Phonon-2
Open English speech recognition model, 164 MB, runs locally on Mac, Linux, Windows and GPUs.
About Phonon-2
Phonon-2: Open Speech Recognition in a 164 MB Download
Phonon-2 is an open speech recognition model for English from Fermion Research, released as a single 164 MB file that runs on Macs, Linux, Windows and NVIDIA GPUs. It averages 5.21% word error on the seven English test sets of the Open ASR Leaderboard and writes punctuated, capitalized text, from an hour of speech in about 20 seconds on a MacBook Air.
Key Features
Compact Model File: A 164 MB download (177 MB on disk), with an encoder that stores each weight in about 2.1 bits.
Benchmark Accuracy: 5.21% average word error on the leaderboard's seven sets, against 4.96% for its 2,508 MB full-precision teacher, NVIDIA's Parakeet TDT 0.6B v3.
Fast on a Laptop: 606 times realtime on the Neural Engine of an M5 MacBook Air with Phonon-2 Core ML, and 174 times realtime on the same Mac's GPU.
Server and GPU Throughput: About 143 times realtime on eight Zen 5 cores, and 6,680 times realtime on one NVIDIA H100 in batches of 128.
Open Weights: Weights under CC-BY-4.0, inherited from Parakeet TDT 0.6B v3; the command line and public code are Apache-2.0.
One-Line Install:
pip install fermion-research, then transcribe a file or dictate live in the terminal; the first run downloads and verifies the model.OpenAI-Compatible Server: Serve an OpenAI-compatible API locally, with a WebSocket streaming API and sentence-level segment timestamps.
Runs Locally: The model and command line run on your own hardware and send nothing back to Fermion Research.
Use Cases
On-Device Dictation: Power dictation on the Mac, as in Detta, the Fermion Research dictation app built on Phonon-2.
Batch Transcription: Transcribe large audio archives on a single GPU or on CPU servers.
Private Speech Pipelines: Keep audio on your own machines by running speech recognition locally instead of calling a cloud API.
Voice App Prototyping: Swap a local OpenAI-compatible endpoint into existing speech-to-text code.
Getting Started
Website: https://www.fermionresearch.com
Documentation: https://www.fermionresearch.com/docs/
Model weights: https://huggingface.co/FermionResearch/Phonon-2
Phonon-2 gives developers an open, English speech recognition model small enough to ship inside an app and fast enough for batch transcription on a single GPU.
