Spaces:
Running
Running
| title: Deep Voice — Bioacoustic Detection | |
| emoji: 🐳 | |
| colorFrom: purple | |
| colorTo: pink | |
| sdk: gradio | |
| sdk_version: 5.34.1 | |
| python_version: "3.11" | |
| app_file: app.py | |
| pinned: false | |
| license: apache-2.0 | |
| short_description: Multi-species call detection in underwater recordings. | |
| # Deep Voice — Bioacoustic Call Detection | |
| A public demo for running **[Deep Voice](https://huggingface.co/deepvoice1)** call-detection | |
| models on your own audio. Seven models trained in collaboration with marine biologists are | |
| available out of the box: | |
| | Species / call type | Window | Min input SR | Threshold | | |
| |---|---|---|---| | |
| | Arctic cod fish (Boreogadus saida) | 1 s | 2 kHz | 0.5 | | |
| | Greater Caribbean manatee | 0.2 s | 96 kHz | 0.5 | | |
| | Burrunan dolphin — barks | 3 s | 10 kHz | 0.5 | | |
| | Burrunan dolphin — echo | 3 s | 96 kHz | 0.9 | | |
| | Burrunan dolphin — buzz | 0.5 s | 96 kHz | 0.8 | | |
| | Burrunan dolphin — whistle | 1 s | 96 kHz | 0.5 | | |
| | Killer whale (orca) — 5-class | 1.5 s | 24 kHz | 0.5 | | |
| | Humpback whale — Mozambique C1 group | 1 s | 16 kHz | 0.5 | | |
| Uploads at a higher sample rate are resampled automatically; uploads below | |
| the minimum are rejected (the missing high-frequency content can't be | |
| reconstructed). | |
| ## How it works | |
| 1. Pick a model from the dropdown. | |
| 2. Upload `.wav` files (up to **500 MB total**; runs that would exceed **15 min** | |
| of compute on the free CPU tier are rejected with a clear message). | |
| 3. Click **Run inference**. | |
| 4. Download: | |
| - **CSV** — per-window class probabilities (one CSV per file; zipped if multiple). | |
| - **Raven `.txt`** — selection table compatible with [Raven Pro / Lite](https://www.ravensoundsoftware.com/). | |
| When you upload more than one file, the Raven outputs are merged into a single | |
| selection table with a `Begin File` column and per-file time offsets. | |
| Running on the free HuggingFace CPU tier. Throughput varies a lot per model — | |
| low-sample-rate models (e.g. Arctic cod, 2 kHz internal) process ~100× faster | |
| than the 96 kHz dolphin/manatee models. The pre-flight line in the UI shows a | |
| per-model runtime estimate after you upload. | |
| ## Models and code | |
| - **Soundbay** training/inference framework: <https://github.com/deep-voice/soundbay> | |
| - **Checkpoints** (gated): `deepvoice1/bioacoustic-checkpoints` on the HF Hub. | |
| - **This Space** vendors the soundbay package and pulls checkpoints at runtime via `HF_TOKEN`. | |
| If you'd like to use these models in your own pipeline, see the soundbay repo for the full | |
| inference CLI and config templates. | |
| ## License | |
| Apache 2.0 — please credit Deep Voice when publishing detections produced with these models. | |