络混合"
-                    )
-                    protect = gr.Slider(
-                        minimum=0.0,
-                        maximum=0.5,
-                        value=0.33,
-                        step=0.01,
-                        label="保护无声辅音",
-                        info="避免气声过重"
-                    )
-                
-                convert_btn = gr.Button("🎭 转换为老年声", variant="primary", size="lg")
-                status = gr.Textbox(label="状态", interactive=False)
-            
-            with gr.Column(scale=1):
-                # Output
-                audio_output = gr.Audio(
-                    label="转换结果",
-                    type="filepath",
-                )
-                gr.Markdown("""
-                ### 💡 提示
-                - **音频时长**：建议 10-60 秒，过长处理较慢
-                - **采样率**：自动重采样到 16kHz
-                - **音调建议**：
-                  - 男性转老年男声：-3~-5
-                  - 女性转老年女声：-2~-4
-                  - 女性转老年男声：-5~-8 (需配合 index_rate)
-                - **F0 方法选择**：
-                  - `rmvpe`：平衡速度/质量，CPU 可用 ⭐推荐
-                  - `crepe`：质量最好，需 GPU
-                  - `harvest`：传统方法，速度快
-                """)
-        
-        # Model info
-        with gr.Accordion("模型信息", open=False):
-            gr.Markdown(f"""
-            - **设备**: {device.upper()} {'✅' if has_gpu else '⚠️ (CPU 模式较慢)'}
-            - **模型路径**: `{MODEL_PATH or '未找到'}`  
-            - **基础模型**: RVC v2 (768 channels)
-            - **目标音色**: 老年/长者声音特征
-            """)
-        
-        convert_btn.click(
-            fn=process_audio,
-            inputs=[
-                audio_input,
-                pitch_shift,
-                f0_method,
-                index_rate,
-                filter_radius,
-                rms_mix_rate,
-                protect,
-            ],
-            outputs=[audio_output, status],
-            api_name="convert",
-        )
-        
-        # Example section (if we have examples)
-        gr.Examples(
-            examples=[],
-            inputs=audio_input,
-            label="示例音频 (可上传自己的)",
+    # Write result outside the tempdir so Gradio can still read it after cleanup
+    final_path = tempfile.NamedTemporaryFile(suffix=".wav", delete=False).name
+    with open(final_path, "wb") as f:
+        f.write(data)
+    return final_path
+
+
+with gr.Blocks(title="老年声音转换") as demo:
+    gr.Markdown(
+        "# 声音转老年音色\n"
+        "1. 上传要转换的原始录音\n"
+        "2. 上传一段几秒钟的老年人声音样本(参考音色)\n"
+        "3. 点击「开始转换」\n\n"
+        "基于 [OpenVoice V2](https://github.com/myshell-ai/OpenVoice) 的零样本音色转换,"
+        "CPU 上单次转换大约需要 10~30 秒,请耐心等待。"
+    )
+    with gr.Row():
+        source_audio = gr.Audio(label="要转换的原始录音", type="filepath")
+        elderly_reference_audio = gr.Audio(
+            label="老年人参考音频(几秒钟即可)", type="filepath"
         )
-    
-    return demo
+    convert_btn = gr.Button("开始转换", variant="primary")
+    output_audio = gr.Audio(label="转换结果", type="filepath")
 
+    convert_btn.click(
+        fn=convert_voice,
+        inputs=[source_audio, elderly_reference_audio],
+        outputs=output_audio,
+    )
 
 if __name__ == "__main__":
-    demo = create_interface()
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True,
-        share=False,
-    )
+    demo.launch()

