GigaAM Multilingual: Foundation Model for Underrepresented Languages

Researchers at SaluteDevices introduce GigaAM Multilingual, an open-weight Conformer speech-recognition foundation model targeting underrepresented Central Asian languages, especially Kazakh, Kyrgyz, and Uzbek. The encoder uses a HuBERT-style masked-unit prediction objective, Rotary Position Embeddings, and a 600M-parameter, 24-layer architecture trained on 2M hours of audio. Its main innovation is balancing multilingual data at the level of five language clusters during pre-training, then combining language- and domain-aware sampling during CTC fine-tuning to prevent English and Russian from dominating. The training mixture includes open-source, crowdsourced, synthetic, and weakly supervised speech, with synthetic data generated from mC4 text using an in-house TTS system with over 100 voices. On Common Voice, FLEURS, and internal spontaneous-speech tests, GigaAM substantially outperforms Whisper Large v3, Seamless-M4T v2 large, and Omnilingual ASR on the target languages. In matched CTC experiments, even the compact 240M GigaAM encoder reaches an average WER of 12.2%, versus 14.1% for Whisper Large v3 and 16.6% for Omnilingual-1B. For minimal-coverage adaptation using only Common Voice training data, GigaAM achieves 3.8% WER on Georgian and 3.6% on Bashkir, beating both baselines. The results show that structured sampling, rather than scale alone, can make multilingual ASR more effective and efficient for long-tail languages.