zikai
00e2a95fb7
- 音频/视频转双语(英/中)SRT 字幕，Docker 容器化，CPU 开发/GPU 生产同一份代码 - faster-whisper ASR(词级时间戳) + 断句时间戳重算 + NLLB 翻译(模型不共驻) - 分片上传(断点续传) + SQLite 持久化 + 主页/历史/日志页面 - 历史页文件名搜索；缓存定时清理(默认保留7天，可配置) - 双 Dockerfile(cpu/gpu) + setup/start/stop 脚本
66 lines
2.0 KiB
Python
66 lines
2.0 KiB
Python
"""语音识别服务：faster-whisper，输出带词级时间戳的 segments。
|
||
|
||
CPU dev: tiny.en + int8；GPU prod: large-v3-turbo + float16。同一份代码。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import logging
|
||
from pathlib import Path
|
||
|
||
from ..config import get_settings
|
||
from .model_manager import get_model_manager
|
||
from .types import Segment, Word
|
||
|
||
logger = logging.getLogger("audio2text.asr")
|
||
|
||
|
||
def transcribe(wav_path: Path) -> list[Segment]:
|
||
"""转写 wav，返回 segments（含词级时间戳）。
|
||
|
||
Args:
|
||
wav_path: 16kHz mono PCM wav
|
||
|
||
Returns:
|
||
list[Segment]，每个 Segment 带词级 words（若 word_timestamps 启用）。
|
||
"""
|
||
s = get_settings().asr
|
||
if not wav_path.is_file():
|
||
raise FileNotFoundError(f"音频不存在：{wav_path}")
|
||
|
||
model = get_model_manager().get_asr()
|
||
logger.debug("开始转写 %s（model=%s language=%s）", wav_path.name, s.model, s.language)
|
||
|
||
segments_gen, info = model.transcribe(
|
||
str(wav_path),
|
||
language=s.language,
|
||
word_timestamps=s.word_timestamps,
|
||
vad_filter=s.vad_filter,
|
||
beam_size=5,
|
||
)
|
||
logger.debug(
|
||
"音频时长 %.1fs，检测语言=%s（置信度 %.2f）",
|
||
info.duration, info.language, info.language_probability,
|
||
)
|
||
|
||
segments: list[Segment] = []
|
||
for seg in segments_gen:
|
||
words: list[Word] = []
|
||
if s.word_timestamps and getattr(seg, "words", None):
|
||
for w in seg.words:
|
||
words.append(Word(
|
||
text=w.word.strip(),
|
||
start=float(w.start),
|
||
end=float(w.end),
|
||
probability=float(getattr(w, "probability", 1.0)),
|
||
))
|
||
segments.append(Segment(
|
||
text=seg.text.strip(),
|
||
start=float(seg.start),
|
||
end=float(seg.end),
|
||
words=words,
|
||
))
|
||
logger.debug("转写完成：%d 段，%d 词。",
|
||
len(segments), sum(len(s.words) for s in segments))
|
||
return segments
|