AI & Voice Settings

This page covers the AI and voice settings available in Settings. Every AI and voice feature runs entirely on your device — there’s nothing to subscribe to.

AI Settings

Model

NoteTalker runs a local language model for chat and voice. Pick which one from the model selector in Settings → AI. The default is Mistral Nemo 12B (~7.5 GB on Windows and Linux; the macOS build uses a 4-bit MLX version, ~6.9 GB), which needs roughly 10–12 GB of VRAM.

If your GPU has less memory, choose a smaller model:

Model	Download	Notes
Mistral Nemo 12B	~7.5 GB	Default on Windows/Linux; best quality, ~10–12 GB VRAM
Qwen3 4B	~2.5 GB	Runs comfortably on 8 GB VRAM (or a 16 GB Apple Silicon Mac)
Qwen3 8B	~5 GB	A middle option
Qwen3 14B	~9 GB	Higher quality, needs more VRAM
Ministral 3 14B	~8.2 GB	256K-token context for very long conversations

Switching models downloads the new one on first use; the previous file stays on disk, so you can switch back without re-downloading.

The model is held in VRAM only while it’s in use. After about ten minutes of inactivity NoteTalker unloads it automatically to free memory, then reloads it the next time you send a message — so the first reply after a long idle takes a little longer while the model loads back in.

Context Window Size

The local model runs with an 8,192-token context window by default. This is how much conversation history, memory, and tool information NoteTalker can hold in mind at once.

You can raise it — up to around 32K tokens on the default Mistral Nemo 12B model — for longer memory within a conversation. The tradeoff is VRAM: a larger window uses more of it. As a rough guide, a 32K window needs about 12 GB of VRAM with the 12B model, and 64K needs roughly 17 GB. If you have less VRAM to spare, keep the window smaller or switch to a lighter model.

Leave the override set to 0 to let NoteTalker pick a context size automatically for whichever model you’re running. Most users should leave it at the default.

Web Search

Default: Enabled

When enabled, NoteTalker can search the web and fetch web pages to answer questions it can’t find in your vault or its own knowledge. NoteTalker will always ask before searching — it never searches the web automatically.

Turn this off if you prefer NoteTalker to only use your vault and its built-in knowledge, or if you’re working offline and don’t want web search attempts.

Clear Chat History

Click Clear Chat History to reset your current conversation. This starts a fresh context — useful if the conversation has gotten long or you want to change topics completely.

This only clears the active conversation in the UI. Your session transcripts and memory files in the vault are not affected.

For more detail on how context is assembled and managed, see How Context Works.

---

Voice Settings

FFmpeg Status

Before anything else, check that FFmpeg is installed — it appears as a status indicator at the top of the Voice settings section. FFmpeg is required for recording, chunking, and processing audio.

If FFmpeg is not detected, voice features won’t work. See Voice Not Working for installation instructions.

Wake Word Enabled

Toggles “Hey Jarvis” detection on or off. When enabled, you can say “Hey Jarvis” to activate voice input hands-free.

Note

Wake word detection runs entirely on your device using a local ONNX model. Nothing is recorded or transcribed until you actually trigger the wake word and speak a message — and even then, your audio stays on your machine.

Wake Word Sensitivity

Range: 0.3 – 0.8 | Default: 0.5

Controls how easily NoteTalker triggers on “Hey Jarvis”.

• Lower values (0.3 – 0.4): Fewer false activations, but you may need to speak more clearly or loudly
• Higher values (0.6 – 0.8): Triggers more easily, but may activate on similar-sounding words or background speech

If NoteTalker activates unexpectedly, lower the sensitivity. If it consistently misses your wake word, raise it.

Voice Auto-Stop

These settings control how NoteTalker detects when you’ve finished speaking during push-to-talk, conversation mode, and wake word interactions. NoteTalker uses silence detection — when you stop talking for long enough, it automatically ends the recording.

Silence Threshold

Range: 0.01 – 0.15 | Default: 0.05

Sets the audio level below which NoteTalker considers the input to be “silence.” Lower values mean only very quiet audio counts as silence; higher values are more aggressive about cutting off.

• Lower values (0.01 – 0.03): Better for quiet environments — NoteTalker waits for near-total silence before stopping. Use this if NoteTalker cuts you off while you’re pausing to think.
• Higher values (0.08 – 0.15): Better for noisy environments — NoteTalker stops more readily. Use this if NoteTalker keeps recording background noise after you’ve finished speaking.

Tip

If NoteTalker frequently cuts you off mid-sentence, try lowering the silence threshold. If it keeps recording long after you’ve stopped talking, raise it.

Silence Duration

Range: 500 – 5000ms | Default: 1500ms

How long NoteTalker waits in silence before it stops recording. This is the pause length that signals “I’m done talking.”

• 500ms: Very responsive — stops almost immediately when you pause. Good for short commands.
• 1500ms (default): Natural conversational pace — allows brief pauses between sentences without cutting off.
• 3000 – 5000ms: Very patient — lets you take long pauses to think. Good for dictation or brainstorming where you need time between thoughts.

Speech Speed

Range: 0.5 – 2.0 | Default: 1.0

Controls how fast NoteTalker speaks when reading responses aloud.

• 0.5: Half speed — noticeably slow, useful for accessibility or careful listening
• 1.0: Normal speed
• 1.5: Comfortably faster for experienced listeners
• 2.0: Very fast — takes getting used to, but saves time

---

For a walkthrough of using voice to send messages and record voice notes, see Push-to-Talk and Conversation Mode.