How TalkBar uses AI
TalkBar uses AI to write summaries, answer questions, polish dictation and find knowledge in your conversations. You decide where that AI runs: on this computer, on an Ollama server you run, or at a cloud provider with your own API key. Each feature can even use a different model.
Two kinds of AI
TalkBar runs two kinds of models.
Voice and helper models always run on this computer. They are never sent to a provider and need no setup beyond a one-time download:
- Speech-to-text for live transcripts, Enhance transcript and dictation
- Speaker detection and voice activity detection
- Live translation
- Document search for your imported documents
Your audio is never sent to an AI provider.
AI models (language models) do the writing and reasoning. These are the ones you route, and they can run locally or in the cloud.
Features that use an AI model
AI settings group the features into four feature groups. You can give each group its own model.
| Feature group | Feature | What the AI does |
|---|---|---|
| Chat & Ask AI | Chat replies | Answers in Ask TalkBar chats, grounded in your sessions, notes and knowledge |
| Ask AI | Writes or rewrites note text from your prompt in the notes Ask bar | |
| Summaries | Voice summaries | Writes the live and final summary of each session |
| Timeline | Writes the short live timeline bullets during a session | |
| Dictation & voice commands | Dictation polish | Cleans up dictated text before it is inserted (Dictation) |
| Command interpretation | Understands natural-language commands in chat and maps them to actions | |
| Knowledge & titles | Knowledge extraction | Finds memories, people and companies in sessions and documents during knowledge scans |
| Title generation | Writes short titles for notes and sessions when automatic titles are on |
A model has to support what a feature needs. Every feature needs a text model. Dictation polish, title generation and the timeline need nothing more, so small fast models work well there. The other features also need structured output. TalkBar only lets you pick models that can serve a feature.
How routing works
Settings → AI Providers → AI Features decides which model each feature uses.
Default model
One Default model powers every feature you have not customized. There are two:
- Standard: your everyday choice, used while Private AI Mode is off.
- Private AI: must be a local model. Used while Private AI Mode is on.
Test sends a tiny request to check that the default model answers.
Per-feature models
Click Customize under Per-feature models to give a feature group (or a single feature inside it) its own model. For example, a small fast model for dictation polish and a larger one for summaries. Standard and Private AI keep separate choices. A feature you do not customize shows Inherited and uses the default model; Reset returns a customized one to the default.
Choices for one chat or one session
- Chats: each chat can use its own model. Use Change chat model in the chat.
- Sessions: in New session, the Intelligence row lets you pick the model for that session's timeline and summary, or choose Private AI to keep the session on local models. You can also change the model on a session's summary later.
AI paths and what leaves your computer
Where your text goes depends on the provider behind the model.
| AI path | Where it runs | What leaves your computer |
|---|---|---|
| Built-in AI | This computer, with models you download under Private Local AI | Nothing. Works offline after the download. |
| Ollama on this computer | The Ollama app on this computer | Nothing |
| Ollama on another computer | An Ollama server you run on your network | The text a feature works on goes to that computer |
| Cloud provider with your own API key | The provider's servers (for example Anthropic, OpenAI or Google Gemini) | The text a feature works on goes to that provider |
| CLI tool (for example Claude Code or Codex) | The tool's own service | The text a feature works on goes through the tool to its service |
"The text a feature works on" means what that feature needs: for example transcript text for a summary, your question and the sources the chat uses for a chat reply, or the dictated words for polish. Audio is never part of it.
Settings that shape outgoing requests
In Settings → AI Providers → Advanced:
- Redact URLs strips full web addresses from content before it is sent to the AI provider, replacing them with placeholders.
- Log requests saves a local copy of every AI request and response for troubleshooting. Logs stay on your computer.
- Custom AI Instructions adds your own short note (up to 300 characters) to every AI request, such as your role or preferred tone.
- Response Language sets the language AI answers in. Auto (Application UI Language) follows the app language; chat replies in Auto follow the language of your question.
You can also keep one person's words out of AI with Ignore speaker.
Setting up AI
The quickest start is Set up AI (logo menu → Set up AI…, or Settings → AI Providers → Setup AI). It walks you through providers and default models, and you can open it again at any time. Details are in Providers and models and Local models.
When AI does not work
- Open Settings → Doctor and click Run all checks. Each finding links to the setting that fixes it.
- In Settings → AI Providers, use Test on the provider, and check Advanced → Diagnostics for recent errors.
- If a cloud provider fails, an alert may offer Fix now or Use Private AI to switch to local models.
- For local models, check that they are downloaded under Settings → Private Local AI.
See also Troubleshooting.