Transcripta/sherpa-onnx-speaker-diarization-wasm

sherpa-onnx speaker diarization (WebAssembly)

Prebuilt WebAssembly assets for fully on-device speaker diarization ("who spoke when"), redistributed for the TranscriptAI / FileWhirl on-device audio features so they can be fetched with CORS at runtime and cached in the browser. No audio leaves the device.

These are the unmodified prebuilt artifacts from sherpa-onnx v1.13.7 (k2-fsa), speaker-diarization WASM target. The pipeline is pyannote-segmentation-3.0 (ONNX) + a speaker-embedding model + agglomerative clustering, all inside the wasm module. The models are baked into the .data preload.

Files

• sherpa-onnx-wasm-main-speaker-diarization.js / .wasm / .data — the Emscripten build
• sherpa-onnx-speaker-diarization.js — the JS API wrapper (createOfflineSpeakerDiarization)

Usage

Load the three scripts (define Module = {} with onRuntimeInitialized first), then: const sd = createOfflineSpeakerDiarization(Module); const segments = sd.process(float32Mono16k); Each segment is { start, end, speaker } (seconds, integer cluster id).

Attribution & licenses

• sherpa-onnx — Apache-2.0, k2-fsa: https://github.com/k2-fsa/sherpa-onnx (source release v1.13.7)
• pyannote/segmentation-3.0 — MIT: https://huggingface.co/pyannote/segmentation-3.0
• Speaker-embedding model — WeSpeaker / 3D-Speaker; VoxCeleb-trained weights are CC-BY-4.0 (attribution required for redistribution / commercial use).

Redistributed under the above terms with attribution.