BuzzASR is a collection of 102 language-specialized fine-tuned Whisper models. These models were adapted for automatic speech recognition (ASR) in various languages. The models were created using a massively scaled approach to simple fine-tuning on monolingual data. This approach covered 102 languages from the FLEURS dataset, incorporating a complex strategy with tokenizer replacement and data augmentation.
Evaluation results show that the BuzzASR models outperform Whisper-large-v3 on 77 out of 102 languages. The average character error rate (CER) reduction is over 2.8 times. The models achieve state-of-the-art CER among open-source systems on 27 of 102 languages on the combined FLEURS and Common Voice test set.
The tokenizer replacement strategy results in an average 3.3x improvement in compression rate (characters per token) compared to Whisper's multilingual BPE. This strategy achieved gains of up to 21.7x. The project releases all models, code, and detailed results at https://lemn-lab.github.io/buzz-asr.
This work addresses the performance limitations of large, multilingual ASR models. The project provides a scalable solution for language-specific ASR, offering improved accuracy and compression for a wide range of languages.
Source: https://arxiv.org/abs/2609.09554