GigaAM Multilingual: Foundation Model for Underrepresented Languages
Andrei Kuzmenko, Alexandr Maximenko, Aleksandr Kutsakov, Georgii Gospodinov, Dmitrii Bolotov, Oleg Kutuzov, Pavel Bogomolov, Fyodor Minkin
eess.AS, cs.CL
2026-07-11
GigaAM Multilingual pretrains a 600M Conformer on 2M hours with cluster-level data balancing, cutting WER on Kazakh, Kyrgyz and Uzbek to roughly 10-16%, far below Whisper Large v3 and Omnilingual-1B.
Once multilingual ASR scales up, head languages (English, Russian) are recognized well, but long-tail languages stay poor, with error rates too high for downstream use. This paper targets Kazakh, Kyrgyz and Uzbek, which have very little data. The hard part is not only scarcity: head-language data is so abundant that training on raw proportions lets English and Russian dominate, and the tail never learns. Plain per-language balancing is hard because weights for tail languages cannot be estimated reliably.
The authors pretrain a 600M-parameter Conformer encoder with a HuBERT-style masked unit prediction objective on 2M hours of audio (24 layers, 1024 hidden, rotary position embeddings, 25 Hz frame rate). Discrete labels come from k-means with 1,000 clusters.
Two data strategies target head-language dominance. During pretraining, cluster-level balancing: rather than balancing per language, they first group 70-plus languages into 5 clusters using a weighted co-occurrence graph, then tune cluster sampling weights. The selected configuration raises the weight of the cluster containing the Central Asian languages from 0.08 to 0.25 and lowers the English cluster from 0.60 to 0.50. Tuning by cluster is more stable than tuning per language, because co-occurrence within a cluster makes weights easier to estimate.
During fine-tuning, domain-aware sampling uses a CTC objective with a shared character vocabulary across 5 languages, and stratifies by domain within each language to prevent synthetic subsets from dominating. Fine-tuning data totals about 54,800 hours, with Kazakh and Kyrgyz relying heavily on synthetic data (7,896 and 6,415 hours respectively).
Direct comparison with three strong baselines on internal test sets (WER, lower is better):
| Language | GigaAM | Whisper v3 | Omnilingual-1B |
| Kazakh | 15.8 | 65.2 | 32.2 |
| Kyrgyz | 9.8 | 102.2 | 25.0 |
| Uzbek | 12.7 | 120.6 | 30.2 |
| Russian | 6.0 | 10.1 | 14.6 |
Whisper Large v3 posts WER above 100% on Kyrgyz and Uzbek, which is essentially broken; GigaAM brings them to about 10%. On English, Whisper is slightly better (20.0 vs 21.5), the cost of trading head-language accuracy for the tail.
To rule out "it just had more fine-tuning data," an encoder-matched ablation runs the same CTC fine-tuning on all encoders: GigaAM 600M averages 10.2 WER across the five languages, versus 14.1 for Whisper Large v3 and 16.6 for Omnilingual-1B. The pretraining balancing strategy earns its keep. Transfer to languages barely seen in pretraining is also strong: Georgian (93 hours) GigaAM 3.8 vs Omnilingual 7.8 and Whisper 13.4; Bashkir (143 hours) 3.6 vs 8.2 and 11.1.
For anyone building low-resource ASR, this is a reusable recipe. The bottleneck is not model size but how to keep head languages from drowning the tail in large-scale multilingual pretraining. Cluster-level balancing plus domain-aware sampling turns "too little data" from a dead end into a controllable engineering problem. The model and encoder are open-sourced (GitHub salute-developers/GigaAM) and can be fine-tuned directly on new low-resource languages.
The cost shows up in the results: lifting the tail pushes English WER from 14.4 to 15.4. Estimating per-language weights for all 70-plus languages is still infeasible, so balancing can only be done at cluster granularity. Uzbek is an exception where domain-aware sampling gives no clear benefit, which the authors attribute to its smaller dataset and lack of synthetic augmentation. Comparisons on tail languages also use the authors' internal test sets rather than a public benchmark, which limits strict comparability.