Hugging Face user comodoro has drawn more than 860,000 monthly downloads for a Slovak checkpoint built from Meta’s XLS-R 300M model. The Apache 2.0-licensed release gives developers a self-hosted option for Slovak speech recognition without collecting labeled audio and training a model from scratch.

860,000 downloads in context

Hugging Face calculates its displayed download figure from qualifying file requests over a rolling 30-day window. Automated deployments, cache misses, continuous integration jobs, and repeated installations can each generate requests, so the figure does not represent unique users or lifetime installations. It does show sustained retrieval activity around a language-specific community model.

Slovak has less labeled speech data and fewer open checkpoints than widely supported languages such as English. This release packages a Slovak acoustic model, tokenizer, processor, and CTC decoding configuration behind the standard transformers interface.

How XLS-R becomes Slovak

Meta’s Facebook AI team pretrained XLS-R on about 436,000 hours of unlabeled speech spanning 128 languages. Pretraining teaches the encoder to extract useful speech representations from raw waveforms. Fine-tuning then maps those representations to Slovak character tokens using the labeled Common Voice recordings.