GigaAM Uzbek 600M — sherpa-onnx (int8)

ONNX export of rustam1221/uzbek-asr-gigaam checkpoint large_full_600m (GigaAM-Multilingual large CTC, fine-tuned on Uzbek, with punctuation), packaged for sherpa-onnx as a NeMo CTC model (is_giga_am=1).

• model.int8.onnx — dynamic int8 quantization (QUInt8 weights)
• tokens.txt — 74 characters (incl. . , ? !) + <blk>

Output is lowercase with punctuation; capitalize sentence starts yourself if needed.

Usage

import sherpa_onnx
rec = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(model="model.int8.onnx", tokens="tokens.txt")

Check

150 FLEURS uz_uz test clips, same normalization on both sides: PyTorch original 13.0% WER, this int8 ONNX 13.3% WER.

Credits

All training by rustam1221, base model ai-sage/GigaAM-Multilingual. MIT license.