spkrec-ecapa-voxceleb-ggml
a ggml speaker recognition model
2 min readupdated
The ivrit-ai team has converted SpeechBrain's popular spkrec-ecapa-voxceleb model—built on the ECAPA-TDNN architecture—to run via ggml (an inference engine for running AI models locally without heavy dependencies). The model extracts voice embeddings, which are numerical "voiceprints" used to identify who is speaking and when in an audio recording (diarization). Unlike the original PyTorch/SpeechBrain release, this port is optimized to run directly on client devices and ships with compact quantized weights. It is a solid fit for developers building local transcription tools, voice assistants, and mobile apps with speaker recognition.
Key features
- Compact footprint: Available in two weight files—
ecapa-q8_0.binat 23 MB andecapa-f16.binat 42 MB. - Negligible accuracy loss: Voiceprints from the
q8_0version achieve a mean cosine similarity of 0.9998 against the original SpeechBrain model (minimum 0.9994) and group into the exact same speaker clusters. - Optimized layout: Batch norm layers are folded into scale and bias, and convolution weights are unpacked into matrices for fast execution in
ggml.
What you can do
- Embed local speaker diarization (splitting transcripts by speaker) into offline audio and meeting transcription apps.
- Run on-device voice verification on resource-constrained hardware without sending audio to a remote server.
Example
The model takes SpeechBrain Fbank audio features as input (80 mel-filters, 400-point Hamming window STFT with a hop size of 160, decibels, and utterance-level mean normalization)—feature extraction is not bundled into the binary and must be handled by the host application. The weight files use a little-endian header with the IVSP magic signature (version 1) and load directly into ggml.
Read more
DetailsBase model: speechbrain/spkrec-ecapa-voxceleb (revision 0f99f2d0ebe89ac095bcc5903c4dd8f72b367286).
Quantization options: q8_0 (large matrices in q8_0, remaining tensors in f32/f16 — 23 MB) and f16 (f16 and f32 — 42 MB).
Accuracy: 0.9998 mean cosine similarity and 0.9994 worst-case against the original.
License: Apache-2.0.










Comments