File size: 2,313 Bytes
a960dea | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | import base64
import io
import sys
import tempfile
from pathlib import Path
import torch
import torchaudio
REPO_DIR = Path(__file__).resolve().parent
sys.path.insert(0, str(REPO_DIR))
sys.path.insert(0, str(REPO_DIR / "third_party" / "Matcha-TTS"))
from accent_config import ACCENT_INSTRUCTIONS # noqa: E402
class EndpointHandler:
def __init__(self, path=""):
self.model_dir = Path(path or __file__).resolve()
if self.model_dir.is_file():
self.model_dir = self.model_dir.parent
from cosyvoice.cli.cosyvoice import AutoModel
self.model = AutoModel(
model_dir=str(self.model_dir),
fp16=torch.cuda.is_available(),
load_vllm=False,
)
def __call__(self, data):
text = data.get("inputs", "")
parameters = data.get("parameters", {})
accent = parameters.get("accent", "singapore")
if not text:
raise ValueError("inputs must contain Chinese text")
if accent not in ACCENT_INSTRUCTIONS:
raise ValueError(f"accent must be one of: {', '.join(ACCENT_INSTRUCTIONS)}")
prompt_path = self.model_dir / "zero_shot_prompt.wav"
prompt_audio = parameters.get("prompt_audio_base64")
temp_path = None
if prompt_audio:
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as handle:
handle.write(base64.b64decode(prompt_audio))
temp_path = Path(handle.name)
prompt_path = temp_path
try:
chunks = [item["tts_speech"] for item in self.model.inference_instruct2(
text,
ACCENT_INSTRUCTIONS[accent],
str(prompt_path),
stream=False,
speed=float(parameters.get("speed", 1.0)),
)]
speech = torch.cat(chunks, dim=1).cpu()
buffer = io.BytesIO()
torchaudio.save(buffer, speech, self.model.sample_rate, format="wav")
return {
"audio_base64": base64.b64encode(buffer.getvalue()).decode("ascii"),
"sample_rate": self.model.sample_rate,
"accent": accent,
}
finally:
if temp_path is not None:
temp_path.unlink(missing_ok=True)
|