Sinhala speech recognition model for VocalDraft
This repository mirrors the unmodified Sinhala ASR model by Deshan Alahakoon for use by VocalDraft's in-browser, on-device dictation. The original project, documentation, limitations and evaluation are at ninjawerk/sinhala-asr. The model is licensed under Apache 2.0; please retain the included LICENSE and NOTICE files.
weights.wasm contains the original raw external model weights (SHA-256 b2b427122c556d7f2577964c382398792c0b6d9e88cea1ed6a1f3d787531ae3c). Its extension helps browsers download it reliably; it is not a WebAssembly executable. For use with the original ONNX graph, rename it to weights.bin and place it beside sinhala_asr_int8.onnx and vocab.json. VocalDraft verifies the weights before loading them and performs speech recognition on the visitor's device.
The base encoder is Meta's omniASR-W2V-300M, and the model was trained with Google's OpenSLR SLR52 corpus. Consult the original project for the language model, decoding pipeline, accuracy and licensing details.