Release frozen StreamPA seed 1337 with validated standalone inference
f8a1702 verified about 2 months ago
65.6 kB
# Frozen inference function extraction. See source provenance and licenses.
import gc
import math
import re
from collections import Counter
from itertools import groupby
import numpy as np
import librosa
import torch
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
EPS = 1e-8
EPS_TOKEN = '<eps>'

def normalize_phone(phone):
    phone = phone.upper().strip()
    phone = phone.replace('[', '').replace(']', '')
    phone = re.sub(r'[_\d].*', '', phone)
    return phone


def normalize_word(word):
    word = word.upper().strip()
    word = word.replace('[', '').replace(']', '')
    word = re.sub(r"[^A-Z']", '', word)
    return word


def monotonic_align(neg_log_probs, phone_ids):
    frame_num = neg_log_probs.shape[0]
    token_num = len(phone_ids)
    if frame_num < token_num:
        raise ValueError(f'Not enough frames ({frame_num}) to align {token_num} phones.')

    dp = np.full((frame_num, token_num), np.inf, dtype=np.float32)
    back = np.zeros((frame_num, token_num), dtype=np.int8)
    dp[0, 0] = neg_log_probs[0, phone_ids[0]]

    for t in range(1, frame_num):
        max_token = min(t, token_num - 1)
        for tok in range(max_token + 1):
            stay_cost = dp[t - 1, tok]
            move_cost = dp[t - 1, tok - 1] if tok > 0 else np.inf
            if move_cost < stay_cost:
                dp[t, tok] = move_cost + neg_log_probs[t, phone_ids[tok]]
                back[t, tok] = 1
            else:
                dp[t, tok] = stay_cost + neg_log_probs[t, phone_ids[tok]]

    path = np.zeros(frame_num, dtype=np.int32)
    tok = token_num - 1
    for t in range(frame_num - 1, -1, -1):
        path[t] = tok
        if t > 0 and back[t, tok] == 1:
            tok -= 1
    return path

def select_visible_frames(probs, keep_mask, audio_end, frame_step):
    frame_limit = max(int(math.ceil(audio_end / max(frame_step, EPS))), 1)
    frame_mask = np.zeros(len(probs), dtype=bool)
    frame_mask[: min(frame_limit, len(probs))] = True
    visible_mask = keep_mask & frame_mask
    return np.flatnonzero(visible_mask), probs[visible_mask]

def softmax_1d(values):
    arr = np.asarray(values, dtype=np.float64)
    if arr.size == 0:
        return arr.astype(np.float32)
    pos_inf = np.isposinf(arr)
    if np.any(pos_inf):
        out = np.zeros_like(arr, dtype=np.float64)
        out[pos_inf] = 1.0 / float(np.sum(pos_inf))
        return out.astype(np.float32)
    finite = np.isfinite(arr)
    if not np.any(finite):
        return np.full(arr.shape, 1.0 / float(arr.size), dtype=np.float32)
    min_finite = np.min(arr[finite])
    arr = np.where(finite, arr, min_finite - 50.0)
    arr = arr - np.max(arr)
    exp_arr = np.exp(arr)
    return (exp_arr / np.clip(exp_arr.sum(), EPS, None)).astype(np.float32)


def entropy_np(probs):
    probs = np.asarray(probs, dtype=np.float64)
    probs = probs[probs > 0]
    if probs.size == 0:
        return 0.0
    return float(-np.sum(probs * np.log(np.clip(probs, EPS, None))))


def top_margin(probs):
    probs = np.asarray(probs, dtype=np.float32)
    if probs.size == 0:
        return 0.0, 0.0
    top = np.sort(probs)[::-1]
    top1 = float(top[0])
    top2 = float(top[1]) if top.size > 1 else 0.0
    return top1, float(top1 - top2)


def js_divergence(p, q):
    p = np.asarray(p, dtype=np.float64)
    q = np.asarray(q, dtype=np.float64)
    p = p / np.clip(p.sum(), EPS, None)
    q = q / np.clip(q.sum(), EPS, None)
    m = 0.5 * (p + q)
    kl_pm = np.sum(p * (np.log(np.clip(p, EPS, None)) - np.log(np.clip(m, EPS, None))))
    kl_qm = np.sum(q * (np.log(np.clip(q, EPS, None)) - np.log(np.clip(m, EPS, None))))
    return float(0.5 * (kl_pm + kl_qm))


def safe_json(value):
    if isinstance(value, dict):
        return {str(key): safe_json(val) for key, val in value.items()}
    if isinstance(value, (list, tuple)):
        return [safe_json(item) for item in value]
    if isinstance(value, np.ndarray):
        return safe_json(value.tolist())
    if isinstance(value, (np.floating, np.integer)):
        return value.item()
    return value


def softmax_np_matrix(values):
    arr = np.asarray(values, dtype=np.float64)
    if arr.size == 0:
        return arr.astype(np.float32)
    arr = arr - np.max(arr, axis=-1, keepdims=True)
    exp_arr = np.exp(arr)
    return (exp_arr / np.clip(np.sum(exp_arr, axis=-1, keepdims=True), EPS, None)).astype(np.float32)


def audio_array_logits(audio, processor, model, sample_rate, device):
    audio = np.asarray(audio, dtype=np.float32)
    if hasattr(processor, 'audio_preprocess'):
        inputs = processor.audio_preprocess(audio, sr=sample_rate)
        inputs = torch.tensor(inputs, dtype=torch.float32).unsqueeze(0).to(device)
        model_inputs = {'input_values': inputs}
    else:
        model_inputs = processor(audio, sampling_rate=sample_rate, return_tensors='pt')
        model_inputs = {key: value.to(device) for key, value in model_inputs.items()}
    with torch.no_grad():
        logits = model(**model_inputs).logits[0].detach().cpu().numpy()
    return softmax_np_matrix(logits), float(len(audio) / sample_rate)


def non_overlapping_occurrences(tokens, pattern):
    positions = []
    pattern_size = len(pattern)
    cursor = 0
    while cursor <= len(tokens) - pattern_size:
        if tokens[cursor : cursor + pattern_size] == pattern:
            positions.append(cursor)
            cursor += pattern_size
        else:
            cursor += 1
    return positions


def find_repeated_ngram(
    words,
    min_repeats,
    max_ngram_size=12,
    strict=False,
    min_coverage=0.6,
    dominant_token_ratio=0.5,
):
    if min_repeats <= 0:
        return None
    tokens = [str(word['text']) for word in words]
    if not tokens:
        return None

    for ngram_size in range(1, min(max_ngram_size, len(tokens)) + 1):
        span = ngram_size * min_repeats
        for start in range(0, len(tokens) - span + 1):
            pattern = tokens[start : start + ngram_size]
            repeats = 1
            cursor = start + ngram_size
            while tokens[cursor : cursor + ngram_size] == pattern:
                repeats += 1
                cursor += ngram_size
            if repeats >= min_repeats:
                return {
                    'start': int(start),
                    'ngram_size': int(ngram_size),
                    'repeats': int(repeats),
                    'pattern': pattern,
                    'detection': 'contiguous_cycle',
                }

    if not strict:
        return None

    token_counts = Counter(tokens)
    dominant_token, dominant_count = token_counts.most_common(1)[0]
    token_ratio = float(dominant_count) / float(len(tokens))
    if dominant_count >= max(min_repeats * 2, 8) and token_ratio >= dominant_token_ratio:
        return {
            'start': None,
            'ngram_size': 1,
            'repeats': int(dominant_count),
            'pattern': [dominant_token],
            'coverage': token_ratio,
            'detection': 'dominant_token',
        }

    max_size = min(max_ngram_size, len(tokens) // max(min_repeats, 1))
    for ngram_size in range(max_size, 1, -1):
        candidates = Counter(
            tuple(tokens[start : start + ngram_size])
            for start in range(0, len(tokens) - ngram_size + 1)
        )
        for pattern, raw_count in candidates.most_common():
            if raw_count < min_repeats:
                break
            positions = non_overlapping_occurrences(tokens, list(pattern))
            if len(positions) < min_repeats:
                continue
            coverage = float(len(positions) * ngram_size) / float(len(tokens))
            if coverage >= min_coverage:
                return {
                    'start': int(positions[0]),
                    'ngram_size': int(ngram_size),
                    'repeats': int(len(positions)),
                    'pattern': list(pattern),
                    'coverage': coverage,
                    'detection': 'high_coverage_periodicity',
                }
    return None


