Repository navigation
Add support for new giga asr models #87
Description
Recently two ASR models were published:
https://huggingface.co/ai-sage/GigaAM-Multilingual
https://huggingface.co/ai-sage/GigaChat3.1-Audio-10B-A1.8B
In Russian, Kazakh, Kyrgyz, and Uzbek tests they performed better than whisper large. It will be nice to have it in Handy, maybe it's worth to replace older GigaAM models.
Activity
cjpais commentedon Jul 16, 2026
I think GigaAM-Multilingual will probably come in first mainly because it's probably simpler to port and runs on more systems
ekhodzitsky commentedon Jul 22, 2026
there's already an int8 onnx export of the multilingual model on hf: istupakov/gigaam-multilingual-ctc-onnx (~225mb). it's encoder-only ctc, decode is just argmax + collapse — much simpler than the rnnt heads. the 25s thing only happens if you run the encoder once on the whole file, split on silence and it's fine for hours
koloved commentedon Aug 31, 2026
@ekhodzitsky I tried using this model. Firstly, it doesn't put punctuation marks or grammar/punctuation, and it supports multilingualism very poorly. You cannot use words from different languages during use: English and Russian. For some reason, it continues to transcribe English words in Cyrillic.
ekhodzitsky commentedon Aug 31, 2026
@koloved both of those are expected for this head, not a transcribe.cpp bug.
No punctuation. Character CTC has no punctuation in the vocab — it emits lowercase letters only. In gigastt that’s a separate RuPunct pass (
--punctuation on). Handy’s GigaAM-v3 e2e heads bake punctuation in; multilingual CTC does not.English coming out in Cyrillic / no mixing. “Multilingual” here means five languages (ru/en/kk/ky/uz), usually one language per clip. It is not trained to code-switch inside a sentence. On Russian-dominant audio, English words get spelled in Cyrillic (
google→гугл). A Russian-only GigaAM-v3 head cannot emit Latin at all — its vocab is Cyrillic. Whisper is better at mixed ru+en because of a large multilingual BPE vocab.So: monolingual Russian with punctuation → v3 e2e (or CTC + RuPunct). Kazakh/Kyrgyz/Uzbek or English-only clips → multilingual CTC. Mixed ru+en in one utterance → Whisper still wins.