Introducing Phonon-2.5

Phonon-2.5 is an open speech recognition model for 25 languages in a 156 MB download. On the FLEURS test sets it averages 11.71 % word error across the 25 languages, against 12.18 % for Parakeet Redux, the other low-bit model of the same size, and 13.26 % for Parakeet TDT 0.6B v3, the 2.5 GB full-precision model it derives from. It is ahead of Redux in 14 of the 25 languages and runs on a laptop CPU with no GPU.

We are also releasing Phonon-2.5 English, an English-only model of the same size. Across the Open ASR Leaderboard’s seven English sets it averages 4.82 % word error, against 4.94 % for Underdog Bark 3.0, the other low-bit model of this size, and 5.21 % for Phonon-2. It is better than Phonon-2 on all seven sets, and it is our most accurate English model.

Both models keep the extreme low-bit Phonon-2 format: the same file runs on Mac, Linux (x86-64 and Arm) and Windows CPUs with no GPU, on NVIDIA GPUs and in the CPU and CUDA Docker images, and the weights are released under CC-BY-4.0, the licence of NVIDIA’s Parakeet TDT 0.6B v3.

25 languages

A word error rate counts the words a model gets wrong, as a share of the words spoken: 10 % means one word in ten is missing, added or different. FLEURS is a public test set of read sentences in each language, and Table 1 scores all three models on the full test set of each one with the same code.

Phonon-2.5 is the lowest of the three in 6 languages and within one point of the lowest in most of the rest. The largest gap is Greek, where Phonon-2.5 scores 20.83 % against 33.25 % for Redux and 36.11 % for Parakeet v3.

Against its own base, that average is 101.8 % of the word accuracy of the 2.5 GB full-precision Parakeet TDT 0.6B v3 (11.71 % against 13.26 % word error), and Phonon-2.5 is ahead of the base outright in 14 of the 25 languages: Greek 20.8 against 36.1, Latvian 14.9 against 23.7, Slovenian 18.5 against 24.7. A standard post-training 4-bit Parakeet v3 (472 MB) goes the other way, worse than the base in all 25 languages at 14.48 % on average.

It is also a lower average word error rate across the 25 languages than Whisper large-v3 (20 times its size), Meta MMS-1B-all (25 times), NVIDIA Parakeet TDT 0.6B v3 (16 times) and Whisper large-v3-turbo (10 times). Whisper base is about the same size as Phonon-2.5 and makes 4.6 times as many errors (53.71 % against 11.71 %).

English

Phonon-2.5 English is trained for English alone. On the seven English test sets of the Open ASR Leaderboard, which cover audiobooks, meetings, earnings calls, podcasts, financial dictation and parliamentary speech, it averages 4.82 % word error, better than Phonon-2 on every one of the seven. Table 2 puts it beside the other models of its size and the full-precision Parakeet v3. The 25-language Phonon-2.5 scores 5.06 % on the same seven sets, so the multilingual model is also a better English model than Phonon-2 was, while covering 24 more languages from a smaller download (156 MB against Phonon-2’s 164 MB release file).

Why a separate English model? Some workloads want English text and nothing else: dictation, captions, compliance and search pipelines built around English. Phonon-2.5 English is our most accurate English model.

Speed

Phonon-2.5 and Phonon-2.5 English have the architecture, the file format and the engines of Phonon-2, so they run at about the same speed as Phonon-2, and it will get a lot faster in the coming weeks. For reference, these are Phonon-2’s published figures, from its release post: 174 times realtime on a base M5 MacBook Air, which is an hour of audio in about 20 seconds; 143 times realtime on eight AMD Zen 5 cores; and 6,680 times realtime batched on an H100.

Run it

The fermion command line installs from PyPI. It transcribes files, adds a start and end time for every word with --json, and serves an OpenAI-compatible endpoint. There is no language setting: Phonon-2.5 detects the language from the audio. Name the model: phonon-2.5 for the 25 languages, phonon-2.5-en for English. Phonon-2 stays available as phonon-2.

$ pip install fermion-research
$ fermion transcribe phonon-2.5 meeting.wav
$ fermion transcribe phonon-2.5-en meeting.wav
$ fermion transcribe phonon-2.5 meeting.wav --json
$ fermion serve phonon-2.5 --port 8010

On Apple silicon the models run on the GPU through MLX.

$ pip install "fermion-research[mlx]"

On Linux (x86-64 and Arm) and Windows the same package runs the CPU engine; install torch from its CPU wheel index first to skip the GPU build.

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

The weights are on Hugging Face at FermionResearch/Phonon-2.5 and FermionResearch/Phonon-2.5-English. The model pages, with the specifications and every table, are at fermionresearch.com/models/phonon-2-5 and fermionresearch.com/models/phonon-2-5-english. The command line and the server, including word timestamps with --json, are documented at fermionresearch.com/docs/speech. The Phonon-2 release post, with the format and the engines the new models share, is at fermionresearch.com/research/phonon-2.