You can deploy this model and start earning money today!

talatapp/
wespeaker-voxceleb-resnet34-LM-onnx

Name: wespeaker-voxceleb-resnet34-LM-onnx
Availability: InStock

Wespeaker/wespeaker-voxceleb-resnet34-LM

quantized

apache-2.0

Views

No views yet

wespeaker-voxceleb-resnet34-LM ONNX (single-file, baked fbank + masking)

Re-export of Wespeaker/wespeaker-voxceleb-resnet34-LM to a single ONNX file whose I/O contract matches FluidInference's CoreML wespeaker_v2.mlmodelc:

role	name	dtype	shape
input	`waveform`	f32	`[3, 160_000]`
input	`mask`	f32	`[3, 589]`
output	`embedding`	f32	`[3, 256]`

What's in the graph

The wrapper bakes the full preprocessing chain into the ONNX so the consumer hands it raw 16 kHz mono PCM and gets back a 256-dim embedding per speaker slot:

STFT — 25 ms hamming window, 10 ms hop, kaldi-style snip_edges=False.
80-bin mel filterbank — kaldi-compatible HTK mel scale, slaney normalisation, log power.
Mask resampling — 589-frame pyannote-3.0 mask → 998-frame fbank time base via nearest-neighbour.
Per-utterance CMN over masked frames — subtracts the mean of active frames so silence / other-speaker frames don't drag the centroid.
Masked feature gating — multiplies fbank by the resampled mask so the encoder ignores unmasked frames.
WeSpeaker resnet34-LM encoder + ASP pooling — vanilla upstream encoder, untouched weights.

Slots whose mask.sum() == 0 get an all-zero embedding so the downstream matcher can detect "no signal" without seeing whatever the encoder produces from all-zero input.

Files

File	Size	Use
`voxceleb_resnet34_LM.onnx`	~26 MB	full-precision fp32
`voxceleb_resnet34_LM.int8.onnx`	~7 MB	per-channel int8 dynamic quant — Windows manifest target

Why this exists

The Mac side of talat uses FluidInference's CoreML conversion of WeSpeaker, which bakes the preprocessing chain into the graph. The upstream Wespeaker/wespeaker-voxceleb-resnet34-LM ONNX export does NOT — it expects pre-computed fbank features instead. So Windows hitting the same upstream ONNX blew up at engine startup with "missing 'waveform' input".

This re-export gives both platforms the same I/O surface so the Rust talat-inference::EmbeddingExtractor works unchanged across Mac (CoreML) and Windows (ONNX). The asymmetric quant story matches the rest of the in-process model bundle: fp16 CoreML on Apple silicon (ANE-native), int8 ONNX on Windows (DirectML / CPU EP).

The export script is at scripts/wespeaker-onnx-export/ in the talat repo — convert_wespeaker.py is heavily commented if you need to adapt it.

Numerical equivalence

The cosine similarity against FluidInference's CoreML version is close-but-not-bit-exact because:

FluidInference's CoreML conversion script isn't open-source, so the exact masking strategy inside their graph (post-fbank gate vs. inside-ASP attention weighting) might differ from ours.
int8 weight-only dynamic quantisation on the Windows file introduces per-layer rounding drift that compounds through the resnet stack.

The speaker matcher compares embeddings by cosine similarity in a 256-dim space, which is forgiving of small drift. If cluster decisions diverge between Mac and Windows on the same audio, the masking strategy in the wrapper is the first thing to revisit.

License

Inherits the WeSpeaker (Apache 2.0) and VoxCeleb (CC-BY-4.0 / academic) licenses of the upstream checkpoint.

talatapp/wespeaker-voxceleb-resnet34-LM-onnx