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

SpeechBrain ECAPA VoxCeleb Core ML

This is a reproducible compiled Core ML export of speechbrain/spkrec-ecapa-voxceleb. It produces a local voice embedding for cosine-similarity speaker comparison.

Model

Property	Value
Parameters	20.77 million
Format	Compiled Core ML, FLOAT16 compute
Compiled size	39.8 MiB
Input	SpeechBrain log-mel [1, frames, 80]
Sample rate	16 kHz
Output	192-dimensional L2-normalized embedding
Mel frames	10 to 3,001 (about 0.1 to 30 seconds)
Minimum deployment	macOS 15 / iOS 18

The graph includes sentence mean normalization. Audio-to-mel processing stays outside the graph so apps can share one tested streaming frontend between Core ML and MLX.

Files

File	Description
SpeechBrainECAPAVoxCeleb.mlmodelc/	Precompiled Core ML model
frontend.py	Reproducible SpeechBrain log-mel frontend
config.json	Graph, audio frontend, source revision, and checksums
artifact_manifest.json	SHA-256 and size of every compiled artifact file
validation.json	Conversion parity and local latency measurements
requirements.txt	Pinned standalone runtime dependencies
LICENSE	Apache 2.0 license

Validation

Check	Result	Meaning
Minimum PyTorch/Core ML output cosine	0.999995	1.0 is identical direction
Maximum absolute error	0.000702	Lower is better
Warm local inference	5.0 ms	Median after warm-up on the conversion machine

The official source reports 0.80% equal-error rate on the cleaned VoxCeleb1 test set. That upstream result is not presented as a new benchmark for this conversion. Before changing a speaker-verification threshold, validate the compiled artifact on VoxCeleb1-O and the microphones and languages used by the product.

Speaker embeddings are not secure biometric authentication and do not protect against replay or synthesized-voice attacks.

Usage

import coremltools as ct
import soundfile as sf

from frontend import compute_fbank

model = ct.models.CompiledMLModel("SpeechBrainECAPAVoxCeleb.mlmodelc")
audio, sample_rate = sf.read("voice.wav", dtype="float32")
assert sample_rate == 16000 and audio.ndim == 1
mel = compute_fbank(audio, 80)[None, :, :]
embedding = model.predict({"mel_features": mel})["embedding"]

The exact periodic Hamming window, symmetric SpeechBrain mel filters, decibel conversion, and frame limits are recorded in config.json.

Source

Converted from the official SpeechBrain checkpoint at revision 0f99f2d0ebe89ac095bcc5903c4dd8f72b367286. Checkpoint SHA-256 values and the SpeechBrain source revision used to reconstruct the graph are pinned in config.json.

Links

• speech-swift — Apple SDK
• Docs — install and CLI docs
• soniqo.audio
• blog