enzolabs/ecapa-tdnn-coreml

ECAPA-TDNN speaker embeddings — Core ML

Core ML conversion of SpeechBrain's spkrec-ecapa-voxceleb ECAPA-TDNN speaker-verification model. Enzo uses it on the Mac to recognise the same voice across recordings.

Attribution and licence

• —ECAPA-TDNN speaker model © SpeechBrain, licensed Apache-2.0.
• —Converted to Core ML with coremltools 9.0 (PyTorch 2.8.0).

This bundle is redistributed under the Apache License 2.0 (see LICENSE).

Files

ecapa-tdnn-coreml.zip expands to chirp-speaker-v2/:

• —speaker.mlmodelc — the compiled Core ML model: input mel_features [1, 80, 301] (80-band log-mel features of 16 kHz audio), output embedding [1, 192] (speaker embedding)
• —mel_filterbank.npy, window.npy — the feature extractor's filterbank and window

The folder name is kept for compatibility with existing installations.

History

The archive Enzo first distributed from GitHub (SHA-256 6ffc6dcf6bc8ff8c49f069607825d817130296bcdf692e1f3c76eb478fdbdcfd, 62.0 MB) also carried an unrelated, unused experimental embedder nested inside the model folder. This version removes it (38.8 MB); the speaker model itself is unchanged and produces identical embeddings.