---
title: "Phonon-2.5 vs Deepgram Nova-3"
description: "Deepgram Nova-3 is a hosted speech API from $0.0043 a minute with diarization; Phonon-2.5 is a 156 MB open model you run yourself. Prices, features, trade-offs."
canonical: "https://www.fermionresearch.com/models/phonon-2-5/vs-deepgram/"
source: "Fermion Research"
---

Phonon/facts checked 13 November 2026

# Phonon-2.5 vs Deepgram Nova-3

A per-minute cloud API against an open model on your own machines. This page compares what each offers; it makes no accuracy claim, because we have no like-for-like measurement of Nova-3.

[Phonon-2.5 downloads and benchmarks](https://www.fermionresearch.com/models/phonon-2-5/)

## The short answer

Choose Deepgram Nova-3 if you want a managed API with speaker diarization, keyterm prompting and streaming built in and you are happy to pay per minute. Choose Phonon-2.5 if the audio has to stay on your own machines, your languages are among its 25, or volume makes a per-minute fee your largest cost.

Nova-3 has no row in the Open ASR Leaderboard snapshot we compare against, and we have not scored it with the same protocol, so this page gives no accuracy comparison. Test both on your own recordings.

## Side by side

Deepgram facts from deepgram.com/pricing and developers.deepgram.com, read 2026-11-13. Prices are pay-as-you-go list prices.

|  | Phonon-2.5 | Deepgram Nova-3 |
| --- | --- | --- |
| 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) | Deepgram's language table lists 24 of Phonon-2.5's 25 (not Maltese), plus many others |
| English word error, seven Open ASR sets | 5.06 % (Phonon-2.5 English: 4.82 %) | No row in the snapshot; not measured by us |
| Price per audio hour | No per-hour fee when self-hosted | Pre-recorded $0.0043/min ($0.258) monolingual, $0.0052/min ($0.312) multilingual; streaming $0.0077/min and $0.0092/min |
| Where it runs | Your hardware: CPUs (Linux x86-64 and arm64, Windows x86-64), Apple silicon, NVIDIA GPUs, Android | Deepgram cloud; self-hosted with an Enterprise plan, or through Amazon SageMaker |
| Streaming | Yes, live WebSocket from fermion serve | Yes (Nova-3, and Flux for voice agents) |
| Diarization | No | Yes; included for pre-recorded, $0.0020/min when streaming |
| Word timestamps | Yes (--json output) | Yes (start and end per word) |

## What volume costs

At Deepgram's list price, 1,000 hours of pre-recorded English audio on Nova-3 cost $258, and 1,000 hours of multilingual audio $312. Keyterm prompting adds $0.0013 per minute and redaction $0.0020 per minute.

A self-hosted Phonon model has no per-minute fee; you pay for the machine. On our staging server, Phonon-2 (the earlier English model of the same size) transcribed about 960 audio hours per NVIDIA L4 GPU hour on a realistic mix, a fully loaded cost of $0.00204 per audio hour at 60 % utilisation, GPU and gateway included. Phonon-2.5 has not been costed this way yet.

## Where Deepgram is the better choice

- Speaker labels: diarization is included with pre-recorded audio. Phonon does not diarize.
- Voice agents: Flux is a streaming model with turn detection built in.
- Custom vocabulary on a multilingual model today: keyterm prompting is an add-on. Phonon-2.5 has no hotwords yet (Phonon-2, English only, has them).
- Add-ons in one call: redaction, entity detection and smart formatting.
- No servers to run, and many more languages than Phonon's 25.

## Where Phonon-2.5 is the better choice

- The audio never leaves your network: Phonon runs on a laptop CPU, a Mac, an NVIDIA GPU or an Android phone, with no account.
- No per-minute fee, so cost tracks hardware, not hours of audio.
- Open weights: you can pin a version, run it offline and check its published benchmark numbers yourself.
- Maltese, which is in Phonon-2.5's 25 and not in Deepgram's Nova-3 table.

```text
pip install fermion-research
fermion transcribe phonon-2.5 interview.wav --json
fermion serve phonon-2.5   # POST /v1/audio/transcriptions, GET /v1/audio/stream
```

## Questions

## Related

- [Speech-to-text API cost](https://www.fermionresearch.com/speech/speech-to-text-api-cost/)
- [Self-hosted vs API speech-to-text](https://www.fermionresearch.com/speech/self-hosted-vs-api-speech-to-text/)
- [Self-host a speech-to-text server](https://www.fermionresearch.com/models/phonon-2-5/self-host/)
- [Phonon-2.5 vs AssemblyAI](https://www.fermionresearch.com/models/phonon-2-5/vs-assemblyai/)

## Sources and method

- Deepgram prices and add-ons: deepgram.com/pricing, read 2026-11-13.
- Deepgram languages, Flux, diarization and word timings: developers.deepgram.com/docs/models-languages-overview and /docs/diarization, read 2026-11-13.
- Deepgram self-hosting: developers.deepgram.com/docs/self-hosted-introduction, read 2026-11-13.
- 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.
- Phonon-2 self-hosting cost: Fermion Research measurement on our staging server, 2026-10-08, realistic audio mix, NVIDIA L4.

Canonical: https://www.fermionresearch.com/models/phonon-2-5/vs-deepgram/

Source: [https://www.fermionresearch.com/models/phonon-2-5/vs-deepgram/](https://www.fermionresearch.com/models/phonon-2-5/vs-deepgram/)