class PhoneMapper:
    def __init__(self, lexicon, phn_dict):
        self.lexicon = lexicon
        self.phn_dict = phn_dict
        self.cache = {}
        self.fallback_phone = 'AH' if 'AH' in phn_dict else (next(iter(phn_dict)) if phn_dict else None)
        self.g2p = None
        try:
            from g2p_en import G2p

            self.g2p = G2p()
        except Exception:
            pass

    def word_to_phones(self, word):
        word = normalize_word(word)
        if word in self.lexicon:
            return list(self.lexicon[word]), 'lexicon'
        if word in self.cache:
            return self.cache[word]
        if self.g2p is not None:
            try:
                phones = []
                for raw_phone in self.g2p(word.lower()):
                    phone = normalize_phone(raw_phone)
                    if phone and phone in self.phn_dict:
                        phones.append(phone)
                if phones:
                    result = (phones, 'g2p_en')
                    self.cache[word] = result
                    return result
            except Exception:
                self.g2p = None
        if self.fallback_phone is None:
            result = ([], 'unmapped')
        else:
            approx_len = max(1, int(math.ceil(len(re.sub(r'[^A-Z]', '', word)) / 3.0)))
            result = ([self.fallback_phone] * approx_len, 'fallback_phone')
        self.cache[word] = result
        return result

    def words_to_phone_sequence(self, words):
        phones = []
        phone_to_word = []
        source_counts = Counter()
        for word_idx, word in enumerate(words):
            cur_phones, source = self.word_to_phones(word)
            source_counts[source] += 1
            for phone in cur_phones:
                phones.append(phone)
                phone_to_word.append(word_idx)
        return phones, phone_to_word, dict(source_counts)


class WhisperNBestGenerator:
    def __init__(self, model_name, language, device, nbest, beam_size, max_new_tokens, no_repeat_ngram_size):
        self.processor = AutoProcessor.from_pretrained(model_name)
        torch_dtype = torch.float16 if str(device).startswith('cuda') else torch.float32
        self.model = AutoModelForSpeechSeq2Seq.from_pretrained(model_name, torch_dtype=torch_dtype)
        self.model.to(device)
        self.model.eval()
        self.model_dtype = next(self.model.parameters()).dtype
        self.language = language
        self.device = device
        self.nbest = int(nbest)
        self.beam_size = max(int(beam_size), int(nbest))
        self.max_new_tokens = int(max_new_tokens)
        self.no_repeat_ngram_size = int(no_repeat_ngram_size)
        self.forced_decoder_ids = None
        if hasattr(self.processor, 'get_decoder_prompt_ids'):
            try:
                self.forced_decoder_ids = self.processor.get_decoder_prompt_ids(
                    language=language,
                    task='transcribe',
                )
            except Exception:
                self.forced_decoder_ids = None

    @torch.inference_mode()
    def generate(self, audio, sample_rate, audio_end):
        inputs = self.processor(
            np.asarray(audio, dtype=np.float32),
            sampling_rate=sample_rate,
            return_tensors='pt',
            return_attention_mask=True,
        )
        inputs = {key: value.to(self.device) for key, value in inputs.items()}
        for key in ('input_features', 'input_values'):
            if key in inputs and torch.is_floating_point(inputs[key]):
                inputs[key] = inputs[key].to(dtype=self.model_dtype)
        generate_kwargs = {
            'num_beams': self.beam_size,
            'num_return_sequences': self.nbest,
            'max_new_tokens': self.max_new_tokens,
            'return_dict_in_generate': True,
            'output_scores': True,
            'do_sample': False,
        }
        if self.no_repeat_ngram_size > 0:
            generate_kwargs['no_repeat_ngram_size'] = self.no_repeat_ngram_size
        if self.forced_decoder_ids is not None:
            generate_kwargs['forced_decoder_ids'] = self.forced_decoder_ids
        output = self.model.generate(**inputs, **generate_kwargs)
        if torch.is_tensor(output):
            sequences = output
            sequence_scores = [0.0 for _ in range(int(sequences.shape[0]))]
            transition_scores = []
        else:
            sequences = output.sequences
            if getattr(output, 'sequences_scores', None) is not None:
                sequence_scores = output.sequences_scores.detach().float().cpu().numpy().tolist()
            else:
                sequence_scores = [0.0 for _ in range(int(sequences.shape[0]))]
            transition_scores = self._transition_scores(output)
        texts = self.processor.batch_decode(sequences, skip_special_tokens=True)

        rows = []
        for rank, (text, sequence_score) in enumerate(zip(texts, sequence_scores)):
            token_ids = sequences[rank].detach().cpu().tolist()
            token_logprobs = (
                transition_scores[rank]
                if rank < len(transition_scores)
                else [float(sequence_score) for _ in token_ids]
            )
            token_rows, word_rows = self._token_word_scores(token_ids, token_logprobs)
            words = [row['word'] for row in word_rows]
            length_norm = (
                float(np.mean([row['logprob'] for row in token_rows]))
                if token_rows
                else float(sequence_score)
            )
            rows.append({
                'rank': int(rank),
                'text': ' '.join(words).lower(),
                'words': words,
                'logprob': float(length_norm),
                'sequence_score': float(sequence_score),
                'length_normalized_sequence_score': float(length_norm),
                'token_ids': [int(row['token_id']) for row in token_rows],
                'token_logprobs': [float(row['logprob']) for row in token_rows],
                'token_confidences': [float(row['confidence']) for row in token_rows],
                'word_token_ranges': [[int(row['token_start']), int(row['token_end'])] for row in word_rows],
                'word_logprobs': [float(row['logprob']) for row in word_rows],
                'word_confidences': [float(row['confidence']) for row in word_rows],
                'word_timestamps': estimate_word_timestamps(words, audio_end),
            })
        if str(self.device).startswith('cuda'):
            torch.cuda.empty_cache()
        gc.collect()
        return rows

    def _transition_scores(self, output):
        if getattr(output, 'scores', None) is None:
            return []
        try:
            sequences = output.sequences.detach().cpu()
            step_scores = [score.detach().float().cpu() for score in output.scores]
            if not step_scores:
                return []
            beam_indices = getattr(output, 'beam_indices', None)
            if beam_indices is not None:
                beam_indices = beam_indices.detach().cpu()
            generated_steps = len(step_scores)
            start = max(0, int(sequences.shape[1]) - generated_steps)
            rows = []
            for seq_idx in range(int(sequences.shape[0])):
                cur = []
                for step, logits in enumerate(step_scores):
                    token_pos = start + step
                    if token_pos >= int(sequences.shape[1]):
                        break
                    token_id = int(sequences[seq_idx, token_pos])
                    if token_id < 0 or token_id >= int(logits.shape[-1]):
                        cur.append(0.0)
                        continue
                    score_row = min(seq_idx, int(logits.shape[0]) - 1)
                    if beam_indices is not None and seq_idx < int(beam_indices.shape[0]) and token_pos < int(beam_indices.shape[1]):
                        candidate = int(beam_indices[seq_idx, token_pos])
                        if 0 <= candidate < int(logits.shape[0]):
                            score_row = candidate
                    log_probs = torch.log_softmax(logits[score_row], dim=-1)
                    cur.append(float(log_probs[token_id].item()))
                rows.append(cur)
            return rows
        except Exception:
            return []

    def _token_word_scores(self, token_ids, token_logprobs):
        tokenizer = self.processor.tokenizer
        special_ids = set(getattr(tokenizer, 'all_special_ids', []) or [])
        if len(token_logprobs) != len(token_ids):
            aligned = [0.0 for _ in token_ids]
            offset = max(0, len(token_ids) - len(token_logprobs))
            for idx, value in enumerate(token_logprobs[-len(token_ids) :]):
                pos = offset + idx
                if pos < len(aligned):
                    aligned[pos] = float(value)
            token_logprobs = aligned
        token_rows = []
        words = []
        current = []
        current_token_ids = []
        current_logprobs = []
        current_start = None

