---
license: mit
library_name: voxpath
tags:
- speaker-recognition
- quantizer
- voxpath
- pyannote
---

# voxpath quantizer

Trained BinaryKMeansQuantizer (12 binary k-means levels, grouped into 4 display words of vocab size 8) that maps L2-normalized 512-dim speaker embeddings
from [`pyannote/embedding`](https://huggingface.co/pyannote/embedding) into
hierarchical word paths like `chalk.fjord.bismuth.elm`.

Use with [`voxpath`](https://github.com/DJRHails/voxpath) to assign
LM-readable speaker identities for diarization + transcription pipelines.

## Files

- `voxceleb+librispeech+commonvoice.pyannote-embedding.quantizer.json` — fitted quantizer state (92.4 MB).

## Training corpus

Trained on 14,307 unique English-speaking speakers from VoxCeleb 2 (5,800) + LibriSpeech (3,507) + CommonVoice EN (5,000). Each speaker contributed one ≥3 s utterance, embedded with `pyannote/embedding`, L2-normalized. The merged corpus is available under the same namespace: [`DJRHails/pyannote-embedding-voxceleb`](https://huggingface.co/datasets/DJRHails/pyannote-embedding-voxceleb), [`DJRHails/pyannote-embedding-librispeech`](https://huggingface.co/datasets/DJRHails/pyannote-embedding-librispeech), [`DJRHails/pyannote-embedding-commonvoice-en`](https://huggingface.co/datasets/DJRHails/pyannote-embedding-commonvoice-en).

## Loading

```python
from huggingface_hub import hf_hub_download
from voxpath.quantize import BinaryKMeansQuantizer

path = hf_hub_download(
    "DJRHails/voxpath-bkm-pyannote-embedding",
    "voxceleb+librispeech+commonvoice.pyannote-embedding.quantizer.json",
)
quantizer = BinaryKMeansQuantizer.load(path)

# Then, given a pyannote/embedding output `embedding` (shape (512,)):
voxpath = quantizer.quantize(embedding)
print(voxpath.to_tag())  # e.g. SPEAKER:halite.rill.bismuth.elm
```

## Why model-and-embedder-specific

Speaker embeddings don't translate across embedders — a quantizer
fitted on `pyannote/embedding` outputs has no meaning for embeddings
from `wespeaker`, `ECAPA-TDNN`, or anything else. The repo name pins
both the quantizer family (BKM) and the embedding model so users find
the right artifact at a glance.