Transcription and AI models
Parley uses two kinds of model:
- a transcription model (Whisper), which turns speech into text. It always runs on your computer.
- a language model, which writes summaries, chapters, action items, follow-up emails and briefs, and answers your questions in Ask. You choose where this runs.
Transcription (Whisper)
Settings → Transcription lists every Whisper model under Model catalog, with its accuracy, size and status. Click Download to fetch one (from Hugging Face), Cancel to stop a download and Delete to remove it. Pick the model used for live transcription under Active model → Whisper model; only downloaded models are listed.
| Model | Accuracy | Size |
|---|---|---|
| tiny-q5_1 / tiny | Decent | 31 MB / 74 MB |
| base-q5_1 / base | Good | 57 MB / 142 MB |
| small-q5_1 / small | Good | 181 MB / 466 MB |
| medium-q5_0 / medium | High | 514 MB / 1.5 GB |
| large-v3-turbo-q5_0 (default) | High | 547 MB |
| large-v3-turbo | High | 1.5 GB |
| large-v3-q5_0 | High | 1.0 GB |
| large-v3 | High | 2.9 GB |
The q5 variants are compressed versions that are much smaller and nearly as accurate. The default, large-v3-turbo-q5_0, is the best balance for most people. Smaller models are faster on slow CPUs; the larger large-v3 models are the most accurate.
Tip
Download large-v3-q5_0 as well as your live model. After each recording, Parley then automatically re-transcribes the meeting with it in the background for a more accurate final transcript. It never downloads a model for this on its own.
Every model understands many languages. To fix the language instead of detecting it, set Settings → Summary → Language & summaries → Transcription language to a language code such as en, fr or de. Leave it blank to detect the language automatically.
Transcription uses your GPU when you installed the Vulkan build (or built Parley with CUDA), and the CPU otherwise. There is no setting for this.
Custom vocabulary
Whisper may misspell names, products and acronyms it hasn’t heard of. Under Settings → Transcription → Custom vocabulary, add each Term the way you want it spelled, optionally with what it Sounds like (for example Kubernetes, sounds like cooper netties).
- The Whisper prompt preview shows exactly what Whisper is given. There is room for roughly a dozen or two terms; Parley warns you about terms that don’t fit.
- Include speaker and attendee names also adds the names of your speakers and of the people invited to the meeting’s calendar event.
- Correct near-miss spellings fixes close misspellings of your terms after transcription.
Custom vocabulary is applied when a meeting is refined after recording, re-transcribed or imported. It isn’t applied to the live transcript.
Language model for summaries and Ask
Choose a Provider under Settings → Summary:
| Provider | Runs | What you need |
|---|---|---|
| Built-in AI (local) (default) | On your computer | Nothing: pick a model under Built-in AI models |
| Ollama | On your computer, or a server you run | Ollama installed, and its Endpoint if it isn’t http://localhost:11434 |
| OpenAI, Claude, Groq, OpenRouter | In the provider’s cloud | An API key from that provider |
| Custom OpenAI-compatible | Wherever the server is | Its Endpoint URL, Model name and, if required, an API key |
Built-in AI
Runs entirely on your computer with llama.cpp. There are two models:
- Gemma 3 1B (Fast): about 1 GB to download and 1 GB of memory. The default.
- Gemma 3 4B (Balanced): about 2.3 GB to download and 3.5 GB of memory. Noticeably better summaries if your computer can spare the memory.
Download a model and click Select to use it.
Ollama
Install Ollama, then choose Ollama as the provider and pick one of your installed models. You can also type a model name (such as gemma3:4b) and click Pull model to download it from inside Parley.
With Ollama as the provider, Parley also uses it for semantic search instead of its built-in model. Pull the embedding model once:
ollama pull nomic-embed-text
Cloud providers
Paste your key under API keys, then choose a model from the list Parley fetches from the provider. Keys are saved in Parley’s local database in plain text, so anyone with access to your user account could read them.
Warning
With a cloud provider, the meeting transcript is sent to that provider whenever you generate a summary, ask a question or use any other AI feature. Their privacy policy then applies. The built-in model and a local Ollama keep everything on your computer.
Other summary settings
- Templates → Default summary template: the template pre-selected when you generate a summary. See Meeting notes.
- Auto-summarize after recording: write the summary automatically when a recording stops. Off by default.
- Auto-generate chapters after recording: split each new meeting into chapters automatically. Until you change it, it follows Auto-summarize after recording.
- Show transcription confidence: show a coloured dot by each transcript line for how sure Whisper was of it. On by default.