GigaAM Multilingual beats Whisper with 2M-hour audio pre-training for underrepresented languages

Multilingual automatic speech recognition (ASR) has made great strides, but long-tail languages remain underserved due to severe data scarcity. A team of researchers addresses this with GigaAM Multilingual, a foundation model designed specifically for underrepresented Central Asian languages — Kazakh, Kyrgyz, and Uzbek. The model is a Conformer encoder pre-trained on 2 million hours of audio using a HuBERT-style self-supervised objective. To combat head-language dominance (where high-resource languages skew performance), the team introduces a cluster-level data balancing strategy during pre-training and a domain-aware sampling method during fine-tuning. These techniques ensure that low-resource languages receive proportionate attention in the learning process.

In controlled comparisons, GigaAM Multilingual significantly outperforms strong open pretrained encoders like Whisper Large v3 and Omnilingual-1B on target languages, especially for spontaneous speech — a notoriously difficult domain. The model maintains computational efficiency despite its large pre-training corpus. The researchers release both the foundation encoder and the ASR model, providing a proven recipe for effective multilingual adaptation under realistic data imbalance. Accepted to Interspeech 2026, this work offers a practical blueprint for building robust speech systems for languages that are typically left behind in the AI boom.