app.py

#!/usr/bin/env python3
"""Gradio Space for Supertonic 3 MLX."""

from __future__ import annotations

import os
import time
from functools import lru_cache

import gradio as gr
import numpy as np
from huggingface_hub import snapshot_download

from supertonic_mlx import SupertonicMLX

os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")

MODEL_ID = os.environ.get("SUPERTONIC_MLX_MODEL", "mlx-community/supertonic-3")
LANGUAGES = [
    "en", "ko", "ja", "ar", "bg", "cs", "da", "de", "el", "es", "et", "fi",
    "fr", "hi", "hr", "hu", "id", "it", "lt", "lv", "nl", "pl", "pt", "ro",
    "ru", "sk", "sl", "sv", "tr", "uk", "vi", "na",
]
VOICES = ["M1", "M2", "M3", "M4", "M5", "F1", "F2", "F3", "F4", "F5"]


@lru_cache(maxsize=1)
def load_model() -> SupertonicMLX:
    model_dir = snapshot_download(
        repo_id=MODEL_ID,
        repo_type="model",
        allow_patterns=[
            "graphs/*.json",
            "weights/*.npz",
            "voice_styles/*.json",
            "tts.json",
            "unicode_indexer.json",
            "mlx_manifest.json",
        ],
    )
    return SupertonicMLX.from_pretrained(model_dir)


def synthesize(text: str, lang: str, voice: str, total_step: int, speed: float, seed: int):
    if not text or not text.strip():
        return None, "Enter text to synthesize."
    try:
        start = time.perf_counter()
        was_loaded = load_model.cache_info().currsize > 0
        load_start = time.perf_counter()
        tts = load_model()
        load_elapsed = time.perf_counter() - load_start
        style = tts.get_voice_style(voice)
        infer_start = time.perf_counter()
        wav, duration = tts.synthesize(
            text.strip(),
            lang,
            style,
            total_step=int(total_step),
            speed=float(speed),
            seed=int(seed),
        )
        infer_elapsed = time.perf_counter() - infer_start
        samples = int(tts.sample_rate * float(duration[0]))
        audio = np.clip(wav[0, :samples], -1.0, 1.0)
        audio_i16 = (audio * 32767).astype(np.int16)
        audio_duration = samples / tts.sample_rate
        total_elapsed = time.perf_counter() - start
        load_text = "cached" if was_loaded else f"{load_elapsed:.2f}s"
        return (
            (tts.sample_rate, audio_i16),
            (
                f"Done. elapsed={total_elapsed:.2f}s, "
                f"model_load={load_text}, synth={infer_elapsed:.2f}s, "
                f"audio={audio_duration:.2f}s"
            ),
        )
    except Exception as exc:
        return None, f"Error: {type(exc).__name__}: {exc}"


with gr.Blocks(title="Supertonic 3 MLX", theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        """
# Supertonic 3 MLX

Lightning-fast multilingual text-to-speech using the community MLX conversion.
"""
    )
    with gr.Row():
        with gr.Column(scale=1):
            text = gr.Textbox(
                label="Text",
                lines=5,
                value="Supertonic 3 is running through an MLX graph runtime.",
            )
            with gr.Row():
                lang = gr.Dropdown(LANGUAGES, value="en", label="Language")
                voice = gr.Dropdown(VOICES, value="M1", label="Voice")
            with gr.Accordion("Generation Settings", open=False):
                total_step = gr.Slider(5, 8, value=5, step=1, label="Total Steps")
                speed = gr.Slider(0.7, 2.0, value=1.05, step=0.05, label="Speed")
                seed = gr.Number(value=0, precision=0, label="Seed")
            btn = gr.Button("Generate", variant="primary")
        with gr.Column(scale=1):
            audio = gr.Audio(label="Output", type="numpy")
            status = gr.Textbox(label="Status")
    btn.click(synthesize, [text, lang, voice, total_step, speed, seed], [audio, status])
    gr.Examples(
        examples=[
            ["Supertonic 3 is running on CPU fallback.", "en", "M1", 5, 1.05, 0],
            ["こんにちは。MLX 変換をテストしています。", "ja", "F1", 5, 1.05, 1],
            ["오늘은 MLX 변환을 테스트하고 있습니다.", "ko", "F2", 5, 1.05, 2],
        ],
        inputs=[text, lang, voice, total_step, speed, seed],
    )


if __name__ == "__main__":
    demo.queue(max_size=8).launch()