Reson8 open-sources Rumble-2-Flash, a 0.5B speech recognition model for 40+ languages
Rumble-2-Flash: an open 0.5B speech model for 40+ languages
Reson8 open-sourced Rumble-2-Flash under Apache 2.0, a 0.5B-parameter speech recognition model covering 40+ languages with automatic language detection and native support in Transformers and vLLM. It averages 6.96% error on the Open ASR Leaderboard multilingual track, ahead of Whisper large-v3 and Meta's omniASR-LLM 7B, and reaches 7.86% WER on AppTek's 302-hour call-centre benchmark, on par with NVIDIA's Nemotron 3.5 Transcribe (30B). The model was distilled from the larger Rumble-2 using on-policy distillation with GOLD, which aligns differing tokenizers between teacher and student.
A 0.5B speech recognition model that matches a 30B model on noisy call-centre audio shows how far on-policy distillation can shrink production ASR.
Source:Hugging Face Blog · huggingface.co