Introduction

Phonon-2.5 is a new open speech recognition model from Fermion Research, the small lab behind the Detta dictation app for Mac. Fermion

, with Fermion's Manan Gupta
"a new accuracy-to-size Pareto frontier." The weights are live on Hugging Face, alongside an English-only sibling, Phonon-2.5 English.

The headline is size. Phonon-2.5 transcribes 25 languages from a 156.2 MB download. It's built from NVIDIA's Parakeet TDT 0.6B v3, a 2.5 GB model, and Fermion says the compressed version is actually more accurate on average. Quick heads-up: despite the shared name, this has nothing to do with Gradium's Phonon TTS models. That one makes speech, while this one turns speech into text.

The Numbers

Fermion scored all three models on the full FLEURS test set for each language, with the same harness and text cleanup. Word error rate, lower is better:

  • Phonon-2.5 (156.2 MB): 11.71% average across 25 languages.
  • Parakeet Redux (178 MB): 12.18%. This is the other low-bit model of similar size.
  • Parakeet TDT 0.6B v3 (2,508 MB): 13.26%. This is the full-precision original.

The original still wins on the big European languages like English, French and German. Phonon-2.5 pulls ahead on its average because the original stumbles on smaller ones. Greek is the clearest case at 20.83% for Phonon-2.5 against 36.11% for the original. Against Redux, Phonon-2.5 wins 14 of the 25 languages.

On English, across the seven Open ASR Leaderboard sets (LibriSpeech, AMI, Earnings-22, GigaSpeech, SPGISpeech and VoxPopuli), the model card reports:

  • Phonon-2.5: 5.06% average word error.
  • Phonon-2.5 English: 4.82%, same size, English only.
  • Phonon-2 (September): 5.21%.

What You Get

  • Languages: English, 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.
  • Hardware: the same file runs on Apple silicon (GPU through MLX) and on Linux or Windows CPUs. No GPU needed.
  • Speed: Fermion's docs say Phonon-2 runs at 174 times real time on an M5 MacBook Air, and Phonon-2.5 runs at about the same speed, with more speedups promised in the coming weeks.
  • Word timestamps: the --json flag returns start and end times for every word, which is exactly what you need for subtitles and caption timing.
  • Local server: fermion serve phonon-2.5 --port 8010 starts an OpenAI-compatible /v1/audio/transcriptions endpoint, so tools already written for OpenAI's speech-to-text API can point at your own machine.
  • License: weights are CC-BY-4.0, same as NVIDIA's original. The code and the command-line tool are Apache 2.0.

Why It Matters for Audio Creators

Transcription is the plumbing under a lot of creative audio work. You need it for podcast show notes, video subtitles, searchable interview archives, and for cutting voice-over by text. You also need it to build datasets for voice cloning and TTS fine-tunes. Most good multilingual models either live behind a paid API or need a real GPU.

A 156 MB model that runs on a laptop CPU, handles 25 languages and gives word-level timing changes that math. You can transcribe a whole season of a podcast offline, with no per-minute bill and no audio leaving your machine. Pair it with an open TTS model and you have a fully local voice pipeline.

Keep the limits in mind. All the benchmarks come from Fermion's own harness, the list covers European languages only (no Asian or Arabic languages), and the tokenizer and punctuation style are inherited from Parakeet.

How to Run It Today

You need Python 3.10 or newer. On a Mac, the built-in python3 is 3.9, so install a newer one first.

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

  • Apple silicon: install "fermion-research[mlx]" to run on the GPU through MLX.
  • Linux and Windows: the same package runs the CPU engine. On Linux, install CPU-only torch first with pip install --no-deps torch --index-url https://download.pytorch.org/whl/cpu.
  • File types: wav, flac, ogg and aiff work directly. For mp3 and m4a, the tool prints the ffmpeg command to convert them first.
  • First run: the engine takes 10 to 40 seconds to load, then transcribes at full speed.
  • No terminal: Fermion is rolling Phonon-2.5 into Detta, its Mac dictation app. You hold a key, speak, and the text appears in any field.

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

Conclusion

Phonon-2.5 is a strong example of compression doing real work. It's 16 times smaller than the model it came from, and it's still more accurate on average across 25 languages. It runs on hardware you already own, under a license that allows commercial use with attribution. If you transcribe, subtitle or build voice datasets in European languages, it's worth a test on your own audio this weekend. The thing to watch is the speedup Fermion says is coming in the next few weeks.

—Anabel ♡