Billy D.
12987c6adc
[ray-serve only] - Fix whitespace in docstrings - Add strict=True to zip() calls - Use ternary operators where appropriate - Rename unused loop variables
125 lines
3.4 KiB
Python
125 lines
3.4 KiB
Python
"""
|
|
Ray Serve deployment for Coqui TTS.
|
|
Runs on: elminster (RTX 2070 8GB, CUDA)
|
|
"""
|
|
|
|
import base64
|
|
import io
|
|
import os
|
|
from typing import Any
|
|
|
|
from ray import serve
|
|
|
|
|
|
@serve.deployment(name="TTSDeployment", num_replicas=1)
|
|
class TTSDeployment:
|
|
def __init__(self):
|
|
import torch
|
|
from TTS.api import TTS
|
|
|
|
self.model_name = os.environ.get("MODEL_NAME", "tts_models/en/ljspeech/tacotron2-DDC")
|
|
|
|
# Detect device
|
|
self.use_gpu = torch.cuda.is_available()
|
|
|
|
print(f"Loading TTS model: {self.model_name}")
|
|
print(f"Using GPU: {self.use_gpu}")
|
|
|
|
self.tts = TTS(model_name=self.model_name, progress_bar=False)
|
|
|
|
if self.use_gpu:
|
|
self.tts = self.tts.to("cuda")
|
|
|
|
print("TTS model loaded successfully")
|
|
|
|
async def __call__(self, request: dict[str, Any]) -> dict[str, Any]:
|
|
"""
|
|
Handle text-to-speech requests.
|
|
|
|
Expected request format:
|
|
{
|
|
"text": "Text to synthesize",
|
|
"speaker": "speaker_name",
|
|
"language": "en",
|
|
"speed": 1.0,
|
|
"output_format": "wav",
|
|
"return_base64": true
|
|
}
|
|
"""
|
|
import numpy as np
|
|
from scipy.io import wavfile
|
|
|
|
text = request.get("text", "")
|
|
speaker = request.get("speaker")
|
|
language = request.get("language")
|
|
speed = request.get("speed", 1.0)
|
|
output_format = request.get("output_format", "wav")
|
|
return_base64 = request.get("return_base64", True)
|
|
|
|
if not text:
|
|
return {"error": "No text provided"}
|
|
|
|
# Generate speech
|
|
try:
|
|
# TTS.tts returns a numpy array of audio samples
|
|
wav = self.tts.tts(
|
|
text=text,
|
|
speaker=speaker,
|
|
language=language,
|
|
speed=speed,
|
|
)
|
|
|
|
# Convert to numpy array if needed
|
|
if not isinstance(wav, np.ndarray):
|
|
wav = np.array(wav)
|
|
|
|
# Normalize to int16
|
|
wav_int16 = (wav * 32767).astype(np.int16)
|
|
|
|
# Get sample rate from model config
|
|
sample_rate = (
|
|
self.tts.synthesizer.output_sample_rate
|
|
if hasattr(self.tts, "synthesizer")
|
|
else 22050
|
|
)
|
|
|
|
# Write to buffer
|
|
buffer = io.BytesIO()
|
|
wavfile.write(buffer, sample_rate, wav_int16)
|
|
audio_bytes = buffer.getvalue()
|
|
|
|
response = {
|
|
"model": self.model_name,
|
|
"sample_rate": sample_rate,
|
|
"duration": len(wav) / sample_rate,
|
|
"format": output_format,
|
|
}
|
|
|
|
if return_base64:
|
|
response["audio"] = base64.b64encode(audio_bytes).decode("utf-8")
|
|
else:
|
|
response["audio_bytes"] = audio_bytes
|
|
|
|
return response
|
|
|
|
except Exception as e:
|
|
return {
|
|
"error": str(e),
|
|
"model": self.model_name,
|
|
}
|
|
|
|
def list_speakers(self) -> dict[str, Any]:
|
|
"""List available speakers for multi-speaker models."""
|
|
speakers = []
|
|
if hasattr(self.tts, "speakers") and self.tts.speakers:
|
|
speakers = self.tts.speakers
|
|
|
|
return {
|
|
"model": self.model_name,
|
|
"speakers": speakers,
|
|
"is_multi_speaker": len(speakers) > 0,
|
|
}
|
|
|
|
|
|
app = TTSDeployment.bind()
|