We beat Whisper Large v3 on LibriSpeech with a 634 MB model running entirely on

A developer has published an open-source Swift library claiming their speech recognition model — 634 megabytes running entirely on Apple Silicon — outperforms OpenAI's Whisper Large v3 on the LibriSpeech benchmark. The claim landed on March 22nd with enough specificity to turn heads: a named benchmark, a named competitor, a concrete model size, and a concrete platform constraint. LibriSpeech is a standard academic measure of transcription accuracy, and Whisper Large v3 has been the de facto reference point for on-device speech recognition quality. The library is framed as Swift-native, which would make it a first-class citizen on Apple's hardware rather than a ported afterthought. No independent replication has emerged yet.

If confirmed, here is what this means. Apple Silicon's neural engine has clearly not been fully exploited by the existing transcription ecosystem, and a Swift-native model beating the dominant open-source benchmark at roughly a third of Whisper Large v3's footprint would reframe the entire on-device AI conversation. Developers building iOS, macOS, and visionOS apps have been working around Whisper's awkward integration story for two years. A genuinely faster, smaller, native alternative would not just be convenient — it would change the default. More quietly, it would signal that benchmark leadership on constrained hardware is becoming achievable outside of well-resourced labs, which is the kind of thing OpenAI should notice.

Watch for independent developers reproducing the LibriSpeech numbers on their own Apple Silicon machines, and for any response — positive or skeptical — from the speech recognition research community.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →