SpeechBrain-ECAPA-VoxCeleb-20M-MLX

aufklarer/SpeechBrain-ECAPA-VoxCeleb-20M-MLX

SpeechBrain-ECAPA-VoxCeleb-20M-MLX is a feature extraction model from aufklarer. Use it when you need embeddings to search or compare text. It is set up for mlx. The card lists the license as apache-2.0.

This is a reproducible MLX weight export of speechbrain/spkrec-ecapa-voxceleb for native Apple Silicon inference. It produces a 192-dimensional, L2-normalized voice embedding for cosine-similarity speaker comparison.

Downloads · 30 days

119

60% of all-time downloads

All-time downloads

199

Public

Parameters

20.8M

41.6 MB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors41.6 MB · 100%

At a glance

Task: Feature Extraction
Library: mlx
License: apache-2.0
Model type: speechbrain-ecapa-voxceleb-speaker
Access: Public
Created: Aug 9, 2026
Updated: Aug 9, 2026
SHA: e749e6e0

Base models

speechbrain/spkrec-ecapa-voxceleb

Model card

From the Hugging Face model README

SpeechBrain ECAPA VoxCeleb MLX

This is a reproducible MLX weight export of speechbrain/spkrec-ecapa-voxceleb for native Apple Silicon inference. It produces a 192-dimensional, L2-normalized voice embedding for cosine-similarity speaker comparison.

Model

Property	Value
Parameters	20.77 million
Weights	Safetensors, FLOAT16
Weight size	39.7 MiB
Input	SpeechBrain log-mel `[1, frames, 80]`
Sample rate	16 kHz
Output	192-dimensional L2-normalized embedding

Files

File	Description
`model.safetensors`	MLX-layout ECAPA weights
`mlx_model.py`	Native MLX ECAPA graph
`frontend.py`	Reproducible SpeechBrain log-mel frontend
`config.json`	Architecture, frontend, source revision, and checksums
`artifact_manifest.json`	SHA-256 and size of every artifact file
`validation.json`	PyTorch/MLX parity and local latency measurements
`requirements.txt`	Pinned standalone runtime dependencies
`LICENSE`	Apache 2.0 license

Validation

Check	Result
Minimum PyTorch/MLX output cosine	0.999999
Maximum absolute error	0.000372
Warm local inference	5.2 ms

The source model reports 0.80% equal-error rate on cleaned VoxCeleb1. Run a fresh VoxCeleb1-O evaluation before treating that upstream score as conversion parity or changing a production threshold.

Usage

import mlx.core as mx
import soundfile as sf

from frontend import compute_fbank
from mlx_model import SpeakerModel

audio, sample_rate = sf.read("voice.wav", dtype="float32")
assert sample_rate == 16000 and audio.ndim == 1
features = compute_fbank(audio, 80)[None, :, :]
model = SpeakerModel()
model.load_weights(list(mx.load("model.safetensors").items()), strict=True)
embedding = model(mx.array(features))
mx.eval(embedding)
print(embedding.shape, float(mx.linalg.norm(embedding).item()))

The exact frontend contract is in config.json. Speaker embeddings are not secure authentication and should not be used alone for access control.

Source

Converted from the official SpeechBrain checkpoint at revision 0f99f2d0ebe89ac095bcc5903c4dd8f72b367286. The original checkpoint and graph revisions are pinned in config.json.