Meta returns to open source AI with Omnilingual ASR models that can transcribe 1,600+ languages natively

Meta's recent release of the Omnilingual ASR represents a major leap in automatic speech recognition technology, supporting over 1,600 languages natively, which is a substantial increase compared to OpenAI's Whisper model that only accommodates 99 languages. This new system not only enhances language support but also introduces a feature called zero-shot in-context learning, enabling users to provide examples of audio and text in new languages for transcription without needing retraining. This flexibility allows the model to potentially cover more than 5,400 languages, making it the most extensible speech recognition system available. By open-sourcing the Omnilingual ASR under the Apache 2.0 license, Meta empowers developers and communities to adapt and expand the system's capabilities, fostering innovation and accessibility in multilingual communication.