        def flush():
            nonlocal current, current_token_ids, current_logprobs, current_start
            word = normalize_word(''.join(current))
            if word:
                avg_logprob = float(np.mean(current_logprobs)) if current_logprobs else 0.0
                words.append({
                    'word': word,
                    'token_start': int(current_start if current_start is not None else 0),
                    'token_end': int((current_token_ids[-1] + 1) if current_token_ids else 0),
                    'logprob': avg_logprob,
                    'confidence': float(np.clip(np.exp(avg_logprob), 0.0, 1.0)),
                })
            current = []
            current_token_ids = []
            current_logprobs = []
            current_start = None

        for pos, token_id in enumerate(token_ids):
            if int(token_id) in special_ids:
                continue
            logprob = float(token_logprobs[pos]) if pos < len(token_logprobs) else 0.0
            try:
                text = tokenizer.decode([int(token_id)], skip_special_tokens=True, clean_up_tokenization_spaces=False)
            except TypeError:
                text = tokenizer.decode([int(token_id)], skip_special_tokens=True)
            if not text:
                continue
            token_rows.append({
                'token_id': int(token_id),
                'logprob': logprob,
                'confidence': float(np.clip(np.exp(logprob), 0.0, 1.0)),
            })
            for char in text:
                if char.isalpha() or char == "'":
                    if current_start is None:
                        current_start = pos
                    current.append(char)
                    current_token_ids.append(pos)
                    current_logprobs.append(logprob)
                else:
                    flush()
        flush()
        return token_rows, words


def normalize_text_to_words(text):
    tokens = []
    for token in re.findall(r"[A-Za-z']+", text):
        norm = normalize_word(token)
        if norm:
            tokens.append(norm)
    return tokens


def estimate_word_timestamps(words, audio_end):
    if not words:
        return []
    step = float(audio_end) / max(len(words), 1)
    rows = []
    for idx, word in enumerate(words):
        rows.append({
            'word': word.lower(),
            'start': float(idx * step),
            'end': float(min(audio_end, (idx + 1) * step)),
            'source': 'duration_proportional_estimate',
        })
    return rows


def hypothesis_phone_rows(row, phone_mapper, phn_dict):
    phones = []
    phone_ids = []
    phone_to_word = []
    phone_confidences = []
    source_counts = Counter()
    word_confidences = row.get('word_confidences') or [1.0 for _ in row.get('words', [])]
    for word_idx, word in enumerate(row.get('words', [])):
        cur_phones, source = phone_mapper.word_to_phones(word)
        source_counts[source] += 1
        word_conf = float(word_confidences[word_idx]) if word_idx < len(word_confidences) else 1.0
        word_conf = float(np.clip(word_conf, 0.0, 1.0))
        for phone in cur_phones:
            if phone not in phn_dict:
                continue
            phones.append(phone)
            phone_ids.append(int(phn_dict[phone]))
            phone_to_word.append(int(word_idx))
            phone_confidences.append(word_conf)
    return {
        'phones': phones,
        'phone_ids': phone_ids,
        'phone_to_word': phone_to_word,
        'phone_confidences': phone_confidences,
        'source_counts': dict(source_counts),
    }


def align_hypothesis_with_charsiu(item, row, phone_mapper, phn_dict, phone_to_frame_id, audio_end):
    hyp = hypothesis_phone_rows(row, phone_mapper, phn_dict)
    words = row.get('words', [])
    if not words or not hyp['phone_ids']:
        row['word_timestamps'] = []
        row['timestamp_source'] = 'empty_hypothesis'
        row['phone_times'] = []
        row['phone_acoustic_supports'] = []
        return row

    kept_indices, kept_probs = select_visible_frames(item['probs'], item['keep_mask'], audio_end, item['frame_step'])
    frame_phone_ids = []
    id_to_phone = {idx: phone for phone, idx in phn_dict.items()}
    try:
        for phone_idx in hyp['phone_ids']:
            phone = id_to_phone[int(phone_idx)]
            frame_phone_ids.append(phone_to_frame_id[phone])
        if kept_probs.shape[0] < len(frame_phone_ids):
            raise ValueError('not_enough_visible_frames_for_hypothesis')
        path = monotonic_align(-np.log(np.clip(kept_probs, EPS, None)), frame_phone_ids)
        phone_times = []
        phone_acoustic_supports = []
        for phone_pos, frame_phone_id in enumerate(frame_phone_ids):
            local_frames = kept_indices[path == phone_pos]
            if local_frames.size == 0:
                raise ValueError('empty_aligned_phone_span')
            start_time = float(local_frames.min() * item['frame_step'])
            end_time = float((local_frames.max() + 1) * item['frame_step'])
            support = float(np.mean(item['probs'][local_frames, frame_phone_id]))
            phone_times.append((start_time, min(float(audio_end), end_time)))
            phone_acoustic_supports.append(float(np.clip(support, 0.0, 1.0)))
        word_rows = []
        for word_idx, word in enumerate(words):
            positions = [idx for idx, cur_word in enumerate(hyp['phone_to_word']) if cur_word == word_idx]
            if not positions:
                continue
            start = min(phone_times[idx][0] for idx in positions)
            end = max(phone_times[idx][1] for idx in positions)
            word_rows.append({
                'word': word,
                'start': float(start),
                'end': float(end),
                'source': 'charsiu_hypothesis_alignment',
                'word_logprob': float(row.get('word_logprobs', [0.0] * len(words))[word_idx])
                if word_idx < len(row.get('word_logprobs', []))
                else 0.0,
                'word_confidence': float(row.get('word_confidences', [1.0] * len(words))[word_idx])
                if word_idx < len(row.get('word_confidences', []))
                else 1.0,
            })
        row['word_timestamps'] = word_rows
        row['timestamp_source'] = 'charsiu_hypothesis_alignment'
        row['phone_times'] = phone_times
        row['phone_acoustic_supports'] = phone_acoustic_supports
    except Exception as exc:
        row['word_timestamps'] = estimate_word_timestamps(words, audio_end)
        row['timestamp_source'] = 'duration_proportional_fallback'
        row['timestamp_fallback_reason'] = repr(exc)
        phone_times = []
        phone_acoustic_supports = []
        for phone_idx, word_idx in enumerate(hyp['phone_to_word']):
            if word_idx < len(row['word_timestamps']):
                word_time = row['word_timestamps'][word_idx]
                phone_times.append((float(word_time['start']), float(word_time['end'])))
            else:
                phone_times.append((0.0, float(audio_end)))
            phone_acoustic_supports.append(0.0)
        row['phone_times'] = phone_times
        row['phone_acoustic_supports'] = phone_acoustic_supports
    row.update(hyp)
    return row


def interval_overlap_ratio(a, b):
    if a is None or b is None:
        return 0.0
    a0, a1 = a
    b0, b1 = b
    inter = max(0.0, min(float(a1), float(b1)) - max(float(a0), float(b0)))
    union = max(float(a1), float(b1)) - min(float(a0), float(b0))
    if union <= 0:
        return 0.0
    return inter / union


def align_token_sequences(
    ref_tokens,
    hyp_tokens,
    mismatch_cost=1.0,
    ref_times=None,
    hyp_times=None,
    hyp_confidences=None,
    time_overlap_reward=0.0,
):
    n = len(ref_tokens)
    m = len(hyp_tokens)
    dp = np.zeros((n + 1, m + 1), dtype=np.float32)
    back = {}
    for i in range(1, n + 1):
        dp[i, 0] = dp[i - 1, 0] + 1.0
        back[(i, 0)] = (i - 1, 0, 'delete')
    for j in range(1, m + 1):
        dp[0, j] = dp[0, j - 1] + 1.0
        back[(0, j)] = (0, j - 1, 'insert')
    for i in range(1, n + 1):
        for j in range(1, m + 1):
            sub_cost = 0.0 if ref_tokens[i - 1] == hyp_tokens[j - 1] else mismatch_cost
            if ref_times is not None and hyp_times is not None:
                sub_cost -= float(time_overlap_reward) * interval_overlap_ratio(ref_times[i - 1], hyp_times[j - 1])
            if hyp_confidences is not None:
                sub_cost -= 0.1 * float(np.clip(hyp_confidences[j - 1], 0.0, 1.0))
            sub_cost = max(0.0, sub_cost)
            candidates = [
                (dp[i - 1, j - 1] + sub_cost, i - 1, j - 1, 'match'),
                (dp[i - 1, j] + 1.0, i - 1, j, 'delete'),
                (dp[i, j - 1] + 1.0, i, j - 1, 'insert'),
            ]
            cost, prev_i, prev_j, op = min(candidates, key=lambda item: item[0])
            dp[i, j] = cost
            back[(i, j)] = (prev_i, prev_j, op)

