mms-300m: 300M-parameter speech model trained on speech data from 1,400 languages

This is a wav2vec2-based model pre-trained on approximately 500,000 hours of speech data across more than 1,400 languages. With 300 million parameters, it provides a foundation for speech recognition in diverse linguistic environments.

This is not a ready-to-use inference model but rather pre-trained weights intended for fine-tuning. Additional training tailored to specific tasks such as automatic speech recognition, translation, or classification is required for deployment in production services. Input speech data must strictly adhere to a 16kHz sampling rate.

The broad range of supported languages makes it suitable for building speech processing pipelines for minority languages or multilingual environments. It is licensed under CC-BY-NC 4.0, so commercial use restrictions must be checked, and it is primarily used for speech technology development for research and non-profit purposes.