AI BriefingPhonon-2: Reproducing the speech recognition accuracy of a 2.5GB model with 164MB
FermionResearch/Phonon-2
About the project
Achieves 5.21% English speech recognition accuracy with an ultra-lightweight 164MB footprint. Compared to the original 2.5GB model, the download size is reduced to 1/15th, yet it maintains 100% or more of the original word accuracy even in complex environments such as meetings and parliamentary speeches.
A quantization technique with 5-stage learned weights compresses the model to approximately 2.1 bits. On an M5 MacBook Air, it processes one hour of audio in 20 seconds, and on an H100 GPU, it achieves 6,680x speed with batch processing.
It runs on Apple Silicon as well as CPU environments on Linux and Windows. It serves as the internal engine for the Mac dictation app Detta, allowing users to convert audio files to text with a single terminal command.
FermionResearch/Phonon-2
The original page has no description.
automatic-speech-recognition
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.