---
license: other
task_categories:
- audio-classification
- audio-to-audio
tags:
- neurips2026
- bioacoustics
- sound-event-detection
- synthetic
private: true
---

# Synthetic Detect Diarize

This dataset contains synthetic bioacoustic 10-second WAV audio and source
annotation metadata. It does not include generated language-pair conversations,
captions, QA pairs, or other `data-synth` outputs.

Rows: 1,000,000
Shards: 50
Maximum rows per shard: 20,000

## Files

- `data-20k/train-*.tar`: WebDataset-style shards containing `audio/<audio_name>`
  and `selection_tables/<selection_table_name>` entries.
- `metadata.parquet`: one row per audio file, with `id`, `audio_name`, and
  `selection_table_name`.

The selection tables include diarization metadata such as `Speaker`, `SNR (dB)`, `Source Stem`, and `Source File` when present.

## Source

Canonical source: `gs://foundation-model-data/synthetic/synthetic_sed_diarization_16k`