Transaction-Science/open-standards
Watch
1
0
Fork
You've already forked open-standards
 0
Code Issues Pull requests Projects Releases 17 Packages Wiki Activity Actions 189
open-standards/aac/aac-rs/data/asr/LICENSES.md
Transaction Science d6d20dbf10 feat(aac listener): listening ends 1 s after the speaker stops; 12 languages 
With the prototype's 500 ms rule, listening ended at the pause between
two sentences (Vietnamese 15 of 25 FLEURS clips, Hindi 4, German and
Portuguese 3): the partner was cut off, and only 6 languages passed on
our runtime. Cutting someone off is worse than about half a second more
wait, so the rule is now 1 s (runtime revision aac3).

Measured again with the worker as shipped (the same 25 clips per
language, the app's worker on WebGPU; identical, clip for clip, to the
earlier 1 s trial): ready are Spanish 4.5, Italian 6.9, Arabic 8.4,
English 9.8, Korean 10.7 (characters), Portuguese 11.4, French 11.7,
German 12.9, Ukrainian 13.8, Vietnamese 14.1, Russian 14.3, Japanese
14.6 (characters). Not yet: Turkish 16.2, Dutch 16.2, Polish 17.4, Hindi
17.5 (3 of its clips still cut, at pauses over a second). 4 of the 400
clips are still cut. The 500 ms results, and a run with no end at all,
stay in tools/asr/eval as the record of why.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-01 22:34:05 -04:00

2.2 KiB
Raw Permalink Blame History

Hearing your partner: licences

The listener in AuraSense AAC (aac-web) downloads one speech model and a voice activity detector, on request, and runs them on the person's device, on its graphics chip. What is heard there is turned into text there: nothing is recorded or sent anywhere.

Nemotron 3.5 ASR Streaming 0.6B

  • NVIDIA Nemotron 3.5 ASR Streaming 0.6B by NVIDIA, in soniqo's fp16 ONNX export, its graphs edited to run on WebGPU. Licence: OpenMDW-1.1 (https://openmdw.ai/license/1-1/); the text is licences/OpenMDW-1.1.2ab44b683654.txt.
  • Silero VAD v6.2.3 by the Silero Team, its graph edited to run on WebGPU. Licence: MIT (https://github.com/snakers4/silero-vad/blob/v6.2.3/LICENSE); the text is licences/MIT-silero-vad.2e63e9a38b6e.txt.
  • Source: huggingface.co/soniqo/Nemotron-3.5-ASR-Streaming-Multilingual-0.6B-ONNX-FP16@76daabfd0a (NVIDIA's huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b@ea30d66deb), with https://github.com/snakers4/silero-vad/tree/v6.2.3 (silero_vad_16k_sequence.onnx); graphs edited to run on WebGPU.
  • What was changed: the graphs only, so that every node runs on WebGPU in onnxruntime-web 1.30 (shapes folded, the attention mask rebuilt in int32, the mel front end merged into the encoder, the prediction net, joint and ArgMax fused into one step, the LSTMs unrolled, the VAD's first convolution padded to 132 channels); the weights are NVIDIA's and Silero's own, in 24 files of at most 64 MiB, each named by its own SHA-256 (1.24 GB).
  • soniqo's model card names the NVIDIA Open Model License, the licence NVIDIA's model had when soniqo exported it; NVIDIA has since published the model under OpenMDW-1.1 (its model card, revision ea30d66d), which this app follows.
  • Ready in: Arabic (ar), German (de), English (en), Spanish (es), French (fr), Italian (it), Japanese (ja), Korean (ko), Portuguese (pt), Russian (ru), Ukrainian (uk), Vietnamese (vi).

The runtime that runs it (onnxruntime-web, MIT) is vendored with its own licence, notice and ONNX Runtime's third-party notices (ThirdPartyNotices.txt) under vendor/onnxruntime-web-<version>-<revision>/.