utils/tts.py

import time
import yaml
from pathlib import Path
import torch
from fish_audio.sdk import TextToSpeech, Vocoder
from pydub import AudioSegment

# Load config
config_path = Path(__file__).parent.parent / "config" / "tts_config.yaml"
with open(config_path) as f:
    config = yaml.safe_load(f)

# Initialize models
tts_model = TextToSpeech(
    model_name=config["tts"]["model"],
    device=config["tts"]["device"],
    precision=config["tts"]["precision"],
)

vocoder = Vocoder(
    model_name=config["tts"]["vocoder"],
    device=tts_model.device,
)

def generate_speech(text: str, language: str = "zh") -> str:
    """Generate speech from text using Fish Audio SDK"""
    # Format text with language tags
    lang_template = config["generation"]["language_mapping"][language]
    processed_text = lang_template.format(text=text)
    
    # Generate mel spectrogram
    mel = tts_model.generate(
        text=processed_text,
        temperature=config["generation"]["temperature"],
        top_k=config["generation"]["top_k"],
        max_length=config["generation"]["max_length"],
    )
    
    # Convert mel to waveform
    waveform = vocoder(mel)
    
    # Create audio segment
    audio = AudioSegment(
        waveform.numpy().tobytes(),
        frame_rate=vocoder.sample_rate,
        sample_width=2,
        channels=1,
    )
    
    # Save output
    output_path = f"temp/outputs/output_{int(time.time())}.wav"
    audio.export(output_path, format="wav")
    
    return output_path