
This is a 0.5B parameter open-source model for speech recognition that supports 40+ languages. It runs on standard Transformers or vLLM stacks without requiring custom code or forks. The model achieves a 6.96% average error rate, outperforming larger competitors like Meta's 7B omniASR-LLM and Whisper large-v3.
It is worth testing if you need high-throughput transcription on modest hardware, as it processes an hour of audio in seconds. The main limitation is its focus on standard phone and clean audio, which may not suit specialized or highly noisy industrial environments. Avoid it if your workflow requires non-English languages not in its 40-language support list.
more: https://huggingface.co/blog/rumble-2-flash
my other side: ivo.blogspot.com
Leave a Reply