Phonon-2.5

Phonon-2.5

Transcribes speech in 25 languages with a 156.2 MB model that needs no graphics card. An open model from FermionResearch for Apple silicon Macs and ordinary Linux or Windows PCs.

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 is a compact, open model from FermionResearch that transcribes speech in 25 languages, and the whole download is 156.2 MB. It derives from NVIDIA's Parakeet TDT 0.6B v3 and is compressed to low-bit weights. One file works on Apple silicon and also on ordinary Linux and Windows processors, so a graphics card is not required.

Who it suits: People who want to turn recordings into text on their own computer, and developers who need a local transcription endpoint.

Industry | Who uses it | How they use it | Example
Media | Podcasters and editors | Transcribe interviews and episodes on their own machine | Turning a recording.wav file into text
Software | Developers | Run an OpenAI-compatible transcription endpoint on a local port | Starting the serve command on port 8010
Research | Speech researchers | Compare word error rates across many languages | Setting Phonon-2.5 beside Parakeet Redux

If you only need English, the card points to a separate English-only model of the same size.

Where to use it:

• Apple silicon Macs, on the GPU through MLX
• Linux on x86-64 and Arm, on the CPU
• Windows, on the CPU

Use case | Who | What they get back
File transcription | Anyone with audio | The spoken words as text
Word timing | Editors and developers | Per-word timestamps, each with a beginning and an end
Local API | Developers | An endpoint that speaks the OpenAI transcription format

How to set it up:

• Python 3.10 or newer, since the macOS built-in python3 is only 3.9
• The fermion-research package from pip
• On Apple silicon, the mlx extra to run on the GPU

1. Install Python 3.10 or newer.
2. Install the package with the first command below.
3. On Apple silicon, install the mlx extra with the second block instead.
4. On Linux, the card suggests installing the CPU build of torch first, to skip the GPU build.

pip install fermion-research

pip install "fermion-research[mlx]"

How to use it: You run the transcribe command with the model name and an audio file. The json flag adds beginning and end timestamps to each word, and the serve command starts an OpenAI-compatible transcription endpoint on a port you choose.

fermion transcribe phonon-2.5 recording.wav

fermion serve phonon-2.5 --port 8010

Tool | What it does
fermion transcribe | Turns an audio file into text
--json | Adds beginning and end timestamps to each word
fermion serve | Starts an OpenAI-compatible transcription endpoint

Example: You run the transcribe command on recording.wav. The spoken words come back as text, and with the json flag each word also carries its start and end time.

For newcomers and experts: Start by installing the package and transcribing one short recording to see the text it returns. No graphics card is needed.

Phonon-2.5 keeps the tokenizer and output conventions of the NVIDIA model it comes from, such as punctuation, casing and numerals. On the card's FLEURS test sets it averages 11.71 percent word error over 25 languages, and it is ahead of Parakeet Redux in 14 of them. On seven English sets from the Open ASR Leaderboard it averages 5.06 percent.

Install fermion-research
Run transcribe on an audio file
The model decodes the speech
Get text, with word times if asked

Title: Average word error over 25 languages, in percent
Phonon-2.5: 11.71%
Parakeet Redux: 12.18%
Parakeet TDT 0.6B v3: 13.26%

Title: Download size in megabytes
Phonon-2.5: 156.2 MB
Parakeet Redux: 178 MB
Parakeet TDT 0.6B v3: 2,508 MB

Good to know:

• License: CC-BY-4.0 for the model, and Apache 2.0 for the command line and this repository's code
• Cost to run: none named, and no GPU is needed
• Data it reaches: the audio files you give it, and a local endpoint if you start the serve command
• The larger Parakeet TDT 0.6B v3 model scores better in many rows of the card's table, English included

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: mlx
• Downloads in the last 30 days: 1,697
• Likes: 51
• License: cc-by-4.0
• Languages: en, es, fr, de, it, pt, nl, pl
• Based on: nvidia/parakeet-tdt-0.6b-v3
• Quantized versions: 3
• Fine-tunes: 2
• Tags: parakeet_tdt_five_value, apple-silicon, speech-to-text, multilingual, low-bit, on-device, automatic-speech-recognition

Links

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

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