---
license: cc0-1.0
task_categories:
- audio-classification
tags:
- speaker-embeddings
- speaker-recognition
- pyannote
---

# Pre-computed speaker embeddings

Pre-computed 512-dim L2-normalized speaker embeddings extracted with
[`pyannote/embedding`](https://huggingface.co/pyannote/embedding) over
LibriSpeech train.100 + train.360 (1172 speakers via openslr/librispeech_asr). One utterance per speaker, minimum 3 s duration.

## Contents

- `librispeech.pyannote-embedding.npz` — numpy `.npz` archive with:
  - `embeddings`: `(3507, 512)` float32
  - `speaker_ids`: `(3507,)` string IDs from the source corpus
  - `metadata_json`: per-speaker metadata (accent / age / gender / source URL)
    — populated for 1172 / 3507 speakers
  - `n_speakers`, `source` for provenance

## Loading

```python
import numpy as np
data = np.load("librispeech.pyannote-embedding.npz", allow_pickle=True)
embeddings = data["embeddings"]            # (N, 512)
speaker_ids = list(data["speaker_ids"])    # length N
```

## Regenerating

This file was produced by [`voxpath`](https://github.com/DJRHails/voxpath)
via:

```bash
voxpath corpus build commonvoice --max-speakers 3507 \
    --output librispeech.pyannote-embedding.npz
```

`voxpath corpus build` streams the source audio, embeds each speaker's
first valid (≥ 3 s) utterance with `pyannote/embedding`, L2-normalises,
and writes the `.npz`.

## Why model-specific

Speaker embeddings are not portable across embedders. A `wespeaker`
embedding and a `pyannote/embedding` embedding for the same audio lie
in different spaces and can't be compared or quantized together. This
repo is named after the embedding model so users can find the right
artifact for their pipeline at a glance.