Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking (syfi.cs.washington.edu)

🤖 AI Summary
Murmur is a newly introduced system that enhances long-context automatic speech recognition (ASR) by using a hierarchical chunking approach, enabling efficient inference of audio recordings up to an hour long. It achieves an average speedup of 2.5 times (up to 2.9 times) compared to single-pass methods, while significantly lowering the time-constrained word error rate (tcpWER) by 7 to 51 points when transcribing English and Mandarin meetings. This is particularly significant as existing long-context models often suffer from slow processing and inaccuracies due to repetition loops or difficulties in stitching together outputs from smaller chunks. The Murmur system intelligently splits audio into chunks of around five minutes, while simultaneously using a lightweight ASR model to create draft transcriptions for shorter 30-second segments. This dual approach allows Murmur to verify and refine those drafts with the long-context model, enhancing speed without sacrificing accuracy. The results indicate that Murmur consistently outperforms existing models like WhisperX and retains a competitive edge against single-pass systems, marking a notable advancement in the performance and efficiency of long-context ASR technologies.
Loading comments...
loading comments...