handler.py

from bark import generate_audio, preload_models
import torch
import numpy as np
import scipy
import io
import base64

class EndpointHandler:
    def __init__(self, path=""):
        preload_models()

    def __call__(self, data):
        text = data.get("inputs", "")
        if not text.strip():
            return {"error": "No input text provided."}

        audio_array = generate_audio(text, history_prompt="v2/ru_speaker_0")

        byte_io = io.BytesIO()
        scipy.io.wavfile.write(byte_io, rate=24000, data=audio_array)
        byte_io.seek(0)

        audio_base64 = base64.b64encode(byte_io.read()).decode("utf-8")

        return {
            "audio_base64": audio_base64,
            "content_type": "audio/wav"
        }