Phonon-2.5 English

Phonon-2.5 English

Transcribes English speech to text in a 157.6 MB download that runs on Apple silicon, Linux and Windows CPUs without a GPU. It suits teams doing dictation, captions, compliance or search.

Made by FermionResearch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Phonon-2.5 English turns English speech into text. FermionResearch released it as an open model. It is built on NVIDIA's parakeet-tdt-0.6b-v3 and keeps that model's tokenizer and output conventions for punctuation, casing and numerals.

The download is 157.6 MB. The same download works on Apple silicon Macs and on Linux and Windows machines using only the CPU. A separate Phonon-2.5 of the same size covers 25 languages.

Who it suits: Teams with English-only workloads suit it best, such as dictation, captions, compliance and search pipelines.

Industry | Who uses it | How they use it | Example
Media | Caption teams | Turn English recordings into text for captions | Transcribe a wav file from the command line
Compliance | Compliance teams | Transcribe English audio for review | Add start and end times for every word
Search | Search teams | Index spoken English in a pipeline | Run a local transcription server for your apps

Anyone who needs other languages can skip it and look at the 25-language Phonon-2.5.

Where to use it: It runs from a terminal, and through a local server for your own apps.

• Apple silicon Macs, on the GPU through MLX
• Linux on x86-64 and Arm, on the CPU engine
• Windows, on the CPU engine
• Apps that speak the OpenAI transcription format, through the local server

Use case | Who | What they get back
Transcribing a recording | Anyone | The English text of the speech
Word timing | Compliance teams | The moment each word begins and ends
Local transcription service | Developers | An endpoint your apps can call

How to set it up: You need Python 3.10 or newer, since the built-in python3 on macOS is 3.9, and the fermion-research package.

• Python 3.10 or newer
• The fermion-research package, installed with pip
• A recording to transcribe

1. Install Python 3.10 or newer.
2. Install the fermion-research package with pip.
3. On Apple silicon, add the mlx extra to run on the GPU.
4. On Linux, put the CPU-only torch wheel in place before the package, so no GPU build is pulled in.

pip install fermion-research
fermion transcribe phonon-2.5-en recording.wav

Use the mlx extra on Apple silicon and the GPU runs the model through MLX.

pip install "fermion-research[mlx]"

How to use it: The transcribe command takes the model name and a recording. The json flag attaches timing to each word, giving its start and end. The serve command starts a local transcription endpoint on the port you choose.

fermion serve phonon-2.5-en --port 8010

Tool | What it does
fermion transcribe | Transcribes a recording with the model you name
--json | Adds the start and end moments of each word
fermion serve | Starts an OpenAI-compatible transcription endpoint

Example: You run fermion transcribe with phonon-2.5-en and a file called recording.wav. The English text of the recording comes back, and with the json flag each word carries its start and end time.

For newcomers and experts: Think of the model as a small listener that types out what an English speaker says. Your first small try is to record a short sentence, save it as a wav file and run the transcribe command on it.

Under the hood, the card scores it on seven English sets from the Open ASR Leaderboard. It averages 4.82 percent word error. Underdog Bark 3.0 averages 4.94 and Phonon-2 averages 5.21. The original Parakeet row in its table is a 2,508 MB download, against 157.6 MB here.

A recording goes in
The command line or local server takes it
The model runs on MLX or the CPU engine
English text comes out, with word times on request

Good to know: The card says Phonon-2.5 English beats Phonon-2 on all seven English sets. Its own table lists word error rates where lower is better.

Title: Average word error rate on seven English sets, lower is better
Phonon-2.5 English: 4.82%
Underdog Bark 3.0: 4.94%
Parakeet TDT 0.6B v3: 4.96%
Phonon-2: 5.21%
Parakeet Redux: 5.69%

• Language: English only
• Other versions: Hugging Face lists 4 quantized and 1 fine-tuned version made by other people
• License: CC-BY-4.0 for the model, Apache 2.0 for the command line and the repository code
• Cost to run: none named
• Data it reaches: the audio files you transcribe, and the model files downloaded from Hugging Face

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: mlx
• Downloads in the last 30 days: 1,349
• Likes: 22
• License: cc-by-4.0
• Languages: en
• Based on: nvidia/parakeet-tdt-0.6b-v3
• Quantized versions: 4
• Fine-tunes: 1
• Tags: parakeet_tdt_five_value, apple-silicon, speech-to-text, low-bit, on-device, automatic-speech-recognition

Links

• Model: https://huggingface.co/FermionResearch/Phonon-2.5-English
• Files: https://huggingface.co/FermionResearch/Phonon-2.5-English/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download FermionResearch/Phonon-2.5-English

Then follow the usage in its model card. Nothing in this kit runs by itself.