| import math |
| from pathlib import Path |
| import struct |
| import wave |
|
|
| import gradio as gr |
|
|
|
|
| DEFAULT_MAX_SECONDS = 300 |
| GROUNDING = "Temporal grounding (when?)" |
| DETECTION = "Event detection (does it occur?)" |
|
|
|
|
| def create_fake_examples() -> list[list[str]]: |
| """Create short synthetic clips so the UI examples work without model data.""" |
| example_dir = Path("/tmp/spotsound_fake_examples") |
| example_dir.mkdir(parents=True, exist_ok=True) |
| specs = [ |
| ("fake_hair_dryer.wav", "hair dryer drying", (900, 1250)), |
| ("fake_ambulance_siren.wav", "ambulance siren", (520, 690)), |
| ("fake_train_horn.wav", "train horning", (180, 240)), |
| ("fake_police_siren.wav", "police car siren", (600, 850)), |
| ] |
|
|
| examples: list[list[str]] = [] |
| sample_rate = 16_000 |
| duration = 2 |
| for filename, query, frequencies in specs: |
| path = example_dir / filename |
| with wave.open(str(path), "wb") as wav: |
| wav.setnchannels(1) |
| wav.setsampwidth(2) |
| wav.setframerate(sample_rate) |
| frames = bytearray() |
| for index in range(sample_rate * duration): |
| time = index / sample_rate |
| frequency = frequencies[int(time * 3) % len(frequencies)] |
| signal = 0.16 * math.sin(2 * math.pi * frequency * time) |
| signal += 0.04 * math.sin(2 * math.pi * frequency * 1.5 * time) |
| frames.extend(struct.pack("<h", int(signal * 32767))) |
| wav.writeframes(frames) |
| examples.append([str(path), query]) |
| return examples |
|
|
|
|
| FAKE_EXAMPLES = create_fake_examples() |
|
|
|
|
| def spot( |
| audio_path: str | None, |
| query: str, |
| task: str = GROUNDING, |
| max_audio_seconds: int = DEFAULT_MAX_SECONDS, |
| max_new_tokens: int = 128, |
| ): |
| """UI-only stand-in for the original SpotSound inference callback.""" |
| return None, "", None |
|
|
|
|
| CSS = """ |
| #col-container { |
| max-width: 1180px; |
| margin: 0 auto; |
| height: calc(100vh - 32px); |
| overflow: visible; |
| } |
| .dark .gradio-container { color: var(--body-text-color); } |
| """ |
|
|
| with gr.Blocks( |
| title="SpotSound — fine-grained audio temporal grounding", fill_height=True |
| ) as demo: |
| with gr.Column(elem_id="col-container"): |
| gr.Markdown( |
| "# 🔍🔊 SpotSound — fine-grained audio temporal grounding" |
| ) |
| gr.Markdown( |
| """ |
| Find *when* a described sound happens inside a long recording. SpotSound is a LoRA on top of |
| **Audio Flamingo 3** that interleaves explicit timestamps into the audio stream, so the model can |
| answer with precise start/end times for short events buried in dense background noise. |
| |
| [Paper](https://huggingface.co/papers/2604.13023) · [Project page](https://loiesun.github.io/spotsound/) · |
| [Code](https://github.com/LoieSun/SpotSound) · [Model](https://huggingface.co/Loie/SpotSound) · |
| [Benchmark](https://huggingface.co/datasets/Loie/SpotSound-Bench) |
| """ |
| ) |
| with gr.Row(): |
| with gr.Column(scale=1): |
| audio_in = gr.Audio( |
| label="Audio recording", |
| type="filepath", |
| sources=["upload", "microphone"], |
| ) |
| query_in = gr.Textbox( |
| label="Sound to look for", |
| placeholder="e.g. dog barking, police car siren, hair dryer drying…", |
| lines=1, |
| ) |
| run_btn = gr.Button("Spot it", variant="primary") |
| with gr.Accordion("Advanced settings", open=False): |
| task_in = gr.Radio( |
| choices=[GROUNDING, DETECTION], |
| value=GROUNDING, |
| label="Task", |
| ) |
| max_seconds_in = gr.Slider( |
| 30, |
| 600, |
| value=DEFAULT_MAX_SECONDS, |
| step=30, |
| label="Truncate audio to (seconds)", |
| ) |
| max_tokens_in = gr.Slider( |
| 16, 512, value=128, step=16, label="Max new tokens" |
| ) |
| with gr.Column(scale=1): |
| plot_out = gr.Image( |
| label="Predicted temporal window(s)", type="pil" |
| ) |
| answer_out = gr.Textbox(label="Model answer", lines=4) |
| segment_out = gr.Audio( |
| label="Spotted segment(s) — full clip if nothing was found", |
| type="numpy", |
| ) |
|
|
| gr.Examples( |
| examples=FAKE_EXAMPLES, |
| inputs=[audio_in, query_in], |
| outputs=[plot_out, answer_out, segment_out], |
| fn=spot, |
| cache_examples=False, |
| label="Fake examples (UI only)", |
| ) |
|
|
| inputs = [audio_in, query_in, task_in, max_seconds_in, max_tokens_in] |
| outputs = [plot_out, answer_out, segment_out] |
| run_btn.click(fn=spot, inputs=inputs, outputs=outputs, api_name="spot") |
| query_in.submit(fn=spot, inputs=inputs, outputs=outputs, api_name=False) |
|
|
| if __name__ == "__main__": |
| demo.launch( |
| server_name="0.0.0.0", |
| server_port=7860, |
| show_error=True, |
| theme=gr.themes.Citrus(), |
| css=CSS, |
| mcp_server=True, |
| ) |
|
|
