modal/app.py

import modal
import io


image = modal.Image.debian_slim(python_version="3.12").apt_install("espeak-ng")
image = image.pip_install(
    "kokoro>=0.9.4", "soundfile", "fastapi[standard]", "spacy==3.8.0"
)
image = image.run_commands("python -m spacy download en_core_web_sm")
app = modal.App("kokoro-api", image=image)

with image.imports():
    import os
    from kokoro import KPipeline
    from IPython.display import display, Audio
    import soundfile as sf
    import torch
    from fastapi.responses import StreamingResponse, Response, FileResponse
    import numpy as np
    import uuid


@app.cls(gpu="t4", scaledown_window=60 * 2, enable_memory_snapshot=True)
@modal.concurrent(max_inputs=30)
class kokoro:
    @modal.enter()
    def load(self):
        self.pipeline = KPipeline(lang_code="a")

    @modal.fastapi_endpoint(docs=True, method="POST")
    def generate(self, text: str, voice: str = "af_heart"):

        if len(text) == 0:
            return Response(content="Text is empty", status_code=400)

        generator = self.pipeline(text, voice)
        audio_bytes = io.BytesIO()
        random_name = str(uuid.uuid4())
        file_path = f"{random_name}.wav"
        audio_combined = []

        for _, _, audio in generator:
            audio_combined.append(audio)

        audio_combined = np.concatenate(audio_combined)

        sf.write(file_path, audio_combined, 24000)

        return FileResponse(path=file_path, media_type="audio/wav", filename=file_path)
    
    @modal.fastapi_endpoint(docs=True, method="GET")
    def wake_up(self):
        return Response(content="Kokoro is awake", status_code=200)