    pairs = []
    i, j = n, m
    while i > 0 or j > 0:
        prev_i, prev_j, op = back[(i, j)]
        if op == 'match':
            pairs.append((i - 1, j - 1))
        elif op == 'delete':
            pairs.append((i - 1, None))
        else:
            pairs.append((None, j - 1))
        i, j = prev_i, prev_j
    pairs.reverse()
    return pairs


def slot_top_phone(slot_counts, eps_index):
    best_idx = None
    best_mass = -1.0
    for phone_idx, mass in slot_counts.items():
        if phone_idx == eps_index:
            continue
        if mass > best_mass:
            best_idx = phone_idx
            best_mass = mass
    return best_idx


def slot_time(slot):
    mass = max(float(slot.get('time_mass', 0.0)), EPS)
    if mass <= EPS:
        return None
    return float(slot.get('start_sum', 0.0) / mass), float(slot.get('end_sum', 0.0) / mass)


def add_slot_phone_mass(slot, phone_idx, phone_mass, eps_index, eps_mass=0.0, phone_time=None, confidence=1.0):
    if phone_mass > 0:
        slot['counts'][int(phone_idx)] += float(phone_mass)
    if eps_mass > 0:
        slot['counts'][eps_index] += float(eps_mass)
    if phone_time is not None and phone_mass > 0:
        slot['start_sum'] = slot.get('start_sum', 0.0) + float(phone_time[0]) * float(phone_mass)
        slot['end_sum'] = slot.get('end_sum', 0.0) + float(phone_time[1]) * float(phone_mass)
        slot['time_mass'] = slot.get('time_mass', 0.0) + float(phone_mass)
    slot['confidence_sum'] = slot.get('confidence_sum', 0.0) + float(confidence) * max(float(phone_mass), 0.0)
    slot['confidence_mass'] = slot.get('confidence_mass', 0.0) + max(float(phone_mass), 0.0)


def insert_slot(slots, insert_at, eps_index, prior_mass):
    slot = {'counts': Counter(), 'start_sum': 0.0, 'end_sum': 0.0, 'time_mass': 0.0}
    if prior_mass > 0:
        slot['counts'][eps_index] += float(prior_mass)
    slots.insert(insert_at, slot)
    return slot


def build_pcn_from_hypotheses(hypotheses, phone_mapper, phn_dict):
    eps_index = len(phn_dict)
    phone_dim = len(phn_dict) + 1
    hyp_logprobs = [float(row.get('length_normalized_sequence_score', row.get('logprob', 0.0))) for row in hypotheses]
    hyp_weights = softmax_1d(hyp_logprobs)
    id_to_phone = {idx: phone for phone, idx in phn_dict.items()}

    hyp_phone_rows = []
    for row, weight in zip(hypotheses, hyp_weights):
        if 'phone_ids' not in row:
            row = align_hypothesis_with_charsiu(
                item={'probs': np.zeros((0, phone_dim), dtype=np.float32), 'keep_mask': np.zeros((0,), dtype=np.int32), 'frame_step': 0.02},
                row=row,
                phone_mapper=phone_mapper,
                phn_dict=phn_dict,
                phone_to_frame_id={},
                audio_end=0.0,
            )
        phone_ids = [int(phone_idx) for phone_idx in row.get('phone_ids', [])]
        phone_confidences = [float(x) for x in row.get('phone_confidences', [1.0] * len(phone_ids))]
        phone_acoustic = [float(x) for x in row.get('phone_acoustic_supports', [0.0] * len(phone_ids))]
        phone_times = row.get('phone_times', [None] * len(phone_ids))
        hyp_phone_rows.append({
            'phones': row.get('phones', []),
            'phone_ids': phone_ids,
            'phone_to_word': list(row.get('phone_to_word', [])),
            'phone_confidences': phone_confidences,
            'phone_acoustic_supports': phone_acoustic,
            'phone_times': phone_times,
            'weight': float(weight),
            'source_counts': row.get('source_counts', {}),
        })

    slots = []
    total_mass = 0.0
    for hyp_idx, hyp in enumerate(hyp_phone_rows):
        phone_ids = hyp['phone_ids']
        weight = float(hyp['weight'])
        if hyp_idx == 0:
            for local_idx, phone_idx in enumerate(phone_ids):
                slot = {'counts': Counter(), 'start_sum': 0.0, 'end_sum': 0.0, 'time_mass': 0.0}
                quality = float(np.clip(
                    0.5 * hyp['phone_confidences'][local_idx] + 0.5 * hyp['phone_acoustic_supports'][local_idx],
                    0.0,
                    1.0,
                ))
                add_slot_phone_mass(
                    slot,
                    phone_idx,
                    weight * quality,
                    eps_index,
                    eps_mass=weight * (1.0 - quality),
                    phone_time=hyp['phone_times'][local_idx] if local_idx < len(hyp['phone_times']) else None,
                    confidence=hyp['phone_confidences'][local_idx],
                )
                slots.append(slot)
            total_mass += weight
            continue

        consensus_slot_ids = []
        consensus_phone_ids = []
        for slot_idx, slot in enumerate(slots):
            top_idx = slot_top_phone(slot['counts'], eps_index)
            if top_idx is not None:
                consensus_slot_ids.append(slot_idx)
                consensus_phone_ids.append(top_idx)
        consensus_times = [slot_time(slots[slot_idx]) for slot_idx in consensus_slot_ids]
        pairs = align_token_sequences(
            consensus_phone_ids,
            phone_ids,
            ref_times=consensus_times,
            hyp_times=hyp.get('phone_times'),
            hyp_confidences=hyp.get('phone_confidences'),
            time_overlap_reward=0.3,
        )
        cursor = 0
        matched_slots = set()
        for ref_pos, hyp_pos in pairs:
            if ref_pos is None:
                insert_at = consensus_slot_ids[cursor] if cursor < len(consensus_slot_ids) else len(slots)
                slot = insert_slot(slots, insert_at, eps_index, total_mass)
                quality = float(np.clip(
                    0.5 * hyp['phone_confidences'][hyp_pos] + 0.5 * hyp['phone_acoustic_supports'][hyp_pos],
                    0.0,
                    1.0,
                ))
                add_slot_phone_mass(
                    slot,
                    phone_ids[hyp_pos],
                    weight * quality,
                    eps_index,
                    eps_mass=weight * (1.0 - quality),
                    phone_time=hyp['phone_times'][hyp_pos] if hyp_pos < len(hyp['phone_times']) else None,
                    confidence=hyp['phone_confidences'][hyp_pos],
                )
                cursor += 1
            else:
                slot_idx = consensus_slot_ids[ref_pos]
                matched_slots.add(slot_idx)
                cursor = ref_pos + 1
                if hyp_pos is None:
                    slots[slot_idx]['counts'][eps_index] += weight
                else:
                    quality = float(np.clip(
                        0.5 * hyp['phone_confidences'][hyp_pos] + 0.5 * hyp['phone_acoustic_supports'][hyp_pos],
                        0.0,
                        1.0,
                    ))
                    add_slot_phone_mass(
                        slots[slot_idx],
                        phone_ids[hyp_pos],
                        weight * quality,
                        eps_index,
                        eps_mass=weight * (1.0 - quality),
                        phone_time=hyp['phone_times'][hyp_pos] if hyp_pos < len(hyp['phone_times']) else None,
                        confidence=hyp['phone_confidences'][hyp_pos],
                    )
        for slot_idx, slot in enumerate(slots):
            if slot_idx not in matched_slots and sum(slot['counts'].values()) < total_mass + weight - EPS:
                if slot_top_phone(slot['counts'], eps_index) is not None:
                    slot['counts'][eps_index] += weight
        total_mass += weight

    if not slots:
        slot = {'counts': Counter()}
        slot['counts'][eps_index] = 1.0
        slots = [slot]
        total_mass = 1.0

    cn_post = np.zeros((len(slots), phone_dim), dtype=np.float32)
    top_phone_ids = []
    slot_times = []
    slot_confidences = []
    for slot_idx, slot in enumerate(slots):
        slot_mass = sum(slot['counts'].values())
        if slot_mass < total_mass:
            slot['counts'][eps_index] += total_mass - slot_mass
        for phone_idx, mass in slot['counts'].items():
            cn_post[slot_idx, int(phone_idx)] = float(mass) / max(float(total_mass), EPS)
        cn_post[slot_idx] /= np.clip(cn_post[slot_idx].sum(), EPS, None)
        top_phone_ids.append(slot_top_phone(slot['counts'], eps_index))
        slot_times.append(slot_time(slot))
        conf_mass = max(float(slot.get('confidence_mass', 0.0)), EPS)
        slot_confidences.append(float(np.clip(slot.get('confidence_sum', 0.0) / conf_mass, 0.0, 1.0)))

    return {
        'cn_post': cn_post,
        'top_phone_ids': top_phone_ids,
        'slot_times': slot_times,
        'slot_confidences': slot_confidences,
        'hyp_phone_rows': hyp_phone_rows,
        'hyp_weights': hyp_weights.tolist(),
        'id_to_phone': id_to_phone,
        'eps_index': eps_index,
    }


def top_hyp_word_ids_for_slots(pcn):
    slot_word_ids = np.zeros((len(pcn['top_phone_ids']),), dtype=np.int32) - 1
    if not pcn['hyp_phone_rows']:
        return slot_word_ids
    top_hyp = pcn['hyp_phone_rows'][0]
    hyp_phone_ids = list(top_hyp['phone_ids'])
    hyp_phone_to_word = list(top_hyp['phone_to_word'])
    consensus_slot_ids = []
    consensus_phone_ids = []
    for slot_idx, phone_idx in enumerate(pcn['top_phone_ids']):
        if phone_idx is not None:
            consensus_slot_ids.append(slot_idx)
            consensus_phone_ids.append(phone_idx)
    pairs = align_token_sequences(consensus_phone_ids, hyp_phone_ids)
    for ref_pos, hyp_pos in pairs:
        if ref_pos is None or hyp_pos is None:
            continue
        slot_idx = consensus_slot_ids[ref_pos]
        if hyp_pos < len(hyp_phone_to_word):
            slot_word_ids[slot_idx] = int(hyp_phone_to_word[hyp_pos])
    return slot_word_ids


def pcn_stats(cn_post, top_phone_ids, prev_top_phone_ids, eps_index):
    stats = np.zeros((cn_post.shape[0], 5), dtype=np.float32)
    stable_prefix = 0
    for idx, phone_idx in enumerate(top_phone_ids):
        if idx < len(prev_top_phone_ids) and phone_idx == prev_top_phone_ids[idx]:
            stable_prefix += 1
        else:
            break
    prefix_stability = float(stable_prefix) / float(max(len(top_phone_ids), 1))
    for idx, probs in enumerate(cn_post):
        non_eps = np.delete(probs, eps_index)
        top1, margin = top_margin(non_eps)
        stats[idx] = np.array(
            [
                float(probs[eps_index]),
                entropy_np(probs),
                top1,
                margin,
                prefix_stability,
            ],
            dtype=np.float32,
        )
    return stats, prefix_stability


def validate_pcn(pcn):
    cn_post = np.asarray(pcn['cn_post'], dtype=np.float32)
    if not np.isfinite(cn_post).all():
        raise ValueError('pcn_contains_nan_or_inf')
    sums = cn_post.sum(axis=-1)
    if not np.allclose(sums, 1.0, atol=1e-3):
        raise ValueError(f'pcn_posterior_not_normalized:max_abs_err={float(np.max(np.abs(sums - 1.0)))}')
    eps_index = int(pcn['eps_index'])
    if eps_index < 0 or eps_index >= cn_post.shape[-1]:
        raise ValueError('invalid_epsilon_index')
    eps_probs = cn_post[:, eps_index]
    if np.any(eps_probs < -1e-6) or np.any(eps_probs > 1.0 + 1e-6):
        raise ValueError('epsilon_probability_out_of_range')
    for row in pcn.get('hyp_phone_rows', []):
        for value in row.get('phone_confidences', []):
            if not np.isfinite(value) or value < -1e-6 or value > 1.0 + 1e-6:
                raise ValueError('phone_confidence_out_of_range')
    for value in pcn.get('slot_confidences', []):
        if not np.isfinite(value) or value < -1e-6 or value > 1.0 + 1e-6:
            raise ValueError('slot_confidence_out_of_range')


def cumulative_candidate_mask_from_asr(pcn_word_id, word_timestamps, commit_time, audio_end, is_final):
    mask = np.zeros((pcn_word_id.shape[0],), dtype=np.float32)
    committed_words = set()
    for word_idx, word_time in enumerate(word_timestamps or []):
        end_time = float(word_time.get('end', 0.0))
        if end_time <= commit_time + 1e-6 or (is_final and end_time <= audio_end + 1e-6):
            committed_words.add(int(word_idx))
    for slot_idx, word_idx in enumerate(pcn_word_id.tolist()):
        if int(word_idx) in committed_words:
            mask[slot_idx] = 1.0
    return mask


def align_previous_commits(prev_state, current_top_phone_ids):
    if prev_state is None:
        return {}, [], []
    prev_top_phone_ids = prev_state.get('top_phone_ids', [])
    prev_committed = set(int(x) for x in prev_state.get('committed_slots', []))
    pairs = align_token_sequences(prev_top_phone_ids, current_top_phone_ids, mismatch_cost=1.0)
    prev_to_current = {}
    for prev_idx, cur_idx in pairs:
        if prev_idx is None or cur_idx is None:
            continue
        if prev_idx < len(prev_top_phone_ids) and cur_idx < len(current_top_phone_ids):
            if prev_top_phone_ids[prev_idx] == current_top_phone_ids[cur_idx]:
                prev_to_current[int(prev_idx)] = int(cur_idx)
    mapped_old = sorted(prev_to_current[idx] for idx in prev_committed if idx in prev_to_current)
    dropped = sorted(idx for idx in prev_committed if idx not in prev_to_current)
    return prev_to_current, mapped_old, dropped


def _top_word_timestamps(word_timestamps):
    rows = word_timestamps or []
    if rows and isinstance(rows[0], list):
        rows = rows[0]
    return rows if isinstance(rows, list) else []


def _commit_word_metadata(word_timestamps, word_idx):
    rows = _top_word_timestamps(word_timestamps)
    row = rows[word_idx] if 0 <= int(word_idx) < len(rows) and isinstance(rows[word_idx], dict) else {}
    text = normalize_word(str(row.get('word', ''))).casefold()
    try:
        start = float(row.get('start'))
        end = float(row.get('end'))
    except (TypeError, ValueError):
        start = None
        end = None
    return {'text': text, 'start': start, 'end': end}


def _commit_time_iou(left, right):
    if any(item.get(key) is None for item in (left, right) for key in ('start', 'end')):
        return 0.0
    overlap = max(0.0, min(left['end'], right['end']) - max(left['start'], right['start']))
    union = max(left['end'], right['end']) - min(left['start'], right['start'])
    return overlap / union if union > 0 else 0.0


def _word_level_lineage(prev_state, pcn_word_id, current_to_prev, word_timestamps):
    """Build deterministic one-to-one predecessor decisions without using supervision."""
    current_positions = {
        word_idx: np.flatnonzero(pcn_word_id == word_idx)
        for word_idx in sorted(set(int(value) for value in pcn_word_id.tolist() if int(value) >= 0))
    }
    if prev_state is None:
        return current_positions, {
            word_idx: {
                'candidate_prev_word_id': None,
                'selected_prev_word_id': None,
                'predecessor_word_ids': [],
                'vote_counts': {},
                'mapping_status': 'first_seen',
                'split_competitor_count': 0,
                'reliable': True,
                'text_equal': False,
                'time_iou': 0.0,
            }
            for word_idx in current_positions
        }

