139 lines
4.2 KiB
Python
139 lines
4.2 KiB
Python
#!/usr/bin/env python
|
|
# -*- coding: utf-8 -*-
|
|
"""微信语音转文字模块（silk 解码 + faster-whisper ASR）。
|
|
|
|
链路：.silk → silk_v3_decoder.exe 解码为 .pcm → 加 wav 头 → faster-whisper 转文字。
|
|
|
|
用法：
|
|
from voice2text import VoiceToText
|
|
v2t = VoiceToText()
|
|
text = v2t.transcribe_silk("/path/to/xxx.silk") # 返回文字，失败返回 None
|
|
text = v2t.transcribe_voice(username, local_id, db) # 直接从微信库提取并转文字
|
|
|
|
说明：
|
|
- whisper 模型懒加载（首次调用才加载，约几十秒；之后常驻内存）。
|
|
- 中文识别，small 模型在 CPU 上平衡速度与准确率。
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import os
|
|
import subprocess
|
|
import threading
|
|
import wave
|
|
|
|
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
|
|
DECODER_EXE = os.path.join(
|
|
BASE_DIR, "third_party", "silk-v3-decoder", "silk_v3_decoder.exe"
|
|
)
|
|
WHISPER_MODEL = os.environ.get("V2T_MODEL", "small")
|
|
SAMPLE_RATE = 24000 # silk_v3_decoder 默认输出采样率
|
|
|
|
_model = None
|
|
_model_lock = threading.Lock()
|
|
_hf_ready = False
|
|
|
|
|
|
def _ensure_hf_env():
|
|
"""国内环境用 hf-mirror 下载模型，并禁用 xet（避免 401）。"""
|
|
global _hf_ready
|
|
if _hf_ready:
|
|
return
|
|
os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com")
|
|
os.environ.setdefault("HF_HUB_DISABLE_XET", "1")
|
|
_hf_ready = True
|
|
|
|
|
|
def _get_model():
|
|
global _model
|
|
with _model_lock:
|
|
if _model is None:
|
|
_ensure_hf_env()
|
|
from faster_whisper import WhisperModel
|
|
_model = WhisperModel(WHISPER_MODEL, device="cpu", compute_type="int8")
|
|
return _model
|
|
|
|
|
|
def _silk_to_pcm(silk_path: str, pcm_path: str) -> bool:
|
|
if not os.path.exists(DECODER_EXE):
|
|
return False
|
|
try:
|
|
r = subprocess.run(
|
|
[DECODER_EXE, silk_path, pcm_path, "-quiet"],
|
|
capture_output=True, timeout=60,
|
|
)
|
|
return r.returncode == 0 and os.path.exists(pcm_path) and os.path.getsize(pcm_path) > 0
|
|
except Exception:
|
|
return False
|
|
|
|
|
|
def _pcm_to_wav(pcm_path: str, wav_path: str, sample_rate: int = SAMPLE_RATE):
|
|
with open(pcm_path, "rb") as f:
|
|
data = f.read()
|
|
with wave.open(wav_path, "wb") as w:
|
|
w.setnchannels(1)
|
|
w.setsampwidth(2)
|
|
w.setframerate(sample_rate)
|
|
w.writeframes(data)
|
|
|
|
|
|
def transcribe_silk(silk_path: str, sample_rate: int = SAMPLE_RATE):
|
|
"""把 .silk 语音文件转成文字；失败返回 None。"""
|
|
if not silk_path or not os.path.exists(silk_path):
|
|
return None
|
|
pcm_path = silk_path + ".pcm"
|
|
wav_path = silk_path + ".wav"
|
|
try:
|
|
if not _silk_to_pcm(silk_path, pcm_path):
|
|
return None
|
|
_pcm_to_wav(pcm_path, wav_path, sample_rate)
|
|
segments, _info = _get_model().transcribe(
|
|
wav_path, language="zh", beam_size=5,
|
|
)
|
|
text = "".join(s.text for s in segments).strip()
|
|
return text or None
|
|
except Exception:
|
|
return None
|
|
finally:
|
|
for p in (pcm_path, wav_path):
|
|
try:
|
|
if os.path.exists(p):
|
|
os.remove(p)
|
|
except OSError:
|
|
pass
|
|
|
|
|
|
def transcribe_voice(username: str, local_id: int, db, save_dir: str = None):
|
|
"""从微信库提取语音并转文字：db 为 WeChatDB 实例。"""
|
|
try:
|
|
from wechatauto.media import MediaDownloader
|
|
md = MediaDownloader(db, save_dir=save_dir)
|
|
silk = md.download_voice(username, local_id, save_dir=save_dir)
|
|
if not silk:
|
|
return None
|
|
return transcribe_silk(silk)
|
|
except Exception:
|
|
return None
|
|
|
|
|
|
class VoiceToText:
|
|
"""可复用的语音转文字器（模型懒加载、线程安全）。"""
|
|
|
|
def __init__(self, model: str = None):
|
|
global WHISPER_MODEL
|
|
if model:
|
|
WHISPER_MODEL = model
|
|
|
|
def transcribe_silk(self, silk_path: str):
|
|
return transcribe_silk(silk_path)
|
|
|
|
def transcribe_voice(self, username: str, local_id: int, db, save_dir: str = None):
|
|
return transcribe_voice(username, local_id, db, save_dir)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
import sys
|
|
if len(sys.argv) < 2:
|
|
print("用法: python voice2text.py <xxx.silk>")
|
|
sys.exit(1)
|
|
print(transcribe_silk(sys.argv[1]))
|