app_simple.py

@@ -1,227 +0,0 @@
-#!/usr/bin/env python3
-"""Simple Elderly Voice Converter using librosa for pitch shift + simple formant shift."""
-import os
-import tempfile
-import gradio as gr
-import numpy as np
-import soundfile as sf
-import librosa
-import librosa.effects
-
-TITLE = "🎭 Elderly Voice Converter (Pitch + Formant Shift)"
-DESCRIPTION = """
-基于 **音调偏移 + 共振峰偏移** 的轻量级声音转换，模拟老年人声线特征。
-
-**原理：**
-- **降低基频 (Pitch)**：使用 librosa PSOLA 算法降低音调
-- **降低共振峰**：通过频谱包络偏移模拟声道变长
-- **添加气声**：模拟声带闭合不全的气息感
-
-**使用方法：**
-1. 上传音频文件 (wav, mp3, flac 等)
-2. 调整参数：
-   - **Pitch Shift (半音)**：建议 -3 到 -6 (男声) / -2 到 -5 (女声)
-   - **Formant Shift (半音)**：建议 -2 到 -4
-   - **Breathiness**：0-0.3，添加气声感
-3. 点击转换
-"""
-
-
-def convert_voice_simple(
-    input_audio_path: str,
-    output_audio_path: str,
-    pitch_shift: float = -4.0,      # semitones
-    formant_shift: float = -3.0,    # semitones
-    breathiness: float = 0.15,
-):
-    """
-    Convert voice to elderly-like using pitch + formant shifting.
-    
-    Args:
-        input_audio_path: Input audio file
-        output_audio_path: Output audio file
-        pitch_shift: Pitch shift in semitones (negative = lower)
-        formant_shift: Formant shift in semitones (negative = lower)
-        breathiness: Amount of breathiness to add (0-0.5)
-    """
-    # Load audio
-    y, sr = librosa.load(input_audio_path, sr=None, mono=True)
-    
-    # 1. Pitch shift using librosa's PSOLA (reliable)
-    if pitch_shift != 0:
-        y = librosa.effects.pitch_shift(y, sr=sr, n_steps=pitch_shift)
-    
-    # 2. Formant shift via spectral envelope manipulation
-    if formant_shift != 0:
-        y = shift_formants(y, sr, formant_shift)
-    
-    # 3. Add breathiness
-    if breathiness > 0:
-        noise = np.random.normal(0, breathiness * 0.1 * np.std(y), len(y))
-        y = y * (1 - breathiness) + noise * breathiness
-    
-    # Normalize to prevent clipping
-    y = y / (np.max(np.abs(y)) + 1e-6) * 0.95
-    
-    # Save
-    sf.write(output_audio_path, y, sr, subtype='PCM_16')
-    return output_audio_path
-
-
-def shift_formants(y, sr, formant_shift_semitones):
-    """
-    Shift formants by manipulating spectral envelope.
-    Positive = higher formants, Negative = lower formants.
-    """
-    if formant_shift_semitones == 0:
-        return y
-    
-    # Convert semitones to frequency ratio
-    ratio = 2 ** (formant_shift_semitones / 12.0)
-    
-    # STFT
-    n_fft = 2048
-    hop_length = 512
-    stft = librosa.stft(y, n_fft=n_fft, hop_length=hop_length)
-    magnitude = np.abs(stft)
-    phase = np.angle(stft)
-    
-    # Shift spectral envelope along frequency axis
-    # Simple approach: interpolate magnitude spectrum
-    n_freqs = magnitude.shape[0]
-    new_freqs = np.linspace(0, n_freqs - 1, int(n_freqs / ratio))
-    
-    if ratio < 1:  # Lower formants (stretch)
-        # Interpolate to expand
-        from scipy.interpolate import interp1d
-        interp_func = interp1d(np.arange(n_freqs), magnitude, axis=0, kind='linear', fill_value='extrapolate')
-        magnitude = interp_func(new_freqs)
-        # Truncate or pad to original size
-        if magnitude.shape[0] > n_freqs:
-            magnitude = magnitude[:n_freqs]
-        else:
-            pad = np.zeros((n_freqs - magnitude.shape[0], magnitude.shape[1]))
-            magnitude = np.vstack([magnitude, pad])
-    else:  # Higher formants (compress)
-        from scipy.interpolate import interp1d
-        interp_func = interp1d(np.arange(n_freqs), magnitude, axis=0, kind='linear', fill_value='extrapolate')
-        magnitude = interp_func(new_freqs)
-        if magnitude.shape[0] < n_freqs:
-            pad = np.zeros((n_freqs - magnitude.shape[0], magnitude.shape[1]))
