Add new GigaAM models

[mikeozornin]

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

[cjpais]

Please add an issue to handy-computer/transcribe.cpp

[mikeozornin]

Oh, thanks for a such quick answer. Here is it: handy-computer/transcribe.cpp#87

[cjpais]

Thank you! Will track there

[Roffild]

These models currently don't transcribe for more than 25 seconds.

Whisper understands me better.

[ekhodzitsky]

On the "does not transcribe past ~25 seconds" problem: that is the integration pattern, not the model. These models consume fixed-length mel windows, so a single encoder pass tops out around 25-30s. The fix is chunked long-form decoding — split on silence, carry decoder state across chunks — after which hours of audio work fine.

On GigaAM-Multilingual: it is an encoder-only CTC model, which makes it simpler to integrate than the RNN-T variant (no decoder/joiner, greedy decode is just argmax + collapse). Accuracy across the five supported languages is solid — roughly 4.4-9.2% clean-read WER (Russian 4.44, English 4.63, Kazakh 6.52, Kyrgyz 7.39, Uzbek 9.21).

(Disclosure: I ship both heads in a server I maintain — https://github.com/ekhodzitsky/gigastt — the chunked decoding and CTC integration code is there if useful as a reference.)