| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - audio-classification |
| - audio-to-audio |
| tags: |
| - neurips2026 |
| - bioacoustics |
| - animal-vocalizations |
| - animalspeak |
| - pseudovox |
| - train-unseen |
| private: true |
| --- |
| |
| # AnimalSpeak Pseudovox Train-Unseen |
|
|
| This dataset contains the train-unseen split of AnimalSpeak Pseudovox. Each |
| example is a short, silence-trimmed, single-vocalization WAV clip plus compact |
| per-clip metadata. It does not include generated conversations, captions, QA |
| pairs, or MCQ answers. |
|
|
| Rows: 346,907 |
| Shards: 18 |
| Maximum rows per shard: 20,000 |
|
|
| ## Files |
|
|
| - `data-20k/train-*.tar`: WebDataset-style shards containing |
| `audio/<audio_name>` WAV entries. |
| - `metadata.parquet`: one row per audio file, including `id`, `audio_name`, |
| duration, species metadata, and classifier confidence metadata. |
| - `croissant.json`: MLCommons Croissant metadata for the sharded export. |
|
|
