lang-id-voxlingua107-ecapa-coreml

ECAPA-TDNN VoxLingua107 CoreML

CoreML conversion of speechbrain/lang-id-voxlingua107-ecapa for native inference on Apple Silicon (macOS 14+ / iOS 17+).

Identifies 107 spoken languages from log-mel spectrogram. No Python required at runtime.

Model Details

Benchmark Results

Tested on Apple Silicon (M1, Metal GPU, .cpuAndGPU):

15-50x faster than MMS-LID-256 with comparable accuracy.

Usage (Swift)

import CoreML

let compiledURL = try MLModel.compileModel(at: URL(fileURLWithPath: "EcapaTdnnLid107.mlpackage"))
let config = MLModelConfiguration()
config.computeUnits = .cpuAndGPU  // ANE provides no benefit for this model
let model = try MLModel(contentsOf: compiledURL, configuration: config)

// melFrames: [[Float]] — log-mel spectrogram [T][60]
let T = melFrames.count
let inputArray = try MLMultiArray(shape: [1, NSNumber(value: T), 60], dataType: .float32)
for t in 0..<T {
    for f in 0..<60 {
        inputArray[[0, NSNumber(value: t), NSNumber(value: f)]] = NSNumber(value: melFrames[t][f])
    }
}

let input = try MLDictionaryFeatureProvider(dictionary: [
    "mel_features": MLFeatureValue(multiArray: inputArray),
])
let output = try model.prediction(from: input)

Mel Spectrogram

The model expects a log-mel spectrogram computed with SpeechBrain-compatible parameters. This must be computed on-device (not baked into the CoreML model).

Implementation Notes

1. DFT size 400 is not power-of-2 — Apple's vDSP_fft_zrip silently computes wrong results. Use manual DFT via cblas_sgemv with precomputed twiddle factors.
2. SpeechBrain filterbank differs from HTK — uses symmetric triangular filters: band[m] = hz[m+1] - hz[m] applied equally to both sides.
3. Periodic Hamming window — vDSP_hamm_window generates symmetric windows. Create N+1 symmetric, take first N for periodic.
4. CMVN stats are identity (all zeros) — normalization can be skipped.

Files

• EcapaTdnnLid107.mlpackage/ — CoreML model
• ecapa_tdnn_lid107_labels.json — language label mapping (index → "code: Name")

Conversion

Converted via torch.jit.trace → coremltools 9.0. See conversion script.

Full Inference Code

Complete Swift CLI with audio loading, mel spectrogram (Accelerate/vDSP), and inference: github.com/beshkenadze/lid-bench

License

Apache 2.0 (same as the original model)