Spaces:
Running
Running
| from __future__ import annotations | |
| import spaces # imported first, as ZeroGPU requires | |
| import tempfile | |
| import uuid | |
| from pathlib import Path | |
| import gradio as gr | |
| from pyharp import ModelCard, build_endpoint | |
| from pyharp.tags import Channels, SampleRate, Subcategory | |
| from pocket_runtime import load_model, synthesize | |
| MAX_TEXT_LENGTH = 500 | |
| OUTPUT_ROOT = Path(tempfile.gettempdir()) / "pocket_tts_outputs" | |
| model_card = ModelCard( | |
| name="Pocket TTS", | |
| description=( | |
| "Generate English speech with Kyutai's compact Pocket TTS model using " | |
| "CPU inference and predefined voices, as 24 kHz mono audio. The code is MIT, but the " | |
| "model weights are CC-BY-4.0, and their model card prohibits uses such as cloning or " | |
| "impersonating a voice without consent, deception, and harmful content. Repository: " | |
| "https://github.com/kyutai-labs/pocket-tts Paper: " | |
| "https://arxiv.org/abs/2509.06926" | |
| ), | |
| author="Simon Rouard, Manu Orsini, Axel Roebel, et al.", | |
| tags=[Subcategory.TEXT_TO_SPEECH, SampleRate(24000), Channels(1), 'streaming-tts', 'domain:speech'], | |
| ) | |
| def process_fn(text: str | None, voice: str) -> str: | |
| # Checked before the GPU is reserved, so a request the model cannot take spends none of the | |
| # caller's ZeroGPU allowance | |
| clean_text = (text or "").strip() | |
| if not clean_text: | |
| raise gr.Error("Please enter text to synthesize.") | |
| if len(clean_text) > MAX_TEXT_LENGTH: | |
| raise gr.Error( | |
| f"Text must be no longer than {MAX_TEXT_LENGTH} characters. " | |
| f"Received {len(clean_text)} characters." | |
| ) | |
| return speak(clean_text, voice) | |
| def speak(clean_text: str, voice: str) -> str: | |
| output_dir = OUTPUT_ROOT / uuid.uuid4().hex | |
| output_dir.mkdir(parents=True, exist_ok=True) | |
| output_path = output_dir / "pocket_tts_output.wav" | |
| try: | |
| synthesize(clean_text, voice, output_path) | |
| except Exception as exc: | |
| raise gr.Error(f"Pocket TTS synthesis failed: {exc}") from exc | |
| return str(output_path) | |
| # pyharp runs process_fn in a worker, which imports this file as __mp_main__ and starts with the | |
| # server, so the model is loaded there, before the first request rather than during it. | |
| if __name__ != "__main__": | |
| load_model() | |
| if __name__ == "__main__": | |
| with gr.Blocks(title="Pocket TTS") as demo: | |
| input_components = [ | |
| gr.Textbox( | |
| label="English Text", | |
| lines=4, | |
| max_lines=8, | |
| placeholder="Enter the sentence you want to hear...", | |
| ) | |
| .harp_required(True) | |
| .set_info("English text, up to 500 characters."), | |
| gr.Dropdown( | |
| choices=[ | |
| "alba", | |
| "anna", | |
| "bill_boerst", | |
| "eve", | |
| "george", | |
| "jane", | |
| "mary", | |
| "michael", | |
| "paul", | |
| "stuart_bell", | |
| "vera", | |
| ], | |
| value="alba", | |
| label="Voice", | |
| ).set_info("A predefined English voice supplied by the model authors."), | |
| ] | |
| output_components = [ | |
| gr.Audio( | |
| type="filepath", | |
| label="Synthesized Speech", | |
| ).set_info("24 kHz mono WAV generated on CPU."), | |
| ] | |
| build_endpoint( | |
| model_card=model_card, | |
| input_components=input_components, | |
| output_components=output_components, | |
| process_fn=process_fn, | |
| ) | |
| demo.queue(default_concurrency_limit=1).launch(show_error=True, pwa=True) | |