Fermion drops Phonon-2, a 164MB speech model that beats giants

Phonon-2 is a featherweight open-source English speech recognition model from Fermion Research, weighing in at just 164MB for the complete download. It is built specifically to transcribe speech right on your local device without eating up your hard drive or bogging down your processor.

It is the exact engine that powers Detta, the local dictation app for macOS. If you want fast local transcription without relying on cloud services or multi-gigabyte models, this is designed for your workflow.

What it does

Despite shrinking down to 164MB, Phonon-2 gives up almost nothing on recognition accuracy.

It derives from NVIDIA's 2.5GB parakeet-tdt-0.6b-v3 base model, cutting down the download footprint by roughly 15 times. Across the seven English benchmark sets on the Open ASR Leaderboard, Phonon-2 hits an average word error rate of 5.21%. That is remarkably close to the original base model's 4.96%. On parliamentary speeches, it matches 100.8% of the base model's word accuracy, and on meeting audio, it actually outperforms the original 2.5GB model.

Compared to much larger alternatives, Phonon-2 holds a clear edge. Whisper large-v3-turbo sits at 1,618MB with a 6.58% error rate, Voxtral Mini 4B Realtime takes roughly 8,000MB with a 6.12% error rate, and the previous 415MB Phonon-1 lands at 6.56%. Under 900MB, Phonon-2 currently claims the lowest error rate among open English ASR models.

Inference speed is blisteringly fast on modest personal hardware. On an M5 MacBook Air, it can process one hour of audio in about 20 seconds, which translates to roughly 174 times realtime. It clocks 143 times realtime on eight Zen 5 CPU cores (16 vCPU), and scales up to 6,680 times realtime on a single NVIDIA H100 with batching.

How to run it

Setting up Phonon-2 locally is straightforward across multiple operating systems.

On Apple silicon, you can install the python package with MLX support to tap into native unified memory and compute. A dedicated Core ML version is also available for the Neural Engine. Passing a simple flag yields word-level timestamps in JSON format for subtitles or search indexing.

Beyond macOS, Fermion Research provides Android arm64-v8a prebuilts with C headers, along with support for standard Linux and Windows setups. Prebuilt Docker containers are available for both CPU and CUDA environments for instant server deployment.

On licensing, the model weights retain the original CC-BY-4.0 license, while the repository code and CLI tools are released under Apache 2.0. A free demo is hosted on Hugging Face Spaces if you want to test audio files before pulling the repo.

Why it matters

We think Phonon-2 is a textbook example of smart model compression done right. For years, developers building local transcription into mobile or desktop apps had to choose between inaccurate tiny models or heavy Whisper models that chew through system memory and battery.

If you build dictation apps, local meeting assistants, or podcast tools, Phonon-2 is an immediate no-brainer to test. It turns offline transcription into a trivial background task even on entry-level laptops. The only clear catch is scope: it is laser-focused on English audio, so multilingual pipelines will still need to look elsewhere.