Cohere Transcribe: Open-source 2B speech recognition model (huggingface.co)

Cohere has announced the launch of its new open-source speech recognition model, Cohere Transcribe, which has claimed the top position on the Hugging Face Open ASR Leaderboard for English and performs comparably or better across 13 other languages. Designed with a production focus, this 2B parameter encoder-decoder transformer model leverages optimized efficiency and accuracy, utilizing a Fast-Conformer encoder and a lightweight decoder. The model's architecture prioritizes real-time processing speed without sacrificing performance, achieving significant gains in both benchmark tests and real-world evaluations. The significance of Cohere Transcribe lies in its competitive edge over existing proprietary and open-source models, marking a substantial advancement in multilingual speech recognition technology. With a robust dataset of 0.5 million hours of curated audio-transcript pairs and innovative enhancements for serving variable-length audio efficiently, the model aims to ensure scalable deployment suitable for diverse applications. Additionally, the incorporation of a customizable punctuation feature and background noise management underscores its versatility. This release not only strengthens Cohere's portfolio in audio technologies but also offers the AI/ML community a state-of-the-art tool for transcription tasks, fostering further advancements in natural language processing and understanding.