Phonon 2: A 164MB Speech Recognition Model That Runs on Any CPU
Phonon 2 is a 164MB CPU-only speech-to-text model built on Nvidia Parakeet. Here's how to install it and how it performs in real tests.
What is Phonon 2?
Phonon 2 is a compact automatic speech recognition (ASR) model from Fermion Research that compresses Nvidia’s Parakeet speech model down to roughly 164MB, about 15 times smaller than the original, while running entirely on CPU with no GPU required. According to its model card, it averages a 5.21% word error rate across the seven English test sets of the Open ASR leaderboard, putting it in the same performance range as models many times its size.
TL;DR
- Phonon 2 is a distilled, heavily compressed version of Nvidia’s Parakeet ASR model, packaged by Fermion Research to run on CPU-only hardware with a download size around 164MB.
- The model card reports a 5.21% average word error rate across the seven English datasets in the Open ASR leaderboard, a strong result for a model this small.
- It installs through a dedicated Phonon CLI alongside a CPU build of PyTorch, and the first run is slower because it downloads the model, verifies a checksum, and compiles a CPU runtime.
- Once cached, transcription on a short clip took about 14 seconds, most of which was engine startup rather than actual decoding time.
- In hands-on tests it handled formal narration, period literary prose, poor-quality personal recordings, and emotionally inflected speech, getting transcripts right word for word in several cases.
- The current release tested here is English-only; a related Fermion Research model, Phonon 2.5, adds multilingual support across more than two dozen languages built on Nvidia’s parakeet-tdt-0.6b-v3 base.
- The model ships under a CC-BY-4.0 weights license with Apache 2.0 code, so anyone considering commercial use should check license terms before deploying it.
Remy.
How does Phonon 2 achieve such a small footprint?
Phonon 2 starts from Nvidia’s Parakeet architecture, an ASR model already covered for its accuracy relative to size. Fermion Research takes that base and compresses it aggressively, likely through a combination of quantization and distillation techniques, down to a package around 164MB. That is small enough to sit comfortably on a laptop, a Raspberry Pi class device, or a cheap cloud VM with no GPU attached.
The tradeoff with this kind of compression is usually accuracy. Phonon 2’s model card claims an average word error rate of 5.21% across the seven English sets in the Open ASR leaderboard, a widely used benchmark suite for comparing speech recognition systems. That number sits in competitive territory for a model that fits in a fraction of the storage most ASR systems need, which is the entire pitch here: near full-size accuracy, a small fraction of the size.
The related Phonon 2.5 release builds on Nvidia’s parakeet-tdt-0.6b-v3 and pushes further into multilingual territory, supporting languages including Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Romanian, Czech, Slovak, Bulgarian, Croatian, Swedish, Danish, Finnish, Estonian, Latvian, Lithuanian, Hungarian, Slovenian, Maltese, and Greek, alongside English. That version also ships with an MLX build, meaning it is optimized to run efficiently on Apple Silicon in addition to standard CPU inference.
How do you install and run Phonon 2 locally?
Getting Phonon 2 running locally is a short process on a Linux machine like Ubuntu:
- Install PyTorch for CPU. Fermion Research also offers a GPU and CUDA build if you want to run the model on a GPU instead, but the entire point of Phonon 2 is that you don’t need one.
- Install the Phonon CLI and its dependencies. This step is quick and installs cleanly through standard package management.
- Check your PATH. You may need to add the CLI’s install location to your shell path before the
phononcommand is recognized. - List available models. Running the CLI’s list command shows every model Fermion offers; Phonon 2 is one of several, but it’s the lightweight CPU-first option.
- Run a transcription. Point the CLI at an audio file and it produces a text transcript.
The first time you run a transcription, the CLI downloads the model weights, verifies a checksum, and compiles a CPU runtime tailored to your machine. That setup process took around 50 seconds in testing. Every run after that uses the cached model and compiled runtime, cutting the time down to about 14 seconds for a short clip, most of which is model loading rather than actual speech decoding.
Phonon 2 can also run as a local server with an OpenAI-compatible API endpoint, which means tools built against OpenAI’s transcription API format can point at it with minimal changes. Spinning up the server takes a bit of extra time on first load, after which you can hit a health check endpoint to confirm it’s running and start sending transcription requests. Requests to an already-running server are noticeably faster than the CLI’s cold-start workflow.
How accurate is Phonon 2 in real-world audio tests?
Across a range of test clips, Phonon 2 performed consistently well on English audio:
- A short, clearly spoken clip about “the general principles of art” was transcribed word for word.
- An 18-second clip with more complex vocabulary and sentence structure, referencing art history and Michelangelo, came through accurately.
- A narrative-style passage with descriptive, literary language was transcribed correctly.
- A passage including the invented name “Polychrome” (from classic fantasy literature) was transcribed correctly despite the unusual proper noun, a reasonable stress test given the model’s tiny size.
- A personal recording with noticeably lower audio quality was still transcribed well, correctly capturing phrases like “the difference is whether the stillness comes from peace or from exhaustion.”
- A short, emotionally charged clip with repeated, overlapping speech (“you’ve hit me four times now”) was transcribed correctly despite an unclear start.
- A longer emotional monologue about receiving good news after years of effort was also transcribed accurately, including natural speech patterns and a mid-sentence restart.
The model struggled outside English. Testing against several European languages, as well as Arabic, Hindi, Urdu, Indonesian, and Malay, produced inconsistent results, occasionally correct for European languages but unreliable overall. That’s expected: this version of Phonon 2 is fine-tuned specifically for English, not multilingual use. Anyone needing broader language coverage should look at the newer Phonon 2.5 release instead.
Is Phonon 2 worth running locally?
For English-only, CPU-bound transcription, Phonon 2 is a strong option if your priority is small footprint and no GPU dependency. A 164MB model that delivers accurate transcripts on noisy personal recordings, literary prose, and emotional speech, without needing specialized hardware, is useful for local tooling, embedded devices, edge deployments, or any situation where shipping a multi-gigabyte ASR model isn’t practical.
It’s not the right choice if you need multilingual support out of the box (Phonon 2.5 covers that) or if you’re building a product for commercial distribution without checking license terms first. The weights are released under a CC-BY-4.0 license with code under Apache 2.0, which is permissive but still requires attribution and review against your specific use case.
Frequently Asked Questions
What is Phonon 2 built on?
Phonon 2 is built on Nvidia’s Parakeet speech recognition model, compressed and optimized by Fermion Research to run entirely on CPU hardware at a fraction of the original size.
How big is the Phonon 2 model file?
The model download is approximately 164MB, about 15 times smaller than the Parakeet model it’s derived from.
Does Phonon 2 support languages other than English?
This version of Phonon 2 is English-only. Fermion Research’s related release, Phonon 2.5, built on Nvidia’s parakeet-tdt-0.6b-v3, adds support for more than two dozen additional languages and includes an MLX build for Apple Silicon.
Do I need a GPU to run Phonon 2?
No. Phonon 2 is designed to run on CPU only. A GPU and CUDA build is also available from Fermion Research for those who prefer it, but it isn’t required.
How accurate is Phonon 2 compared to larger ASR models?
- Patient records✓ Granted
- Surgery schedule✓ Granted
- Billing export✕ Denied
Its model card reports a 5.21% average word error rate across the seven English datasets in the Open ASR leaderboard, a competitive score for a model of this size, and it performed reliably in hands-on tests across clean speech, noisy recordings, and emotionally varied audio.