105 lines
3.4 KiB
Python
105 lines
3.4 KiB
Python
"""Audio tool operations: TTS (edge-tts) and ASR (faster-whisper)."""
|
|
import asyncio
|
|
import os
|
|
import re
|
|
|
|
FFMPEG_BIN = "ffmpeg"
|
|
MAX_TTS_CHARS = 2000
|
|
MAX_ASR_MB = 30
|
|
|
|
VOICE_FILES = {
|
|
"zh_CN-huayan-medium": "晓妍(女·标准中文)",
|
|
"en_US-amy-medium": "Amy(女·英语)",
|
|
}
|
|
|
|
_asr_model = None
|
|
|
|
|
|
def _get_asr():
|
|
global _asr_model
|
|
if _asr_model is None:
|
|
from faster_whisper import WhisperModel
|
|
_asr_model = WhisperModel("small", device="cpu", compute_type="int8")
|
|
return _asr_model
|
|
|
|
|
|
PIPER_DIR = os.environ.get("PIPER_DIR", "/app/piper-models")
|
|
_piper_voices = {}
|
|
|
|
|
|
def _load_voice(voice: str):
|
|
from piper import PiperVoice
|
|
if voice not in _piper_voices:
|
|
model = os.path.join(PIPER_DIR, voice + ".onnx")
|
|
if not os.path.exists(model):
|
|
raise ValueError("音色不存在: " + voice)
|
|
_piper_voices[voice] = PiperVoice.load(model)
|
|
return _piper_voices[voice]
|
|
|
|
|
|
def run_tts(text: str, voice: str, rate_pct: int, pitch_hz: int, out_path: str):
|
|
import subprocess
|
|
import tempfile
|
|
import wave
|
|
text = (text or "").strip()
|
|
if not text:
|
|
raise ValueError("请输入文字内容")
|
|
if len(text) > MAX_TTS_CHARS:
|
|
raise ValueError(f"文字过长(最多 {MAX_TTS_CHARS} 字)")
|
|
if voice not in VOICE_FILES:
|
|
voice = "zh_CN-huayan-medium"
|
|
rate_pct = max(-50, min(100, int(rate_pct)))
|
|
v = _load_voice(voice)
|
|
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tf:
|
|
wav_path = tf.name
|
|
try:
|
|
with open(wav_path, "wb") as f:
|
|
with wave.open(f, "wb") as wf:
|
|
v.synthesize_wav(text, wf)
|
|
# post-process: speed (atempo) + pitch-free, output mp3
|
|
tempo = 1 + rate_pct / 100.0
|
|
tempo = max(0.5, min(2.0, tempo))
|
|
argv = [FFMPEG_BIN, "-y", "-i", wav_path, "-filter:a", f"atempo={tempo:.4f}",
|
|
"-c:a", "libmp3lame", "-b:a", "192k", out_path]
|
|
env = dict(os.environ)
|
|
env.pop("HTTPS_PROXY", None)
|
|
r = subprocess.run(argv, capture_output=True, timeout=120, env=env)
|
|
if r.returncode != 0:
|
|
# fallback: plain wav->mp3
|
|
r = subprocess.run([FFMPEG_BIN, "-y", "-i", wav_path, "-c:a", "libmp3lame",
|
|
"-b:a", "192k", out_path], capture_output=True, timeout=120, env=env)
|
|
if r.returncode != 0:
|
|
raise RuntimeError("音频转码失败")
|
|
finally:
|
|
try:
|
|
os.remove(wav_path)
|
|
except Exception:
|
|
pass
|
|
if not os.path.exists(out_path) or os.path.getsize(out_path) < 100:
|
|
raise RuntimeError("TTS 生成失败")
|
|
|
|
|
|
def run_asr(audio_path: str, language: str):
|
|
model = _get_asr()
|
|
language = language if language in ("zh", "en", "ja", "yue") else None
|
|
segments, info = model.transcribe(audio_path, language=language, beam_size=2, vad_filter=True)
|
|
out = []
|
|
texts = []
|
|
for s in segments:
|
|
item = {"start": round(s.start, 1), "end": round(s.end, 1), "text": s.text.strip()}
|
|
out.append(item)
|
|
texts.append(s.text.strip())
|
|
return {
|
|
"text": "\n".join(texts),
|
|
"segments": out[:200],
|
|
"language": info.language,
|
|
"probability": round(info.language_probability, 2),
|
|
}
|
|
|
|
|
|
# voices served by local piper models
|
|
VOICES = [
|
|
{"value": k, "label": v, "lang": k.split("_")[0]}
|
|
for k, v in VOICE_FILES.items()
|
|
]
|