---
title: How to Run Phonon 2 Locally on CPU for Speech Recognition
url: "https://goremy.ai/blog/articles/phonon-2-speech-recognition-run-locally-cpu"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-01T11:40:06.662Z"
updated: "2026-11-01T11:40:06.662Z"
topic: Open Model Run Guides
description: "Run Fermion Research's 164 MB Phonon 2 ASR model on CPU only. Install steps, CPU timings, the OpenAI-compatible server, accuracy numbers and license notes."
---

# How to Run Phonon 2 Locally on CPU for Speech Recognition

> Run Fermion Research's 164 MB Phonon 2 ASR model on CPU only. Install steps, CPU timings, the OpenAI-compatible server, accuracy numbers and license notes.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-01

## How do you run Phonon 2 locally on a CPU, and how accurate is it?

Install the CPU build of PyTorch, then `pip install fermion-research`, and call the `fermion transcribe` command on an audio file. No GPU is required. Phonon 2 is a 164 MB English-only speech recognition model from Fermion Research, built on NVIDIA's Parakeet. Its model card reports a 5.21% average word error rate across the seven English sets of the Open ASR Leaderboard. In one CPU-only test, a short clip took about 50 seconds on first run and about 14 seconds once cached.

## TL;DR

- Phonon 2 is a **164 MB English-only ASR model** derived from NVIDIA's Parakeet, roughly 15 times smaller than the original download.
- It reports **5.21% average word error rate** across the seven English sets of the Open ASR Leaderboard.
- On CPU alone, the **first run took about 50 seconds** (download, checksum, runtime compile) and a **cached run about 14 seconds**, mostly model load.
- The CLI can also start an **OpenAI-compatible transcription server**, which answers quickly once the model is loaded.
- It handles clean read speech, narrative text and some noisy audio well, but **non-English audio performs poorly**.
- **Check the license** before commercial use. The newer Phonon-2.5 card lists CC-BY-4.0 for the model and Apache 2.0 for the CLI.

## What is Phonon 2?

Phonon 2 is a compressed speech recognition model from Fermion Research. It starts from NVIDIA's Parakeet and squeezes it into a download of 164 MB. The Phonon-2.5 model card describes the base as parakeet-tdt-0.6b-v3, a 2,508 MB full-precision model. It calls the result a low-bit model of a much smaller size.

Phonon 2 is tuned for English only. Fermion Research has since released Phonon-2.5, a 156.2 MB model covering 25 languages, plus a Phonon-2.5 English variant at the same size. If you need other languages, look at 2.5 rather than 2.

## How do you install Phonon 2 on a CPU?

You need Python 3.10 or newer. On Linux, install the CPU wheel of PyTorch first so pip does not pull the much larger GPU build. Then install the Fermion package.

```bash
pip install --no-deps torch --index-url https://download.pytorch.org/whl/cpu
pip install fermion-research
```

Depending on your setup, the CLI may not be on your PATH after install. If the `fermion` command isn't found, add the pip script directory to PATH. The same package runs a CPU engine on Linux (x86-64 and Arm) and Windows. On Apple silicon, the `[mlx]` extra runs the model on the GPU through MLX instead.

The CLI has a command that lists the models available from Fermion. Use it to confirm the exact name for Phonon 2. The model card for 2.5 uses the pattern below, so swap in the name your CLI shows.

```bash
fermion transcribe phonon-2.5 recording.wav
```

Add `--json` to get a start and end time for every word.

## How fast is Phonon 2 on CPU only?

A creator who tested it on an Ubuntu machine with no GPU load timed a short English clip twice with the shell's `time` command.

| Run | Wall time | What happened |
|---|---|---|
| First | about 50 seconds | Downloaded the model, checked its checksum, compiled the CPU runtime |
| Second | about 14 seconds | Used the cached model |

The 14 seconds is mostly engine startup. Loading the model takes roughly 10 to 14 seconds, so decoding itself is a small slice of that. This is one machine and one short clip, so treat the numbers as a rough guide. Your CPU will change them.

The practical lesson is to avoid launching a fresh process per file. If you transcribe many clips, keep the model loaded, which is what the server mode does.

