Phonon-2: Fast Local Speech-to-Text for Apple Silicon
Phonon-2 is a 164 MB speech-to-text model built specifically for Apple Silicon. Developed by Fermion Research, it runs locally through Apple’s MLX framework and is designed around one clear goal: transcribe English audio as quickly as possible without sending anything to the cloud.
Its headline claim is 174x real-time transcription speed. In practical terms, that means a one-hour audio file can be transcribed in roughly 21 seconds.
That speed comes with a deliberate trade-off. Phonon-2 supports only English. By narrowing the problem instead of trying to handle dozens of languages, Fermion can keep the model small and optimize it more aggressively for Apple hardware.
This article looks at how Phonon-2 achieves that size and speed, how to install and use it, and how its performance holds up outside the company’s own benchmarks. It also compares those results with independent tests and looks at where Phonon-2 makes more sense than broader speech-to-text alternatives.
How Phonon-2 got to 164 MB
Both Phonon-2 and an earlier model called Parakeet Redux started from the same parent: NVIDIA's Parakeet TDT 0.6B v3, a 2.5 GB model. Both compress it to roughly 178 MB, but they took different approaches.
Parakeet Redux (from developer Moondream) uses 1.58-bit ternary weights: each parameter is represented as -1, 0, or +1. This produces maximum compression while keeping the model multilingual across 25 languages, spreading capacity across all of them.
Phonon-2 uses a five-value alphabet instead: each weight is quantized to one of five learned values (approximately -0.720, -0.284, 0, +0.284, +0.720), roughly 2.1 bits per weight, with remaining parameters in 6-bit integer tables. Same disk size, but different trade-off: Phonon-2 drops all non-English languages and focuses its entire 178 MB on English.
The result is a model that's smaller than OpenAI's Whisper base (145 MB when Phonon-2 was released at 164 MB) but aims for higher accuracy on English specifically.
Installation and usage on Apple Silicon
Phonon-2 requires Python 3.10+ and an Apple Silicon Mac. Install it with the MLX extras:
Transcribe a file:
The first run downloads the ~164 MB model weights from Hugging Face and caches them locally. Subsequent runs use the cache. Output includes a full transcription with precise start and end timestamps for each spoken segment.
Speed: 174x real-time, with caveats
At 174x real-time, a one-hour file transcribes in about 21 seconds. The speed comes from two things: MLX leverages Apple Silicon's unified memory architecture, and the default "fast-path decoder" uses an approximation algorithm that prioritizes speed over exact computation.
Switching to --exact-decode drops speed to 109x real-time but increases accuracy. Running CPU-only drops further to about 40x real-time. All three modes are still substantially faster than real-time.
The "6x faster than Parakeet Redux" claim in Fermion's benchmarks uses Redux running through a third-party Core ML port called FluidAudio that appears to be a performance bottleneck. Redux's own developers report significantly faster speeds, which narrows the gap considerably. The speed advantage over Redux specifically should be verified on your hardware.
Accuracy: good, but read the independent numbers
Fermion's own benchmarks report:
| Model | WER (Fermion) |
|---|---|
| NVIDIA Parakeet (parent) | 4.96% |
| Phonon-2 | 5.21% |
| Parakeet Redux | 5.69% |
| Whisper large-v3-turbo | 6.58% |
The Open ASR Leaderboard, an independent testing group, measured different results:
| Model | WER (independent) |
|---|---|
| NVIDIA Parakeet (parent) | 6.51% |
| Phonon-2 | 7.23% |
| Whisper Turbo | 8.39% |
Phonon-2 is still more accurate than Whisper Turbo in both sets of tests, which is the meaningful comparison for most users. The independent tests also found Whisper performs better than Phonon-2 on audio with significant background noise. For clean audio, Phonon-2 leads; for noisy environments, the gap narrows or reverses.
Features beyond file transcription
Live captions from your microphone:
Starts transcribing from the microphone in real time, printing results to the console.
Hotwords for technical vocabulary:
Increases the likelihood that specific terms get transcribed correctly, useful for domain-specific audio with proper nouns or jargon.
Local OpenAI-compatible server:
Starts a server at http://127.0.0.1:8000 with two endpoints: a REST endpoint at /v1/audio/transcriptions that matches OpenAI's Whisper API shape, and a WebSocket endpoint for streaming real-time transcription.
Applications that currently call OpenAI's Whisper API can switch to a local Phonon-2 instance by changing the base URL. The payload format is identical.
Licensing
The CLI and Python code are Apache 2.0. The model weights are CC-BY-4.0, inherited from NVIDIA's Parakeet license. Commercial use is allowed in both cases; attribution to NVIDIA and Fermion Research is required for the weights.
When to use it and when not to
Phonon-2 is the right choice for English transcription on Apple Silicon where you want offline processing, no API costs, and the lowest possible latency. The OpenAI-compatible server makes it a straightforward drop-in for applications already built against Whisper's API.
Use something else when you need multilingual support (Parakeet Redux handles 25 languages at the same size), when your audio has heavy background noise (Whisper Turbo holds up better there), or when multiple overlapping speakers are present in meeting audio (Redux has a slight edge in some tests). And as with any model, vendor benchmarks should be treated as a starting point rather than a final answer: running Phonon-2 against a sample of your own audio is the only reliable way to know how it will perform on your specific use case.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.