    prev_word_ids = np.asarray(prev_state.get('pcn_word_id', []), dtype=np.int32)
    prev_word_states = prev_state.get('word_states', {})
    decisions = {}
    competitors_by_prev = {}
    for word_idx, positions in current_positions.items():
        votes = Counter()
        for position in positions.tolist():
            old_slot = int(current_to_prev[position]) if position < len(current_to_prev) else -1
            if 0 <= old_slot < len(prev_word_ids):
                prev_word_idx = int(prev_word_ids[old_slot])
                if prev_word_idx >= 0:
                    votes[prev_word_idx] += 1
        strict = [key for key, count in votes.items() if count > len(positions) / 2.0]
        decision = {
            'candidate_prev_word_id': None,
            'selected_prev_word_id': None,
            'predecessor_word_ids': sorted(int(key) for key in votes),
            'vote_counts': {str(key): int(value) for key, value in sorted(votes.items())},
            'mapping_status': 'unmatched_new' if not votes else 'ambiguous_no_majority',
            'split_competitor_count': 0,
            'reliable': not votes,
            'text_equal': False,
            'time_iou': 0.0,
        }
        if len(strict) == 1:
            prev_word_idx = int(strict[0])
            prev_meta = prev_word_states.get(prev_word_idx, {}).get('metadata', {})
            cur_meta = _commit_word_metadata(word_timestamps, word_idx)
            decision.update({
                'candidate_prev_word_id': prev_word_idx,
                'mapping_status': 'candidate',
                'text_equal': bool(prev_meta.get('text') and cur_meta['text'] and prev_meta['text'] == cur_meta['text']),
                'time_iou': float(_commit_time_iou(prev_meta, cur_meta)),
            })
            competitors_by_prev.setdefault(prev_word_idx, []).append((word_idx, decision))
        decisions[word_idx] = decision

    # A strict-majority edge is unique for each current word. Selecting the
    # highest-weight edge for each predecessor is therefore the global maximum
    # weight one-to-one matching for this candidate graph.
    for prev_word_idx, competitors in sorted(competitors_by_prev.items()):
        for _, decision in competitors:
            decision['split_competitor_count'] = len(competitors)
        ranked = sorted(
            competitors,
            key=lambda item: (
                -int(item[1]['vote_counts'][str(prev_word_idx)]),
                -round(float(item[1]['time_iou']), 12),
                -int(item[1]['text_equal']),
                int(item[0]),
            ),
        )
        winner_key = (
            int(ranked[0][1]['vote_counts'][str(prev_word_idx)]),
            round(float(ranked[0][1]['time_iou']), 12),
            int(ranked[0][1]['text_equal']),
        )
        tied = [
            item for item in ranked
            if (
                int(item[1]['vote_counts'][str(prev_word_idx)]),
                round(float(item[1]['time_iou']), 12),
                int(item[1]['text_equal']),
            ) == winner_key
        ]
        if len(tied) == 1:
            _, winner = ranked[0]
            winner['selected_prev_word_id'] = int(prev_word_idx)
            winner['mapping_status'] = 'accepted'
            winner['reliable'] = True
        else:
            for _, decision in tied:
                decision['mapping_status'] = 'ambiguous_competition_tie'
                decision['reliable'] = False
        for _, loser in ranked[1:]:
            if loser['mapping_status'] == 'candidate':
                loser['mapping_status'] = 'split_new_successor'
                loser['reliable'] = True
    return current_positions, decisions


def build_stateful_commit_masks(
    prev_state,
    pcn,
    pcn_word_id,
    word_timestamps,
    commit_time,
    audio_end,
    is_final,
    current_to_prev_override=None,
    commit_rule='legacy',
):
    if commit_rule not in {'legacy', 'merge_aware'}:
        raise ValueError(f"commit_rule must be 'legacy' or 'merge_aware', got {commit_rule!r}")
    word_timestamps = _top_word_timestamps(word_timestamps)
    candidate = cumulative_candidate_mask_from_asr(pcn_word_id, word_timestamps, commit_time, audio_end, is_final)
    if current_to_prev_override is None:
        prev_to_current, mapped_old_slots, dropped_or_revised = align_previous_commits(prev_state, pcn['top_phone_ids'])
        current_to_prev = np.zeros((len(pcn['top_phone_ids']),), dtype=np.int32) - 1
        for prev_idx, cur_idx in prev_to_current.items():
            if 0 <= cur_idx < current_to_prev.shape[0]:
                current_to_prev[cur_idx] = int(prev_idx)
    else:
        current_to_prev = np.asarray(current_to_prev_override, dtype=np.int32).copy()
        prev_to_current = {
            int(prev_idx): int(cur_idx)
            for cur_idx, prev_idx in enumerate(current_to_prev.tolist())
            if int(prev_idx) >= 0
        }
        prev_committed = set(int(value) for value in (prev_state or {}).get('committed_slots', []))
        mapped_old_slots = sorted(cur for prev, cur in prev_to_current.items() if prev in prev_committed)
        dropped_or_revised = sorted(prev for prev in prev_committed if prev not in prev_to_current)
    mapped_old_set = set(mapped_old_slots)
    cumulative = candidate.copy()
    for slot_idx in mapped_old_set:
        if 0 <= slot_idx < cumulative.shape[0]:
            cumulative[slot_idx] = 1.0

    if commit_rule == 'legacy':
        new_mask = np.zeros_like(cumulative)
        for word_idx in sorted(set(int(value) for value in pcn_word_id.tolist() if int(value) >= 0)):
            positions = np.flatnonzero(pcn_word_id == word_idx)
            if positions.size == 0 or not bool(np.all(candidate[positions] > 0)):
                continue
            if any(int(position) in mapped_old_set for position in positions.tolist()):
                continue
            new_mask[positions] = 1.0
        committed_slots = sorted(np.flatnonzero(cumulative > 0).astype(int).tolist())
        new_slots = sorted(np.flatnonzero(new_mask > 0).astype(int).tolist())
        diagnostics = {
            'mapped_old_slots': mapped_old_slots,
            'new_slots': new_slots,
            'dropped_or_revised_slots': dropped_or_revised,
            'state_event_semantics': 'legacy_slot_any_v1',
        }
        next_state = {
            'top_phone_ids': list(pcn['top_phone_ids']),
            'committed_slots': committed_slots,
        }
        return cumulative.astype(np.float32), new_mask.astype(np.float32), current_to_prev, diagnostics, next_state

    current_positions, lineage = _word_level_lineage(
        prev_state,
        np.asarray(pcn_word_id, dtype=np.int32),
        current_to_prev,
        word_timestamps,
    )
    prev_word_states = (prev_state or {}).get('word_states', {})
    next_logical_id = int((prev_state or {}).get('next_logical_word_id', 0))
    next_event_id = int((prev_state or {}).get('next_state_event_id', 0))
    next_word_states = {}
    word_event_rows = []
    new_mask = np.zeros_like(cumulative)

