OpenASR/wespeaker-voxceleb-resnet34-lm

WeSpeaker ResNet34 Speaker Embedder (VoxCeleb) · OpenASR

WeSpeaker ResNet34 speaker embedder for OpenASR diarization — stronger speaker separation, fully on-device

Speaker-diarization support pack for the OpenASR runtime — pure-Rust inference, no Python at inference time.

---

✨ Highlights

• 🗣️ Default OpenASR speaker embedder — install this pack and --diarize labels anonymous speakers for any ASR family
• 🧬 256-dim ResNet34 embeddings — pyannote's WeSpeaker VoxCeleb model with temporal statistics pooling
• 🔒 Diarization, not identification — anonymous session-relative labels; embeddings stay local and are discarded after the request unless you explicitly enroll a local profile
• 🎯 Parity-gated packaging — single raw-f32 build; the Rust forward pass matches the upstream pyannote reference at cosine ≥ 0.9999
• 🦀 Native in OpenASR — .oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU

🚀 Quickstart

# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull the pack
openasr pull wespeaker-voxceleb-resnet34-lm:f32

# 3. Diarize any transcription (works with every OpenASR ASR model)
openasr transcribe meeting.wav --model xasr-zh-en --diarize --format srt

📦 Pack

Quant	File (.oasr)	Size
f32	wespeaker-voxceleb-resnet34-lm-f32.oasr	27 MB

_{Single raw-f32 build: the pure-Rust forward pass consumes f32 directly and the parity gates assert bit-exact outputs vs the upstream weights, so no integer quantization is produced.}

🧠 About WeSpeaker ResNet34 Speaker Embedder (VoxCeleb)

WeSpeaker ResNet34 is the speaker-embedding model used by pyannote for VoxCeleb speaker representations. OpenASR packages it as a local .oasr capability pack and uses it as the default speaker-embedding stage for diarization: speech regions are embedded, clustered, and attributed back onto transcripts from any ASR model. The realtime speaker-change detector now uses this same WeSpeaker embedder, so diarization has one speaker-embedding space across batch and streaming paths.

⚙️ How this pack was made

Converted from pyannote/wespeaker-voxceleb-resnet34-LM with the OpenASR importer:

openasr model-pack import wespeaker <src>.safetensors <out>.oasr \
  --package-id wespeaker-voxceleb-resnet34-lm

The .oasr container is GGUF-backed; every tensor is stored as raw f32 so the pack round-trips bit-identically against the source weights.

⚖️ License

This pack inherits the upstream model's license: CC-BY-4.0 (source). OpenASR packaging retains the upstream copyright; the only modification is format conversion.

🙏 Acknowledgements

This pack redistributes pyannote/wespeaker-voxceleb-resnet34-LM in OpenASR's .oasr runtime format. Credit for the model architecture, training, and original weights belongs to the upstream pyannote/WeSpeaker authors. The upstream model is licensed under CC-BY-4.0; OpenASR packaging retains that license and attribution, with the only modification being format conversion for local runtime loading.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — pyannote/wespeaker-voxceleb-resnet34-LM