app.py

import os
os.environ["KOKORO_NO_ESPEAK_CHECK"] = "1"

import torch, numpy as np, gradio as gr
from kokoro import KPipeline

device = "cuda" if torch.cuda.is_available() else "cpu"
pipeline = KPipeline(lang_code="a", device=device)

def synthesize(text, speed, voice):
    gs, ps, audio = next(pipeline(text, voice=voice, speed=speed))
    # ensure `audio` is numpy (KPipeline already returns numpy in HF demo)
    return (24000, np.asarray(audio))

demo = gr.Interface(
    fn=synthesize,
    inputs=[
        gr.Textbox(lines=3, label="Text"),
        gr.Slider(0.5, 2.0, value=1.0, step=0.1, label="Speed"),
        gr.Dropdown(
            ["af_heart", "af_bella", "af_sarah", "bf_emma", "bm_george"],
            value="af_heart",
            label="Voice"
        )
    ],
    outputs=gr.Audio(type="numpy", label="Speech Output")
)
demo.queue().launch(server_name="0.0.0.0", server_port=7860)