    for word_idx, positions in sorted(current_positions.items()):
        decision = lineage[word_idx]
        selected_prev = decision['selected_prev_word_id']
        predecessor_ids = decision['predecessor_word_ids']
        predecessor_states = [prev_word_states.get(value, {}) for value in predecessor_ids]
        predecessor_committed = [bool(value.get('committed', False)) for value in predecessor_states]
        completed = bool(np.all(candidate[positions] > 0))
        split = int(decision['split_competitor_count']) > 1
        merge = len(predecessor_ids) > 1
        inherited_state = prev_word_states.get(selected_prev, {}) if selected_prev is not None else {}
        has_old_committed_slot = any(int(position) in mapped_old_set for position in positions.tolist())
        prev_word_ids_array = np.asarray((prev_state or {}).get('pcn_word_id', []), dtype=np.int32)
        prev_committed_slots = set(int(value) for value in (prev_state or {}).get('committed_slots', []))
        unassigned_committed_predecessors = 0
        for position in positions.tolist():
            old_slot = int(current_to_prev[position]) if position < len(current_to_prev) else -1
            if old_slot not in prev_committed_slots:
                continue
            if old_slot >= len(prev_word_ids_array) or int(prev_word_ids_array[old_slot]) < 0:
                unassigned_committed_predecessors += 1
        if unassigned_committed_predecessors:
            decision['mapping_reliable'] = False
            decision['reliable'] = False
            decision['mapping_status'] = 'lineage_broken_unassigned_committed_slot'
        legacy_emit = completed and not has_old_committed_slot
        mixed_merge = bool(
            merge
            and predecessor_committed
            and any(predecessor_committed)
            and not all(predecessor_committed)
        )
        partial_word_completed = bool(
            completed
            and selected_prev is not None
            and not inherited_state.get('committed', False)
            and not merge
            and bool(decision['reliable'])
        )
        # A3-Lite is intentionally minimal: reproduce the legacy mask exactly,
        # then add one event for a completed mixed committed/partial merge or
        # for a reliably inherited partial word that has now become complete.
        emit_event = bool(legacy_emit or (completed and mixed_merge) or partial_word_completed)
        committed = bool(completed or inherited_state.get('committed', False))

        if selected_prev is not None and not (completed and mixed_merge):
            logical_id = int(inherited_state['logical_word_id'])
        else:
            logical_id = next_logical_id
            next_logical_id += 1

        if completed and mixed_merge:
            event_reason = 'mixed_committed_partial_merge_completed'
        elif mixed_merge:
            event_reason = 'mixed_committed_partial_merge_partial'
        elif split:
            event_reason = 'split_completed' if completed else 'split_partial'
        elif merge:
            if completed:
                event_reason = 'merge_completed' if emit_event else 'merge_completed_legacy_suppressed'
            else:
                event_reason = 'merge_partial'
        elif unassigned_committed_predecessors and completed and not emit_event:
            event_reason = 'lineage_broken_completed_legacy_suppressed'
        elif selected_prev is not None and inherited_state.get('committed', False):
            event_reason = 'persistent_committed'
        elif selected_prev is not None:
            event_reason = 'partial_became_completed' if completed else 'partial_persistent'
        elif predecessor_ids:
            event_reason = 'ambiguous_completed' if completed else 'ambiguous_partial'
        else:
            event_reason = 'newly_completed' if completed else 'new_partial'

        state_event_id = -1
        if emit_event:
            new_mask[positions] = 1.0
            state_event_id = next_event_id
            next_event_id += 1
        metadata = _commit_word_metadata(word_timestamps, word_idx)
        next_word_states[int(word_idx)] = {
            'logical_word_id': int(logical_id),
            'committed': bool(committed),
            'state_event_id': int(state_event_id if state_event_id >= 0 else inherited_state.get('state_event_id', -1)),
            'metadata': metadata,
        }
        word_event_rows.append({
            'current_word_id': int(word_idx),
            'logical_word_id': int(logical_id),
            'state_event_id': int(state_event_id),
            'completed': bool(completed),
            'emit_state_event': bool(emit_event),
            'event_reason': event_reason,
            'mapping_status': decision['mapping_status'],
            'mapping_reliable': bool(decision['reliable']),
            'selected_prev_word_id': selected_prev,
            'predecessor_word_ids': predecessor_ids,
            'predecessor_committed': predecessor_committed,
            'vote_counts': decision['vote_counts'],
            'split_competitor_count': int(decision['split_competitor_count']),
            'text': metadata['text'],
            'text_equal': bool(decision['text_equal']),
            'time_iou': float(decision['time_iou']),
            'unassigned_committed_predecessor_slots': int(unassigned_committed_predecessors),
        })

    committed_slots = sorted(np.flatnonzero(cumulative > 0).astype(int).tolist())
    new_slots = sorted(np.flatnonzero(new_mask > 0).astype(int).tolist())
    diagnostics = {
        'mapped_old_slots': mapped_old_slots,
        'new_slots': new_slots,
        'dropped_or_revised_slots': dropped_or_revised,
        'word_level_state_events': word_event_rows,
        'state_event_semantics': 'word_predecessor_one_to_one_v1',
    }
    next_state = {
        'top_phone_ids': list(pcn['top_phone_ids']),
        'pcn_word_id': np.asarray(pcn_word_id, dtype=np.int32).tolist(),
        'committed_slots': committed_slots,
        'word_states': next_word_states,
        'next_logical_word_id': int(next_logical_id),
        'next_state_event_id': int(next_event_id),
    }
    return cumulative.astype(np.float32), new_mask.astype(np.float32), current_to_prev, diagnostics, next_state


def acoustic_distribution_for_frames(frame_probs, phone_to_frame_id, phn_dict, phone_dim):
    out = np.zeros((phone_dim,), dtype=np.float32)
    if frame_probs.size == 0:
        return out
    mean_probs = frame_probs.mean(axis=0)
    for phone, phone_idx in phn_dict.items():
        frame_idx = phone_to_frame_id.get(phone)
        if frame_idx is not None and frame_idx < mean_probs.shape[0]:
            out[int(phone_idx)] = float(mean_probs[frame_idx])
    if out.sum() > 0:
        out /= out.sum()
    return out


def build_acoustic_evidence(item, cn_post, top_phone_ids, phone_to_frame_id, phn_dict, audio_end):
    phone_dim = cn_post.shape[-1]
    acoustic_post = np.zeros_like(cn_post, dtype=np.float32)
    acoustic_stats = np.zeros((cn_post.shape[0], 4), dtype=np.float32)
    kept_indices, kept_probs = select_visible_frames(item['probs'], item['keep_mask'], audio_end, item['frame_step'])
    alignable = [
        (slot_idx, phone_idx)
        for slot_idx, phone_idx in enumerate(top_phone_ids)
        if phone_idx is not None
    ]
    if kept_probs.shape[0] < len(alignable) or not alignable:
        for slot_idx in range(cn_post.shape[0]):
            acoustic_stats[slot_idx, 3] = js_divergence(cn_post[slot_idx], acoustic_post[slot_idx])
        return acoustic_post, acoustic_stats, 0

    id_to_phone = {idx: phone for phone, idx in phn_dict.items()}
    frame_phone_ids = [phone_to_frame_id[id_to_phone[phone_idx]] for _, phone_idx in alignable]
    try:
        path = monotonic_align(-np.log(np.clip(kept_probs, EPS, None)), frame_phone_ids)
    except Exception:
        for slot_idx in range(cn_post.shape[0]):
            acoustic_stats[slot_idx, 3] = js_divergence(cn_post[slot_idx], acoustic_post[slot_idx])
        return acoustic_post, acoustic_stats, 0

    for local_idx, (slot_idx, _) in enumerate(alignable):
        tok_frames = kept_indices[path == local_idx]
        if tok_frames.size == 0:
            acoustic_stats[slot_idx, 3] = js_divergence(cn_post[slot_idx], acoustic_post[slot_idx])
            continue
        dist = acoustic_distribution_for_frames(
            item['probs'][tok_frames],
            phone_to_frame_id,
            phn_dict,
            phone_dim,
        )
        acoustic_post[slot_idx] = dist
        non_eps = dist[:-1]
        _, margin = top_margin(non_eps)
        duration = float(tok_frames.size * item['frame_step'])
        acoustic_stats[slot_idx] = np.array(
            [
                entropy_np(dist),
                margin,
                duration,
                js_divergence(cn_post[slot_idx], dist),
            ],
            dtype=np.float32,
        )
    return acoustic_post, acoustic_stats, int(kept_probs.shape[0])