## How do you serve Phonon 2 with an OpenAI-compatible endpoint?

The CLI includes a serve command. It exposes an OpenAI-compatible `/v1/audio/transcriptions` endpoint, so existing client code that targets OpenAI's transcription API can point at your local machine instead.

```bash
fermion serve phonon-2.5 --port 8000
```

The model card's example uses port 8010 and the demo used port 8000. The `--port` flag sets it, so pick whatever is free. The first launch takes a while because the model has to load. After that, you can hit the health check from a second terminal, which returned an OK status in the demo, and send audio for transcription. Requests after the load were described as very fast.

Full CLI and server options are documented at fermionresearch.com/docs/speech.

## How accurate is Phonon 2?

On the Open ASR Leaderboard's seven English sets (LibriSpeech clean and other, AMI, Earnings-22, GigaSpeech, SPGISpeech and VoxPopuli), Phonon 2 averages 5.21% word error rate. For comparison, the model card for its successor reports:

| Model | Open ASR English average WER |
|---|--:|
| Phonon 2 | 5.21% |
| Phonon-2.5 | 5.06% |
| Phonon-2.5 English | 4.82% |

Within 2.5's per-set results, the spread is wide. LibriSpeech clean sits at 1.60%, while AMI, which is meeting speech, reaches 9.12%. Expect clean read speech to be much better than conversational or far-field audio.

Informal tests matched this. The model transcribed a short LibriSpeech-style clip word for word. It also handled an 18-second passage about decorative art, a narrative sentence, and a line containing the unusual name "Polychrome". The tester then tried a 25-second recording of their own with mediocre audio quality, a hard short clip where the opening was unclear, and an emotional monologue. Results on those were described as good. That is anecdotal, not a benchmark, but it fits the leaderboard numbers.

## Does Phonon 2 support other languages?

No. Phonon 2 is English only. The tester tried several European languages, Arabic, Hindi, Urdu, Indonesian and Malay, and found it did not work well. A few European languages came out right here and there, but not reliably.

If you need multilingual output, Phonon-2.5 covers 25 languages. Its card reports an 11.71% average word error rate on FLEURS across those languages, against 12.18% for Parakeet Redux and 13.26% for the full Parakeet TDT 0.6B v3. English on FLEURS is 6.58% for Phonon-2.5. Quality varies by language, and Greek, Danish and Slovenian have the highest error rates among those listed.

## Can you use Phonon 2 commercially?

Check the license first. The Phonon-2.5 card says the model is CC-BY-4.0, the same as the Parakeet original, with a `NOTICE` file listing changes. The CLI and repository code are Apache 2.0. CC-BY-4.0 generally permits commercial use with attribution, but that is the 2.5 card. Read the license on the exact model you download, and keep the attribution and notice requirements.

## Is Phonon 2 worth using?

It makes sense if you want local, private English transcription on modest hardware and can tolerate a startup cost per process. The small download and CPU-only operation make it easy to try on a laptop or a cheap VM. If you need multiple languages, or the best English accuracy available from this family, start with Phonon-2.5 or its English variant. If you need the strongest accuracy regardless of size, the full 2.5 GB Parakeet is still ahead on most languages in the card's table, though by small margins for English.

## Frequently Asked Questions

### Do I need a GPU to run Phonon 2?
No. The model runs on CPU only, and the demo showed no GPU activity. Install the CPU wheel of PyTorch to avoid pulling the GPU build. On Apple silicon you can optionally use MLX to run on the GPU.

### Why did my first transcription take so long?
The first run downloads the model, checks its checksum and compiles the CPU runtime. That took about 50 seconds in testing. Later runs used the cached model and took about 14 seconds, most of it model load.

### Is Phonon 2 compatible with the OpenAI API?
The CLI's serve command exposes an OpenAI-compatible `/v1/audio/transcriptions` endpoint. You choose the port with `--port`. Point any client that speaks the OpenAI transcription format at your local address.

### Should I use Phonon 2 or Phonon-2.5?
Use Phonon 2 if you only need English and want the model covered here. Phonon-2.5 is smaller at 156.2 MB, supports 25 languages and reports a slightly lower English average of 5.06%. Phonon-2.5 English reports 4.82%.