diff --git a/app/services/asr/qwen3_vllm.py b/app/services/asr/qwen3_vllm.py index fe18103..a2a265e 100644 --- a/app/services/asr/qwen3_vllm.py +++ b/app/services/asr/qwen3_vllm.py @@ -11,8 +11,8 @@ import queue import re import threading import time +import unicodedata from dataclasses import dataclass, field -from difflib import SequenceMatcher from typing import Any, Optional import librosa @@ -333,65 +333,58 @@ class VLLMRealtimeState: chunk_id: int = 0 text: str = "" raw_decoded: str = "" - window_text: str = "" - # 记录滚动窗口是否已经丢弃旧音频,供文本快照重叠拼接使用。 - audio_window_start_samples: int = 0 window_rolled: bool = False audio_buffer: np.ndarray = field(default_factory=lambda: np.array([], dtype=np.float32)) audio_accum: np.ndarray = field(default_factory=lambda: np.array([], dtype=np.float32)) -def _transcript_char_views(text: str) -> list[tuple[str, int]]: - """Return normalized speech characters with their original string offsets.""" - return [ - (char.lower(), index) - for index, char in enumerate(text) - if char.isalnum() or "\u4e00" <= char <= "\u9fff" - ] +def _normalize_transcript_with_positions(text: str) -> tuple[str, list[int]]: + """忽略标点比较窗口重叠,同时保留回到原文的位置映射。""" + normalized: list[str] = [] + source_positions: list[int] = [] + for index, char in enumerate(text): + category = unicodedata.category(char) + if char.isspace() or category.startswith(("P", "S")): + continue + folded = char.casefold() + normalized.extend(folded) + source_positions.extend([index] * len(folded)) + return "".join(normalized), source_positions -def _merge_rolling_window_text( - assembled_text: str, - previous_window_text: str, - current_window_text: str, -) -> tuple[str, bool]: - """Append only the new tail after matching adjacent partial-window overlap.""" - if not assembled_text: - return current_window_text.strip(), True - previous_views = _transcript_char_views(previous_window_text) - current_views = _transcript_char_views(current_window_text) - if not previous_views or not current_views: - return assembled_text, False +def _merge_overlapping_transcripts(existing: str, incoming: str) -> Optional[str]: + """Port FunASR's conservative suffix-to-prefix rolling-window merge.""" + existing_norm, existing_positions = _normalize_transcript_with_positions(existing) + incoming_norm, incoming_positions = _normalize_transcript_with_positions(incoming) + if not existing_norm or not incoming_norm: + return None - match = SequenceMatcher( - None, - [char for char, _ in previous_views], - [char for char, _ in current_views], - autojunk=False, - ).find_longest_match(0, len(previous_views), 0, len(current_views)) - min_overlap = min(8, len(previous_views), len(current_views)) - max_edge_gap = max(12, int(min(len(previous_views), len(current_views)) * 0.25)) - previous_tail_gap = len(previous_views) - (match.a + match.size) - if ( - match.size < min_overlap - or previous_tail_gap > max_edge_gap - or match.b > max_edge_gap - ): - return assembled_text, False + minimum_overlap = max(4, min(len(existing_norm), len(incoming_norm)) // 4) + for overlap in range(min(len(existing_norm), len(incoming_norm)), minimum_overlap - 1, -1): + if existing_norm[-overlap:] != incoming_norm[:overlap]: + continue + existing_start = len(existing_norm) - overlap + if ( + existing_start > 0 + and existing_positions[existing_start - 1] == existing_positions[existing_start] + ): + continue + if ( + overlap < len(incoming_positions) + and incoming_positions[overlap - 1] == incoming_positions[overlap] + ): + continue - overlap_end = match.b + match.size - delta_start = ( - current_views[overlap_end][1] - if overlap_end < len(current_views) - else len(current_window_text) - ) - delta = current_window_text[delta_start:].lstrip() - if not delta: - return assembled_text, True - - # Chinese text joins directly; add a separator only between adjacent Latin words. - separator = " " if assembled_text[-1].isalnum() and delta[0].isalnum() else "" - return f"{assembled_text.rstrip()}{separator}{delta}", True + suffix_start = incoming_positions[overlap - 1] + 1 + # Keep punctuation from the incoming tail, but avoid duplicating a mark + # already present at the exact join boundary. + while suffix_start < len(incoming): + category = unicodedata.category(incoming[suffix_start]) + if not category.startswith(("P", "S")) or not existing.endswith(incoming[suffix_start]): + break + suffix_start += 1 + return existing + incoming[suffix_start:] + return None class Qwen3VLLMBackend: @@ -738,18 +731,12 @@ class Qwen3VLLMBackend: parsed_language, parsed_text = _parse_asr_output(prefix + generated, state.language or None) state.raw_decoded = prefix + generated if state.window_rolled and parsed_text: - state.text, overlap_matched = _merge_rolling_window_text( - state.text, - state.window_text, - parsed_text, - ) - # Keep the older anchor when a model revision breaks overlap matching; - # the next window still has another chance to find the shared span. - if overlap_matched: - state.window_text = parsed_text + merged_text = _merge_overlapping_transcripts(state.text, parsed_text) + # FunASR only joins a confirmed suffix-prefix overlap; on mismatch it + # starts a fresh partial snapshot instead of guessing an interior match. + state.text = merged_text if merged_text is not None else parsed_text elif not state.window_rolled: state.text = parsed_text - state.window_text = parsed_text state.language = parsed_language or state.language state.chunk_id += 1 return state @@ -766,7 +753,6 @@ class Qwen3VLLMBackend: samples_to_drop = len(state.audio_accum) - max_window_samples # Partial 只保留最近窗口,完整音频仍由 WebSocket 侧保留用于定稿。 state.audio_accum = state.audio_accum[samples_to_drop:] - state.audio_window_start_samples += samples_to_drop state.window_rolled = True state = self._decode_stream(state) return state @@ -780,7 +766,6 @@ class Qwen3VLLMBackend: samples_to_drop = len(state.audio_accum) - max_window_samples # 收尾 partial 仍按窗口处理;最终整段转写使用外层保留的完整音频。 state.audio_accum = state.audio_accum[samples_to_drop:] - state.audio_window_start_samples += samples_to_drop state.window_rolled = True state = self._decode_stream(state) elif state.chunk_id == 0 and len(state.audio_accum) > 0: