GigaAM-Multilingual

GigaAM Multilingual is a family of Conformer-based foundation models (220M / 600M parameters) pre-trained with a HuBERT-style objective on 2M hours of speech across 70+ languages and fine-tuned for speech recognition with character-wise CTC decoders on 50K hours.

The models provide best-in-class open-source quality on Russian, Kazakh, Kyrgyz, and Uzbek, and moderate quality on English.

GigaAM Multilingual includes the following model variants:

• ssl — 220M self-supervised encoder
• ctc — 220M ASR model with a character-wise CTC decoder
• large_ssl — 600M self-supervised encoder
• large_ctc — 600M ASR model with a character-wise CTC decoder

Model Performance

Word Error Rate (%) on Common Voice (CV), FLEURS, and internal in-the-wild test sets. Utterances longer than 30 s and references containing digits are excluded; references/hypotheses are normalized (lowercasing, punctuation removal, numerals→words); greedy decoding. Best per row in bold.

Usage

from transformers import AutoModel

revision = "ctc"  # any variant: ssl, ctc, large_ssl, large_ctc
model = AutoModel.from_pretrained(
    "ai-sage/GigaAM-Multilingual",
    revision=revision,
    trust_remote_code=True,
)

transcription = model.transcribe("example.wav")
print(transcription)

Recommended versions:

• torch==2.10.*, torchaudio==2.10.*
• transformers==5.*
• (any) hydra-core, omegaconf

Full usage guide can be found in the example.

Fine-tuning to a new language

The ssl / large_ssl backbones can be adapted to a new language — see the fine-tuning guide and the example notebook.

Citation

@misc{gigaam_multilingual,
      title={GigaAM Multilingual: Foundation Model for Underrepresented Languages},
      author={Andrei Kuzmenko and Alexandr Maximenko and Aleksandr Kutsakov and Georgii Gospodinov and Dmitrii Bolotov and Oleg Kutuzov and Pavel Bogomolov and Fyodor Minkin},
      year={2026},
      eprint={2607.10371},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2607.10371}
}