handler.py

import torch
from fish_speech.models.fish_speech import FishSpeech
from fish_speech.inference import infer
import io
import base64
import soundfile as sf

# 加载模型
model = FishSpeech.from_pretrained('fishaudio/fish-speech-1.5')

def predict(inputs: dict):
    text = inputs.get('inputs', 'Hello world')
    # 支持 [singing] 标签
    if "[singing]" in text.lower():
        mode = "singing"
        text = text.replace("[singing]", "")
    else:
        mode = "speech"

    # 生成音频
    audio = infer(model, text, mode=mode)

    # 转 base64 WAV
    buffer = io.BytesIO()
    sf.write(buffer, audio.cpu().numpy(), 24000, format='WAV')
    audio_b64 = base64.b64encode(buffer.getvalue()).decode()

    return {"audio": audio_b64}

def query(payload):
    return predict(payload)