New AI Speeds Up Meeting Transcriptions by 2.5x
Tired of waiting for meeting notes? This AI cuts transcription time by 60%.
Researchers propose Murmur, an inference system for long-form automatic speech recognition built on hierarchical chunking. The article says a long-context ASR model decodes intermediate target chunks in parallel, while a fast short-window ASR on 30-second draft windows accelerates each chunk with lossless, structure-aware speculative decoding; simple format rules draft timestamps and speaker tags, and the long-context model verifies every token.
On two long-context ASR models and four meeting datasets in English and Mandarin, Murmur cuts end-to-end latency per recording by 2.5x on average (1.9–2.9x) relative to single-pass decoding, including diarization, while lowering tcpWER by 1.9 points on average. The article also reports that an intermediate chunk size of about five minutes matched single-pass accuracy while avoiding repetition loops, and says the code is available at https://github.com/uw-syfi/Murmur.
- Murmur transcribes long recordings 2.5 times faster than previous methods.
- It uses a two-step process: a fast draft and a careful check, ensuring accuracy.
- The system is not yet available to the public, but the code is open for developers.
Why It Matters
Faster transcription means quicker access to meeting notes, saving time and boosting productivity for professionals.