Spaces:
Sleeping
Sleeping
| import spaces | |
| # isort: split | |
| from collections.abc import Iterator | |
| from itertools import batched | |
| import gradio as gr | |
| import torch | |
| from silero_vad import load_silero_vad | |
| from src.environment import load_hf_token | |
| from src.shared.transcription import transcribe | |
| from src.space.runtime import SpaceConfig, load_space_components, read_recording | |
| def create_app() -> gr.Blocks: | |
| # --------------------------------------------------------- | |
| # Place model weights on CUDA during startup for ZeroGPU sharing. | |
| # --------------------------------------------------------- | |
| config = SpaceConfig() | |
| components = load_space_components(load_hf_token()) | |
| device = torch.device("cuda") | |
| components.model.to(device) | |
| vad_model = load_silero_vad(onnx=True) | |
| def generate(audio: list[bytes]) -> list[str]: | |
| results = transcribe(components.model, components.feature_extractor, components.tokenizer, config, device, audio) | |
| return [result.text for result in results] | |
| def transcribe_recording(audio_path: str | None) -> Iterator[str]: | |
| segments = read_recording(audio_path, vad_model) | |
| if not segments: | |
| yield "音声内に発話が見つかりませんでした。" | |
| return | |
| texts: list[str] = [] | |
| for batch in batched(segments, 4): | |
| texts.extend(generate(list(batch))) | |
| yield "".join(texts).replace("\r", "").replace("\n", "") | |
| # --------------------------------------------------------- | |
| # Provide one audio input and one transcription action. | |
| # --------------------------------------------------------- | |
| with gr.Blocks(title="LambdaAudio v6", delete_cache=(3600, 3600)) as app: | |
| gr.Markdown("# LambdaAudio v6\n日本語の音声をアップロードするか、マイクで録音してください。\n\n[モデルの詳細](https://huggingface.co/KeisukeMiyamoto/lambda-audio-v6)") | |
| audio = gr.Audio(sources=["microphone", "upload"], type="filepath", format="wav", label="音声", editable=False, elem_id="recording-input") | |
| button = gr.Button("文字起こし", variant="primary", elem_id="transcribe-button") | |
| submit = gr.Button(visible="hidden", elem_id="transcribe-submit") | |
| output = gr.Textbox(label="文字起こし結果", lines=5, interactive=False, buttons=["copy"]) | |
| submit.click(transcribe_recording, inputs=audio, outputs=output, api_name="transcribe", concurrency_limit=1) | |
| # --------------------------------------------------------- | |
| # Stop the browser recorder before submitting its uploaded audio. | |
| # The stop_recording event fires after Gradio finishes uploading. | |
| # --------------------------------------------------------- | |
| button.click( | |
| fn=None, | |
| queue=False, | |
| js="""() => { | |
| const audio = document.getElementById('recording-input'); | |
| if (audio.dataset.submitAfterRecording) return; | |
| const stop = [...audio.querySelectorAll('.stop-button, .stop-button-paused')] | |
| .find(button => button.checkVisibility()); | |
| if (stop) { | |
| audio.dataset.submitAfterRecording = 'true'; | |
| stop.click(); | |
| return; | |
| } | |
| document.getElementById('transcribe-submit').click(); | |
| }""", | |
| ) | |
| audio.stop_recording( | |
| fn=None, | |
| queue=False, | |
| js="""() => { | |
| const audio = document.getElementById('recording-input'); | |
| if (!audio.dataset.submitAfterRecording) return; | |
| delete audio.dataset.submitAfterRecording; | |
| document.getElementById('transcribe-submit').click(); | |
| }""", | |
| ) | |
| return app.queue(max_size=10) | |
| if __name__ == "__main__": | |
| create_app().launch(max_file_size="200mb") | |