-            magnitude = np.vstack([magnitude, pad])
-        else:
-            magnitude = magnitude[:n_freqs]
-    
-    # Reconstruct
-    stft_new = magnitude * np.exp(1j * phase)
-    y_new = librosa.istft(stft_new, hop_length=hop_length, length=len(y))
-    
-    return y_new
-
-
-def process_audio(
-    audio_file,
-    pitch_shift,
-    formant_shift,
-    breathiness,
-):
-    if audio_file is None:
-        return None, "请先上传音频文件"
-    
-    try:
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_out:
-            output_path = tmp_out.name
-        
-        convert_voice_simple(
-            input_audio_path=audio_file,
-            output_audio_path=output_path,
-            pitch_shift=pitch_shift,
-            formant_shift=formant_shift,
-            breathiness=breathiness,
-        )
-        
-        return output_path, "✅ 转换完成！"
-    except Exception as e:
-        import traceback
-        return None, f"❌ 转换失败: {str(e)}"
-
-
-def create_interface():
-    with gr.Blocks(title=TITLE, theme=gr.themes.Soft()) as demo:
-        gr.Markdown(f"# {TITLE}")
-        gr.Markdown(DESCRIPTION)
-        
-        with gr.Row():
-            with gr.Column(scale=1):
-                audio_input = gr.Audio(
-                    label="输入音频",
-                    type="filepath",
-                    sources=["upload", "microphone"],
-                )
-                
-                with gr.Group():
-                    gr.Markdown("### 参数")
-                    pitch_shift = gr.Slider(
-                        minimum=-12,
-                        maximum=0,
-                        value=-4,
-                        step=0.5,
-                        label="Pitch Shift (半音)",
-                        info="负值降低音调，建议 -3 ~ -6"
-                    )
-                    formant_shift = gr.Slider(
-                        minimum=-8,
-                        maximum=0,
-                        value=-3,
-                        step=0.5,
-                        label="Formant Shift (半音)",
-                        info="负值降低共振峰，建议 -2 ~ -4"
-                    )
-                    breathiness = gr.Slider(
-                        minimum=0.0,
-                        maximum=0.5,
-                        value=0.15,
-                        step=0.05,
-                        label="Breathiness (气声感)",
-                        info="模拟声带闭合不全，0.1-0.3 推荐"
-                    )
-                
-                convert_btn = gr.Button("🎭 转换为老年声", variant="primary", size="lg")
-                status = gr.Textbox(label="状态", interactive=False)
-            
-            with gr.Column(scale=1):
-                audio_output = gr.Audio(
-                    label="转换结果",
-                    type="filepath",
-                )
-                gr.Markdown("""
-                ### 💡 参数建议
-                | 声音类型 | Pitch Shift | Formant Shift | Breathiness |
-                |---------|-------------|---------------|-------------|
-                | 男性→老年男声 | -4 ~ -6 | -2 ~ -3 | 0.1 ~ 0.2 |
-                | 女性→老年女声 | -2 ~ -4 | -1 ~ -2 | 0.1 ~ 0.2 |
-                | 女性→老年男声 | -6 ~ -9 | -3 ~ -5 | 0.15 ~ 0.25 |
-                | 儿童→老年声 | -8 ~ -12 | -4 ~ -6 | 0.2 ~ 0.3 |
-                
-                **注意**：这是基于 DSP 的近似模拟，非 AI 声音转换。
-                效果取决于原音频质量，建议使用清晰、单人、少背景噪音的音频。
-                """)
-        
-        convert_btn.click(
-            fn=process_audio,
-            inputs=[audio_input, pitch_shift, formant_shift, breathiness],
-            outputs=[audio_output, status],
-            api_name="convert",
-        )
-    
-    return demo
-
-
-if __name__ == "__main__":
-    demo = create_interface()
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True,
-        share=False,
-    )