def compute_prosody(audio, sample_rate, audio_end, word_count, phone_count):
    audio = np.asarray(audio, dtype=np.float32)
    duration = float(max(audio_end, EPS))
    if audio.size == 0:
        return np.zeros((14,), dtype=np.float32)
    hop_length = max(int(0.01 * sample_rate), 1)
    frame_length = max(int(0.025 * sample_rate), hop_length)
    rms = librosa.feature.rms(y=audio, frame_length=frame_length, hop_length=hop_length)[0]
    log_energy = np.log(np.clip(rms, EPS, None))
    energy_threshold = np.percentile(rms, 20) if rms.size else 0.0
    silence = rms <= max(energy_threshold, EPS)
    silence_ratio = float(np.mean(silence)) if silence.size else 0.0
    pause_count, longest_pause = count_pauses(silence, hop_length / sample_rate)
    try:
        f0, voiced_flag, voiced_prob = librosa.pyin(
            audio,
            fmin=50,
            fmax=500,
            sr=sample_rate,
            frame_length=frame_length,
            hop_length=hop_length,
        )
        voiced = np.isfinite(f0)
        f0_values = f0[voiced]
        voiced_prob_mean = float(np.nanmean(voiced_prob)) if voiced_prob is not None else float(np.mean(voiced))
    except Exception:
        f0_values = np.array([], dtype=np.float32)
        voiced = np.zeros_like(rms, dtype=bool)
        voiced_prob_mean = 0.0
    f0_mean = float(np.mean(f0_values)) if f0_values.size else 0.0
    f0_std = float(np.std(f0_values)) if f0_values.size else 0.0
    f0_slope = slope(f0_values)
    energy_mean = float(np.mean(log_energy)) if log_energy.size else 0.0
    energy_std = float(np.std(log_energy)) if log_energy.size else 0.0
    energy_slope = slope(log_energy)
    voiced_duration = max(float(np.mean(voiced)) * duration, EPS) if voiced.size else EPS
    return np.array(
        [
            duration,
            f0_mean,
            f0_std,
            f0_slope,
            voiced_prob_mean,
            energy_mean,
            energy_std,
            energy_slope,
            silence_ratio,
            float(pause_count),
            float(longest_pause),
            float(word_count) / duration,
            float(phone_count) / duration,
            float(phone_count) / voiced_duration,
        ],
        dtype=np.float32,
    )


def slot_prosody_feature_names():
    return [
        'slot_duration',
        'slot_log_energy_mean',
        'slot_log_energy_std',
        'slot_log_energy_max',
        'slot_f0_mean',
        'slot_f0_std',
        'slot_f0_max',
        'slot_voiced_ratio',
        'slot_position_in_word',
        'word_phone_count',
        'energy_relative_to_word_mean',
        'duration_relative_to_word_mean',
        'f0_relative_to_word_mean',
        'is_vowel',
        'lexical_stress_0',
        'lexical_stress_1',
        'lexical_stress_2',
    ]


def compute_slot_prosody(audio, sample_rate, slot_times, pcn_word_id, top_phone_ids, phn_dict):
    feature_names = slot_prosody_feature_names()
    slot_count = len(top_phone_ids)
    out = np.zeros((slot_count, len(feature_names)), dtype=np.float32)
    if slot_count <= 0:
        return out
    audio = np.asarray(audio, dtype=np.float32)
    id_to_phone = {int(idx): phone for phone, idx in phn_dict.items()}
    hop_length = max(int(0.01 * sample_rate), 1)
    frame_length = max(int(0.025 * sample_rate), hop_length)
    if audio.size:
        rms = librosa.feature.rms(y=audio, frame_length=frame_length, hop_length=hop_length)[0]
        log_energy = np.log(np.clip(rms, EPS, None))
        frame_times = librosa.frames_to_time(np.arange(rms.shape[0]), sr=sample_rate, hop_length=hop_length)
    else:
        log_energy = np.zeros((0,), dtype=np.float32)
        frame_times = np.zeros((0,), dtype=np.float32)
    try:
        f0, _, _ = librosa.pyin(
            audio,
            fmin=50,
            fmax=500,
            sr=sample_rate,
            frame_length=frame_length,
            hop_length=hop_length,
        )
    except Exception:
        f0 = np.zeros_like(log_energy, dtype=np.float32) * np.nan
    if f0 is None:
        f0 = np.zeros_like(log_energy, dtype=np.float32) * np.nan
    f0 = np.asarray(f0, dtype=np.float32)
    common_frames = min(frame_times.shape[0], log_energy.shape[0], f0.shape[0])
    frame_times = frame_times[:common_frames]
    log_energy = log_energy[:common_frames]
    f0 = f0[:common_frames]

    durations = np.zeros((slot_count,), dtype=np.float32)
    energy_means = np.zeros((slot_count,), dtype=np.float32)
    f0_means = np.zeros((slot_count,), dtype=np.float32)
    pcn_word_id = np.asarray(pcn_word_id, dtype=np.int32)
    for slot_idx in range(slot_count):
        if slot_idx < len(slot_times) and slot_times[slot_idx] is not None:
            start, end = slot_times[slot_idx]
            start = float(start)
            end = float(max(end, start))
        else:
            start = 0.0
            end = 0.0
        durations[slot_idx] = max(end - start, 0.0)
        frame_mask = (frame_times >= start) & (frame_times <= end) if frame_times.size else np.zeros((0,), dtype=bool)
        energy_values = log_energy[frame_mask]
        f0_values = f0[frame_mask]
        f0_values = f0_values[np.isfinite(f0_values)]
        if energy_values.size:
            out[slot_idx, 1] = float(np.mean(energy_values))
            out[slot_idx, 2] = float(np.std(energy_values))
            out[slot_idx, 3] = float(np.max(energy_values))
            energy_means[slot_idx] = out[slot_idx, 1]
        if f0_values.size:
            out[slot_idx, 4] = float(np.mean(f0_values))
            out[slot_idx, 5] = float(np.std(f0_values))
            out[slot_idx, 6] = float(np.max(f0_values))
            out[slot_idx, 7] = float(f0_values.size / max(frame_mask.sum(), 1))
            f0_means[slot_idx] = out[slot_idx, 4]
        out[slot_idx, 0] = durations[slot_idx]
        word_idx = int(pcn_word_id[slot_idx]) if slot_idx < pcn_word_id.shape[0] else -1
        same_word = np.flatnonzero(pcn_word_id == word_idx) if word_idx >= 0 else np.asarray([slot_idx], dtype=np.int64)
        if same_word.size:
            position = np.flatnonzero(same_word == slot_idx)
            out[slot_idx, 8] = float(position[0] / max(same_word.size - 1, 1)) if position.size else 0.0
            out[slot_idx, 9] = float(same_word.size)
        phone_idx = top_phone_ids[slot_idx] if slot_idx < len(top_phone_ids) else None
        phone = id_to_phone.get(int(phone_idx), '') if phone_idx is not None else ''
        out[slot_idx, 13] = 1.0 if phone[:1].upper() in {'A', 'E', 'I', 'O', 'U'} else 0.0

    for slot_idx in range(slot_count):
        word_idx = int(pcn_word_id[slot_idx]) if slot_idx < pcn_word_id.shape[0] else -1
        same_word = np.flatnonzero(pcn_word_id == word_idx) if word_idx >= 0 else np.asarray([slot_idx], dtype=np.int64)
        if same_word.size:
            word_energy = float(np.mean(energy_means[same_word]))
            word_duration = float(np.mean(durations[same_word]))
            word_f0 = float(np.mean(f0_means[same_word]))
            out[slot_idx, 10] = out[slot_idx, 1] - word_energy
            out[slot_idx, 11] = durations[slot_idx] / max(word_duration, EPS)
            out[slot_idx, 12] = out[slot_idx, 4] - word_f0
    return out.astype(np.float32)


def slope(values):
    values = np.asarray(values, dtype=np.float32)
    if values.size < 2:
        return 0.0
    x = np.arange(values.size, dtype=np.float32)
    x = x - x.mean()
    y = values - values.mean()
    denom = float(np.sum(x * x))
    if denom <= EPS:
        return 0.0
    return float(np.sum(x * y) / denom)


def count_pauses(silence_mask, frame_step):
    pause_count = 0
    longest = 0.0
    cur = 0
    for value in silence_mask.tolist():
        if value:
            cur += 1
        elif cur > 0:
            pause_count += 1
            longest = max(longest, cur * frame_step)
            cur = 0
    if cur > 0:
        pause_count += 1
        longest = max(longest, cur * frame_step)
    return pause_count, longest