The short answer
If English accuracy is what matters most and a cloud API is acceptable, AssemblyAI Universal-3.5 Pro is ahead: 4.46 % average word error on the Open ASR Leaderboard's seven English sets, against 5.06 % for Phonon-2.5 and 4.82 % for Phonon-2.5 English. Pick Phonon-2.5 when the audio must stay on your hardware, when you need one of the European languages outside AssemblyAI's Universal-3.5 Pro list, or when per-hour fees dominate your cost.
Phonon is ahead on one set, AMI meeting recordings: 9.12 % (Phonon-2.5 English 8.76 %) against 10.63 %.
Side by side
AssemblyAI facts from assemblyai.com/pricing and its model docs, read 2026-11-13.
| Phonon-2.5 | AssemblyAI Universal-3.5 Pro | |
|---|---|---|
| Model type | Open weights on Hugging Face | Hosted API |
| Size | 156 MB download (Phonon-2.5 English: 158 MB) | Not published |
| Languages | 25 European languages (Phonon-2.5 English: English only) | 18 (Universal-2: 99; Universal-3.6 Pro Realtime: 32) |
| English word error, seven Open ASR sets | 5.06 % (Phonon-2.5 English: 4.82 %) | 4.46 % |
| Price per audio hour | No per-hour fee when self-hosted | $0.21 pre-recorded; Universal-3.6 Pro Realtime $0.45 |
| Where it runs | Your hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, Android | AssemblyAI cloud |
| Streaming | Yes, live WebSocket from fermion serve | Through the Universal-Streaming and Universal-3.6 Pro Realtime models |
| Diarization | No | Yes, +$0.02 per hour (standard, pre-recorded) |
| Word timestamps | Yes (--json output) | Yes, with confidence and speaker per word |
| Custom vocabulary | Not on Phonon-2.5 (hotwords are on Phonon-2, English) | Keyterms prompting, up to 1,000 words |
English, set by set
AssemblyAI Universal-3.5 Pro has the lower word error on six of seven sets, with the largest margins on SPGISpeech and LibriSpeech other. Phonon-2.5 and Phonon-2.5 English are lower on AMI.
| Test set | Phonon-2.5 | Phonon-2.5 English | AssemblyAI Universal-3.5 Pro |
|---|---|---|---|
| LibriSpeech clean | 1.60 | 1.64 | 1.10 |
| LibriSpeech other | 3.41 | 3.56 | 2.11 |
| AMI | 9.12 | 8.76 | 10.63 |
| Earnings-22 | 6.72 | 6.22 | 6.05 |
| GigaSpeech | 8.04 | 7.99 | 7.62 |
| SPGISpeech | 3.31 | 3.31 | 1.81 |
| VoxPopuli | 3.23 | 2.28 | 1.92 |
| Average of the seven | 5.06 | 4.82 | 4.46 |
Where AssemblyAI is the better choice
- Lower English word error on six of the seven Open ASR sets.
- Speaker labels for $0.02 per hour more; Phonon does not diarize.
- Keyterms prompting for names and product terms on the current model.
- Languages beyond Europe through Universal-2 (99 languages).
- A managed service: no servers, no model updates to deploy.
Where Phonon-2.5 is the better choice
- Privacy by placement: the model runs where the audio is, including offline laptops and Android phones.
- Cost at volume: no per-hour fee. 10,000 hours on Universal-3.5 Pro at list price is $2,100 before add-ons; a self-hosted model costs what the machine costs.
- 25 languages in one 156 MB model, including Baltic, Slavic and Nordic languages and Maltese and Greek.
- Meetings: lower word error on AMI.
pip install fermion-research
fermion transcribe phonon-2.5 interview.wav --json
fermion serve phonon-2.5 # POST /v1/audio/transcriptions, GET /v1/audio/streamQuestions
- Is AssemblyAI more accurate than Phonon-2.5?
- In English on average, yes: Universal-3.5 Pro averages 4.46 % word error on the Open ASR Leaderboard's seven sets, against 5.06 % for Phonon-2.5 and 4.82 % for Phonon-2.5 English. Phonon is lower on AMI meeting audio.
- How much does AssemblyAI cost per hour?
- AssemblyAI lists Universal-3.5 Pro at $0.21 per hour pre-recorded, Universal-2 at $0.15, and Universal-3.6 Pro Realtime at $0.45, with standard diarization at $0.02 per hour extra (read 2026-11-13).
- Which languages does AssemblyAI Universal-3.5 Pro support?
- AssemblyAI's model page says 18 languages for Universal-3.5 Pro, 32 for Universal-3.6 Pro Realtime and 99 for Universal-2. Phonon-2.5 covers 25 European languages.
- Can I run AssemblyAI models on my own servers?
- AssemblyAI's pricing and model pages describe a hosted API. Phonon-2.5 is open weights and runs on your own CPU, Mac, NVIDIA GPU or Android device.
- Why choose Phonon-2.5 if it has higher English word error?
- Because it is 156 MB, runs with no per-hour fee and no network connection, and covers 25 languages in one model. For workloads where those matter more than about 0.6 points of English word error, it is the lower-cost option.
Related
Sources and method
- AssemblyAI prices and add-ons: assemblyai.com/pricing, read 2026-11-13.
- AssemblyAI languages, keyterms and streaming models: assemblyai.com/docs/getting-started/models, read 2026-11-13; word timestamps: assemblyai.com/docs/pre-recorded-audio/word-level-timestamps, read 2026-11-13.
- Rival English word error: Open ASR Leaderboard, seven English sets, snapshot 2026-09-25 (huggingface.co/spaces/hf-audio/open_asr_leaderboard).
- Phonon-2.5 and Phonon-2.5 English: Fermion Research measurement, Phonon-2.5 release post (2026-10-22), Open ASR Leaderboard scoring code, full test sets of the same seven sets.
Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-assemblyai/