Praat's measurements, with word, phone and speaker tiers filled by speech recognition and exported as a TextGrid.
The demo runs Praat in this page and replays what real models returned for two readings of one sentence, in German and Georgian.
The phone tier is read off the acoustics by an IPA recogniser, with no pronunciation dictionary in the way. The word tier comes from a separate model.
So ანტარქტიკის comes out as ɑ n tʼ ɑ r kʼ tʼ i kʼ i s, ejectives and all, and the method works for languages that have no dictionary.
Pitch, formants, intensity and the spectrogram are Praat's own, compiled to WebAssembly. Every argument is a field you can edit, and the page prints the commands it ran.
Each tier has its own model, and results appear while the job is still going. A lane that is still being computed looks different from one that is empty.
A speaker tier keys on the voice, not the words, so it needs no language code and runs alongside the recogniser.
Here the speaker changes from S1 to S2 at 6.04 s, and the silence between them is left unassigned.
The desktop app has recognition built in. Download a model once and the recording never leaves the computer.
This Dargwa sentence was transcribed in under six seconds on a laptop CPU: words by MMS, phones by ZIPA.
Each model shows its tier, size and licence before anything downloads, and the files are checked against published checksums.
Every version supports hand annotation, TextGrid import from Praat or ELAN, and recording from a microphone. Annotations save as you work.
Praat runs in the page, and the demo replays recorded results from real models. To analyse your own audio, open a file. It stays on your computer.
Open the demoBuilt-in, offline recognition for words, IPA phones and speakers. API keys for online providers go in the system keychain.
Connect the browser version to a GPU server you run, or to a provider such as OpenAI or Cloudflare. Keys stay on the server.
python3 web/serve.py
Read the setup
The installers are not signed with Apple or Microsoft certificates yet, so both systems warn you before the first launch. The install notes explain how to continue.
At least one model per tier runs on an ordinary laptop. Which word model to use depends more on the language than on model size. MMS loads an adapter for each language, so it gets the script right where Omnilingual has to guess.
| Model | Tier | Download | Licence |
|---|---|---|---|
| Whisper base, small, large-v3 turbo | words | 142–547 MB | MIT |
| Omnilingual ASR 300M, 1B | words | 365 MB, 1 GB | Apache 2.0 |
| MMS-1b-all, 1,150+ languages | words | 1 GB, then ~9 MB a language | CC BY-NC 4.0 |
| ZIPA small, large | IPA phones | 71 MB, 310 MB | CC BY 4.0 |
| Speakers (pyannote segmentation + embedding) | speakers | 36 MB | MIT, CC BY 4.0 |
On 40 held-out Common Voice recordings each, MMS made 10.1% character errors on Dargwa and 12.1% on Ossetian. Omnilingual made 38–85% errors on the same recordings, mostly because it wrote them in the wrong script.
Export the tiers to Praat, subtitle tools or a spreadsheet, or open an existing annotation next to the recording and correct it in place.
Machine labels are a first pass. Check them against the spectrogram.