cstr / gigaam-v3-GGUF

GigaAM-v3 — GGUF (ggml conversions)

GGUF conversions of ai-sage/GigaAM-v3 for use with the gigaam backend in CrispStrobe/CrispASR .

GigaAM-v3 is a 220 M-parameter Conformer foundation model for Russian ASR, pretrained with a HuBERT-CTC objective on ~700 K hours of Russian speech. The upstream repo ships five checkpoints as git revisions; the four ASR ones are converted here (the ssl encoder has no head and produces no transcript).

Which one to pick

gigaam-v3-e2e-rnnt-q8_0.gguf unless you have a reason not to — it is the lowest-WER variant, emits punctuation and casing, and its transcript is identical to the PyTorch reference.

Usage

crispasr --backend gigaam -m gigaam-v3-e2e-rnnt-q8_0.gguf -f audio.wav
# or let the registry fetch it:
crispasr --backend gigaam -m auto --auto-download -f audio.wav

Audio is 16 kHz mono. Long inputs are sliced by the CLI's VAD/chunking; the model itself has a ~25 s practical window (full attention, O(T²)).

Verification

Every file was checked against a per-stage PyTorch reference dumped from the upstream modeling_gigaam.py ( crispasr-diff gigaam <model> <ref> <wav> ), on GigaAM's own example.wav :

So: q8_0 is the safe quant ; q4_k is fine for the charwise models and costs a little casing/punctuation fidelity on the two SentencePiece ones.

In every quant the mel filterbank, Hann window, encoder.pre.* subsampling convs and the decode head ( joint.* / decoder.* / head.ctc.* ) are kept at source precision — the mel is un-normalized log-mel, so subsampling rounding error would otherwise cascade through all 16 conformer blocks, and the head is a blank-vs-token argmax where a flipped decision derails the greedy decode.

Conversion

python models/convert-gigaam-to-gguf.py \
    --model ai-sage/GigaAM-v3 --revision e2e_rnnt \
    --output gigaam-v3-e2e-rnnt-f16.gguf
./build/bin/crispasr-quantize gigaam-v3-e2e-rnnt-f16.gguf \
    gigaam-v3-e2e-rnnt-q8_0.gguf q8_0

License

MIT, inherited from ai-sage/GigaAM-v3 . Please cite the upstream model when you use these weights.