inference.py

@@ -1,191 +0,0 @@
-#!/usr/bin/env python3
-"""RVC Voice Conversion Inference for Elderly Voice."""
-import os
-import sys
-import torch
-import numpy as np
-import soundfile as sf
-from pathlib import Path
-
-# Add RVC to path
-rvc_path = Path("/app/rvc")
-sys.path.insert(0, str(rvc_path))
-
-# Import RVC modules with correct paths
-from infer.vc.pipeline import Pipeline
-from infer.hubert import load_hubert_model, extract_hubert_features
-from infer.module.models import SynthesizerTrnMs768NSFsid, SynthesizerTrnMs768NSFsid_nono
-from infer.audio import load_audio
-from tools.cuda_graph import run_cuda_graph
-
-# Model paths
-MODEL_DIR = Path("/app/models")
-ELDERLY_MODEL_DIR = MODEL_DIR / "elderly"
-
-def find_model_files():
-    """Find the elderly voice model files."""
-    pth_files = list(ELDERLY_MODEL_DIR.glob("*.pth"))
-    index_files = list(ELDERLY_MODEL_DIR.glob("*.index"))
-    
-    if not pth_files:
-        pth_files = list((MODEL_DIR / "weights").glob("*.pth")) if (MODEL_DIR / "weights").exists() else []
-    if not index_files:
-        index_files = list((MODEL_DIR / "weights").glob("*.index")) if (MODEL_DIR / "weights").exists() else []
-    
-    model_path = str(pth_files[0]) if pth_files else None
-    index_path = str(index_files[0]) if index_files else None
-    
-    return model_path, index_path
-
-MODEL_PATH, INDEX_PATH = find_model_files()
-
-# Device
-device = "cuda" if torch.cuda.is_available() else "cpu"
-is_half = torch.cuda.is_available()
-print(f"Using device: {device}, is_half: {is_half}")
-print(f"Model: {MODEL_PATH}")
-print(f"Index: {INDEX_PATH}")
-
-# Global objects (initialized lazily)
-_pipeline = None
-_hubert_model = None
-_net_g = None
-_tgt_sr = None
-_if_f0 = None
-_version = None
-
-def load_model():
-    """Load the RVC model and initialize pipeline."""
-    global _pipeline, _hubert_model, _net_g, _tgt_sr, _if_f0, _version
-    
-    if MODEL_PATH is None or not os.path.exists(MODEL_PATH):
-        raise RuntimeError(f"Model not found: {MODEL_PATH}")
-    
-    # Load hubert model
-    _hubert_model = load_hubert_model(device, is_half)
-    
-    # Load the model checkpoint
-    cpt = torch.load(MODEL_PATH, map_location=torch.device("cpu"))
-    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
-    _if_f0 = cpt.get("f0", 1)
-    _version = cpt.get("version", "v1")
-    
-    if _version == "v1":
-        if _if_f0 == 1:
-            _net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=False)
-        else:
-            _net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
-    elif _version == "v2":
-        if _if_f0 == 1:
-            _net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=False)
-        else:
-            _net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
-    
-    del _net_g.enc_q
-    _net_g.load_state_dict(cpt["weight"], strict=False)
-    _net_g = _net_g.float()
-    _net_g.eval().to(device)
-    _net_g.remove_weight_norm()
-    
-    if is_half:
-        _net_g = _net_g.half()
-    else:
-        _net_g = _net_g.float()
-    
-    _tgt_sr = cpt["config"][-1]
-    
-    # Create config object for pipeline
-    class PipelineConfig:
-        def __init__(self, device, is_half):
-            self.device = torch.device(device)
-            self.is_half = is_half
-            if is_half:
-                self.x_pad, self.x_query, self.x_center, self.x_max = 3, 10, 60, 65
-            else:
-                self.x_pad, self.x_query, self.x_center, self.x_max = 1, 6, 38, 41
-    
-    config = PipelineConfig(device, is_half)
-    _pipeline = Pipeline(_tgt_sr, config)
-    
-    print("Model loaded successfully")
-
-def convert_voice(
-    input_audio_path: str,
-    output_audio_path: str,
-    f0_up_key: int = 0,
-    f0_method: str = "rmvpe",
-    index_rate: float = 0.75,
-    filter_radius: int = 3,
-    rms_mix_rate: float = 0.25,
-    protect: float = 0.33,
-):
-    """
-    Convert input audio to elderly voice using RVC Pipeline.
-    
-    Args:
-        input_audio_path: Path to input audio file
-        output_audio_path: Path to save output audio
-        f0_up_key: Pitch shift in semitones (negative = lower pitch)
-        f0_method: Pitch extraction method (pm, rmvpe, fcpe)
-        index_rate: Index influence (0-1)
-        filter_radius: Median filter radius for pitch (not used in pipeline directly)
-        rms_mix_rate: RMS mix rate for volume
-        protect: Protect voiceless consonants
-    """
-    global _pipeline, _hubert_model, _net_g, _tgt_sr, _if_f0, _version
-    
-    # Lazy load model
-    if _pipeline is None:
-        load_model()
-    
-    # Use provided index path or fallback
-    index_path = INDEX_PATH if INDEX_PATH and os.path.exists(INDEX_PATH) else ""
-    
-    # Load audio
-    audio = load_audio(input_audio_path, 16000)
-    
-    # Prepare times array for profiling
-    times = [0.0, 0.0, 0.0]
-    
-    # Run pipeline
-    audio_opt = _pipeline.pipeline(
-        model=_hubert_model,
-        net_g=_net_g,
-        sid=0,
-        audio=audio,
-        times=times,
-        f0_up_key=f0_up_key,
-        f0_method=f0_method,
-        file_index=index_path,
-        index_rate=index_rate,
-        if_f0=_if_f0,
-        tgt_sr=_tgt_sr,
-        resample_sr=0,
-        rms_mix_rate=rms_mix_rate,
-        version=_version,
-        protect=protect,
-    )
-    
-    # Save output
-    sf.write(output_audio_path, audio_opt, _tgt_sr, subtype='PCM_16')
-    return output_audio_path
-
-
-if __name__ == "__main__":
-    import argparse
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--input", required=True, help="Input audio file")
-    parser.add_argument("--output", required=True, help="Output audio file")
-    parser.add_argument("--pitch", type=int, default=-2, help="Pitch shift (semitones, negative for deeper)")
-    parser.add_argument("--method", default="rmvpe", help="F0 method")
-    parser.add_argument("--index-rate", type=float, default=0.75, help="Index rate")
-    args = parser.parse_args()
-    
-    convert_voice(
-        args.input, 
-        args.output, 
-        f0_up_key=args.pitch, 
-        f0_method=args.method, 
-        index_rate=args.index_rate
-    )
-    print(f"Converted: {args.output}")

