Modelpublic

aufklarer/SpeechBrain-ECAPA-VoxLingua107-21M-CoreML

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes31downloads

Model Card

SpeechBrain ECAPA VoxLingua107 Core ML

This is a pinned, validated compiled Core ML export of `speechbrain/lang-id-voxlingua107-ecapa`. It identifies one of 107 spoken-language labels on Apple devices.

Model

Property	Value
Parameters	21.25 million
Format	Compiled Core ML, FLOAT16 compute
Compiled size	40.8 MiB
Input	SpeechBrain log-mel `[1, frames, 60]`
Sample rate	16 kHz
Output	107 log probabilities
Mel frames	10 to 3,001 (about 0.1 to 30 seconds)
Minimum deployment	macOS 15 / iOS 18

Unlike several earlier community conversions, this graph includes the sentence-level mean normalization used by the official SpeechBrain inference pipeline. The release also validates an independent audio frontend against SpeechBrain itself.

Files

File	Description
`SpeechBrainECAPAVoxLingua107.mlmodelc/`	Precompiled Core ML model
`frontend.py`	Reproducible SpeechBrain log-mel frontend
`labels.json`	Stable upstream index, code, and name mapping
`config.json`	Frontend, graph, source revisions, 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
Minimum PyTorch/Core ML output cosine	0.999967
Maximum absolute error	0.217239
Warm local inference	3.7 ms

The upstream card reports 6.7% classification error on the VoxLingua107 development set. That development set contains only 1,609 manually verified clips across 33 of the 107 labels, so it is not a complete 107-language evaluation.

Usage

python

import coremltools as ct
import numpy as np
import soundfile as sf

from frontend import compute_fbank

model = ct.models.CompiledMLModel("SpeechBrainECAPAVoxLingua107.mlmodelc")
audio, sample_rate = sf.read("recording.wav", dtype="float32")
assert sample_rate == 16000 and audio.ndim == 1
mel = compute_fbank(audio, 60)[None, :, :]
log_probabilities = model.predict({"mel_features": mel})["log_probabilities"]

This is a closed-set classifier: it always ranks a known label. Applications need confidence calibration and an unknown-language rejection policy.

Source