tools/asr/build_asr.py publishes what the listener ("Hearing your
partner") downloads: NVIDIA Nemotron 3.5 ASR Streaming 0.6B (soniqo's fp16
ONNX export) and Silero VAD v6.2.3, with the graphs the asr-proto
prototype edited so every node runs on onnxruntime-web 1.30's WebGPU
provider. Each of the 24 files (1.24 GB, at most 64 MiB each) is checked
against its pinned size and SHA-256 and named by its contents; the
catalogue (aac-asr-index/1) lists the three graphs, their weights by the
names the graphs give them, the vocabulary, and the credits: OpenMDW-1.1
for the model and MIT for Silero, with both texts. tools/asr/check_packs.py
matches a directory to the committed catalogue before a site ships it.
A language is offered only with evidence: NVIDIA's published FLEURS error
rate at 320 ms chunks with the language given (as the app runs it), or
ours on the app's own runtime, at most 15 in 100 words (characters for
Chinese, Japanese and Korean), and none of ours above that.
languages.json records each number, its source and the plain note
Settings shows, and each language's prompt slot (the app's locale first:
pt-BR 12, pt-PT 13).
- Ready (15): Arabic, Dutch, English, French, German, Hindi, Italian,
Japanese, Korean, Portuguese, Russian, Spanish, Turkish, Ukrainian,
Vietnamese.
- Not yet, too many errors (9): Polish, Norwegian, Chinese, Finnish,
Czech, Swedish, Romanian, Danish, Hungarian.
- Not yet, not the model's (21): the app's other languages.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
26bad4aa72