cstr/wespeaker-resnet34-lm-GGUF

WeSpeaker ResNet34-LM — GGUF (ggml conversion)

GGUF conversion of `Wespeaker/wespeaker-voxceleb-resnet34-LM`, a 256-dimensional speaker-embedding model, for the --diarize-method foxnose diarizer in [CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR).

⚠ Licence — attribution is required

These weights are CC-BY-4.0, inherited from the upstream model. Several downstream projects describe them as Apache-2.0; that is incorrect — the wenet-e2e/wespeaker code is Apache-2.0, the published weights are CC-BY-4.0 and carry an attribution requirement. If you redistribute these files, keep the attribution.

Source: Wespeaker/wespeaker-voxceleb-resnet34-LM Upstream: https://github.com/wenet-e2e/wespeaker Licence: https://creativecommons.org/licenses/by/4.0/

Files

Conv kernels stay F32 in both, and that is a speed choice. An F16 conv kernel is numerically fine (cosine 0.99999724 against the PyTorch oracle) and would shrink the file to 13.3 MB, but ggml's CPU conv path is 2.2× slower on it — 297 ms vs 133 ms per 1.2 s window, measured back to back over the same 352 windows on an M1. ResNet34 is ~94% of embedding time, so 10 MB of disk is not worth it. F16 on the 2-D linear is free and yields an identical embedding (cosine 0.99999744 vs 0.99999747).

Architecture

ResNet34 [3,4,6,3] over 80-bin Kaldi fbank, TSTP pooling, Linear(5120→256). BatchNorm is folded into every convolution at conversion time (219 → 74 tensors) and the ArcMargin projection head is training-only and dropped (11.25 M → 6.6 M params).

Three details that decide correctness, traced to wespeaker/cli/speaker.py rather than assumed:

• —the waveform is int16-scale (torchaudio.load(normalize=False), since wavform_norm defaults to False), window is hamming, then per-utterance CMN;
• —the 2-D map is height=freq, width=time — TSTP reduces over time and flattens (channel, freq) with freq fastest, which is the order seg_1's 5120 columns are in;
• —TSTP's std uses torch's unbiased (n−1) variance, +1e-7 inside the sqrt;
• —the output is seg_1(stats) raw — no ReLU, no BatchNorm, no L2 normalisation.

Verification

Per-stage against the upstream PyTorch model run as an oracle (crispasr-diff wespeaker), on an 11 s clip:

Discriminative check on real audio: two windows of the same speaker score cosine 0.595, against 0.100 for a different speaker.

End-to-end, the CrispASR diarizer built on this model scores DER 3.93% against the upstream Python pipeline's own output (same pinned speaker count, 0.25 s collar) with zero speaker confusion — the residual is entirely false alarm from a different speech-segmentation source.

⚠ That is a PARITY number, not an accuracy one. It says this port reproduces the reference implementation; it does not say either is 96% right. Measured against HUMAN labels on VoxConverse dev (40 files, --diarize-max- speakers 8, whisper-tiny segments, 0.25 s collar):

Estimating the number of speakers is the weak link, not the embeddings — the embedding matches the PyTorch oracle to cosine 0.99999747. Pass --diarize-num-speakers N when you know the count and the picture improves sharply. Numbers near 3–7% DER quoted elsewhere in this project's history came from an 8-file subset that turned out to be unrepresentative: identical code scores 7.3% there and 33.1% on the fuller corpus.

Usage

crispasr -m <asr-model.gguf> -f audio.wav \
    --diarize --diarize-method foxnose \
    --diarize-embedder wespeaker-resnet34-lm.gguf

Conversion

python models/convert-wespeaker-to-gguf.py \
    --model Wespeaker/wespeaker-voxceleb-resnet34-LM \
    --output wespeaker-resnet34-lm.gguf

The CrispASR runtime is an independent implementation written from the published architecture; no upstream source is incorporated.

Provenance and EU AI Act Art. 53 note

• —Upstream model: Wespeaker/wespeaker-voxceleb-resnet34-LM — published by Wespeaker.
• —Upstream licence: cc-by-4.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• —What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• —Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• —Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.