FermionResearch/Phonon-2 — new model trending #29 on Hugging Face
Fermion Research released Phonon-2, a 164 MB English speech model averaging 5.21% word error.
Fermion Research published Phonon-2, an open English speech-recognition model trending on Hugging Face, with a 164 MB download and an average 5.21% word error rate across the seven English sets of the Open ASR Leaderboard. Its encoder stores weights at about 2.1 bits using five learned levels and matches much of the accuracy of its 2.5 GB full-precision Parakeet TDT 0.6B v3 teacher, which averages 4.96% WER. Reported throughput is about 174 times realtime on an M5 MacBook Air, 143 times on eight Zen 5 cores, and 6,680 times on one H100 at batch 128. The model is based on NVIDIA’s parakeet-tdt-0.6b-v3, licensed CC-BY-4.0, and is the engine in the Detta Mac dictation app.
- Phonon-2 is a 164 MB English ASR model averaging 5.21% word error on seven Open ASR sets.
- Its roughly 2.1-bit encoder reaches about the accuracy of a 2.5 GB Parakeet teacher.
- It transcribes about 174 times realtime on an M5 MacBook Air and 6,680 times on one H100.
- Weights are based on NVIDIA parakeet-tdt-0.6b-v3 and released under CC-BY-4.0.
- It ships inside the Detta Mac dictation app and via the fermion-research package and Docker images.
- Vendors
- Fermion Research
- Organizations
- Fermion ResearchNVIDIA
- AI models
- Phonon-2Parakeet TDT 0.6B v3
From the original article
2 min read at huggingface.coautomatic-speech-recognition · 2,126 downloads · 144 likes · license cc-by-4. 0
By FermionResearch · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.