inference_simple.py

@@ -1,22 +0,0 @@
-#!/usr/bin/env python3
-"""Simple inference module for compatibility."""
-# The conversion logic is in app.py directly
-# This file exists for import compatibility
-
-def convert_voice(input_audio_path, output_audio_path, **kwargs):
-    """Placeholder - actual implementation in app.py"""
-    pass
-
-MODEL_PATH = None
-vc = None
-
-if __name__ == "__main__":
-    import argparse
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--input", required=True)
-    parser.add_argument("--output", required=True)
-    parser.add_argument("--pitch", type=float, default=-4)
-    parser.add_argument("--formant", type=float, default=-3)
-    parser.add_argument("--breathiness", type=float, default=0.15)
-    args = parser.parse_args()
-    print(f"Use app.py for conversion: {args.input} -> {args.output}")

packages.txt

@@ -0,0 +1 @@
+ffmpeg

requirements.txt

@@ -1,24 +1,18 @@
-# Minimal RVC inference dependencies (torch/torchaudio installed separately)
-numpy>=1.24.0,<2
-scipy>=1.10.0,<2
-librosa>=0.10.0,<0.11
-soundfile>=0.12.0,<1
-faiss-cpu>=1.13.0,<2
-praat-parselmouth>=0.4.5,<1
-transformers>=4.49.0,<4.50
-tqdm>=4.67.0,<5
-huggingface_hub>=0.20.0
-ffmpeg-python>=0.2.0,<1
-av>=15.1.0,<16
-onnxruntime>=1.24.4,<2
-scikit-learn>=1.6.0,<2
-PyYAML>=6.0
-
-gradio>=4.44.0
-fastapi>=0.88.0,<0.100
-uvicorn>=0.20.0,<0.23
-pydantic>=1.10.13,<2
-httpx>=0.23.0,<0.24
-anyio>=3.6.2,<4
-starlette>=0.22.0,<0.28
-websockets>=10.4,<11
+torch==2.1.0
+torchaudio==2.1.0
+librosa==0.9.1
+pydub==0.25.1
+wavmark==0.0.3
+numpy==1.22.0
+eng_to_ipa==0.0.2
+inflect==7.0.0
+unidecode==1.3.7
+whisper-timestamped==1.14.2
+pypinyin
+cn2an
+jieba
+langid
+gradio==4.44.0
+huggingface_hub
+soundfile
+git+https://github.com/myshell-ai/OpenVoice.git