Audio and model settings
This page is the practical reference for two parts of Settings: the Audio tab, where you tune how InkSpoke captures your voice, and the AI Models section, where you choose the models that turn that voice into polished text. For the concepts behind these choices — the two model roles and the three sources — read Models and providers first; this page tells you where each control lives and what its default is.
Both areas open from the left navigation of the Settings window:
- Audio → Configuration → the Audio tab.
- AI Models → its own top-level item, split into three tabs: Global Defaults, On-Device, and Providers.
Audio settings
The Audio tab controls the whole capture side of dictation — which mic you use, the sounds you hear, when a recording stops on its own, and how aggressively InkSpoke cleans up the incoming audio.
┌──────────────────────────────────────────────────────────┐
│ Configuration › Audio │
├──────────────────────────────────────────────────────────┤
│ Microphone [ Built-in Microphone ▾ ] ⟳ Refresh │
│ │
│ Sound Effects [ ✓ On ] │
│ ▶ start ▶ stop ▶ cancel (test each) │
│ │
│ Dictation Mode ( ● Standard ◦ Live Preview ) │
│ Silence timeout [ − 30 s + ] │
│ Max recording duration [ − 300 s + ] │
│ │
│ Noise Suppression (DeepFilterNet) [ ✓ On ] │
│ Whisper Mode (quiet speech) [ Off ] │
└──────────────────────────────────────────────────────────┘
Microphone
Pick which input device InkSpoke records from. If you plug in a headset or USB mic after opening Settings, click Refresh to re-scan for it. Your active device is also shown on the listening overlay and can be switched quickly from the system-tray Microphone submenu.
Sound Effects
InkSpoke plays short chimes when a recording starts, stops, and is cancelled, so you get audible confirmation without looking at the overlay. Three test buttons let you preview each cue. Turn the whole set off with a single toggle.
The Sound Effects card only appears on platforms that provide capture sounds. If you don't see it, your build doesn't ship them.
Dictation Mode
This chooses how your speech is transcribed:
- Standard (default) transcribes everything in one pass after you stop — the most accurate option.
- Live Preview streams a running transcript into the overlay while you're still talking, finalizing at sentence boundaries.
Live Preview runs on the on-device speech path only. If you have a cloud speech model active, InkSpoke shows a warning and quietly falls back to Standard rather than firing a stream of API calls.
Select a cloud (Platform or BYOK) speech model and Live Preview can't stream — you'll get Standard behavior instead. Keep Whisper (on-device) active to see the live transcript. More detail in Dictation modes and languages.
Recording limits
Two safety timers stop a session that runs away — for example if you walk off mid-dictation:
| Setting | Default | Range | What it does |
|---|---|---|---|
| Silence timeout | 30 s | 0–300 s | Auto-cancels after this many seconds of continuous silence. 0 disables the timer. |
| Max recording duration | 300 s (5 min) | 0–3600 s | Hard cap on a single dictation. 0 means no limit. |
Noise Suppression (DeepFilterNet)
A neural denoiser that strips background noise — fans, keyboard clatter, ambient chatter — while preserving your voice. It's on by default and adds only a few milliseconds of latency. Because it relies on a downloadable model, the toggle is disabled while that model is still downloading.
Whisper Mode
Whisper Mode (off by default) tunes InkSpoke for dictating quietly — an open office, a shared room, a sleeping partner nearby. When it's on, InkSpoke lowers the threshold for detecting speech, boosts input gain, and asks Whisper to search harder, so soft, near-silent speech still transcribes cleanly. Leave it off for normal-volume dictation. It's also toggleable from the system-tray menu.
Audio settings at a glance
| Setting | Default | What it does |
|---|---|---|
| Microphone | System default until you pick one | Chooses the capture device; Refresh re-scans. |
| Sound Effects | On | Start / stop / cancel chimes, with test buttons. |
| Dictation Mode | Standard | Standard (batch) vs Live Preview (streaming, on-device only). |
| Silence timeout | 30 s | Auto-stop after continuous silence (0 = off). |
| Max recording duration | 300 s | Hard cap per dictation (0 = no limit). |
| Noise Suppression (DeepFilterNet) | On | Neural background-noise removal (needs its model downloaded). |
| Whisper Mode | Off | Quiet-speech tuning (lower threshold, higher gain). |
AI Models
The AI Models area is where you choose the two models every dictation flows through — a speech model that hears you and a text model that refines the result — and where you download local models or connect your own providers. It has three tabs.
┌────────────────────────────────────────────────────────────────┐
│ AI Models │
│ [ Global Defaults ] [ On-Device · PRO ] [ Providers · PRO ] │
└────────────────────────────────────────────────────────────────┘
Global Defaults
Your baseline models — used for every dictation unless a workspace pins its own. Two grouped pickers list every available model under Platform, On-Device, and BYOK headings.
┌────────────────────────────────────────────── ────────────┐
│ AI Models › Global Defaults │
├──────────────────────────────────────────────────────────┤
│ Speech recognition │
│ [ Whisper Small (On-Device) ▾ ] │
│ │
│ Text processing (refinement) │
│ [ Platform AI (Platform) ▾ ] │
│ Workspace-default refinement [ ✓ On ] │
│ Token limit [ − ···· + ] │
│ │
│ (Master AI Refinement is ON) │
└────────────────────── ────────────────────────────────────┘
New installs start on the private-by-default pairing: Whisper Small (on-device, offline, free) for speech and Platform AI (cloud) for refinement during your Pro trial.
The text side adds two controls:
| Control | Default | What it does |
|---|---|---|
| Workspace-default refinement | On | Whether workspaces that don't pin their own text model still get refined by this global default. |
| Token limit | — | Caps how long a refined response can be. |
Both sit under the master AI Refinement switch (on Configuration → General, on by default). When that master switch is off, these controls are greyed out with a reminder, and InkSpoke injects your raw transcript verbatim. See How refinement works.
On-Device (Pro)
PRO (requires Pro or Perpetual) — download and manage models that run entirely on your machine, so your audio and text never leave the computer. A banner reminds you that Whisper Small (244M) is free; every other on-device model requires Pro or Perpetual.
┌───────────────────────────────────────────────────────────┐
│ AI Models › On-Device [PRO] │
│ Small (244M) is free · other models need Pro │
│ Storage ▓▓▓▓▓▓▓░░░░░░░░ used by downloaded models │
├───────────────────────────────────────────────────────────┤
│ Whisper Small 244M ● in use │
│ SPEED ●●●●○ ACCURACY ●●●○○ many languages │
│ [ Delete ] │
│ Whisper Large V3 Turbo │
│ SPEED ●●●○○ ACCURACY ●●●●● [ Download ] │
│ ────────────────────────────────────────────────── │
│ Use GPU for dictation [ ✓ On ] │
│ Apple Neural Engine (macOS only) [ Download ] │
│ ────────────────────────────────────────────────── │
│ Local text model (GGUF) │
│ Max context size [ − 16384 + ] │
└───────────────────────────────────────────────────────────┘
Speech — Whisper models. InkSpoke ships the Whisper family in several sizes. Each card shows its size chip, an in use pill for the active one, SPEED and ACCURACY dot ratings, its language support, and a Download / Delete button. A storage bar tracks total disk use; the model that's currently in use can't be deleted.
| Whisper size | Tier |
|---|---|
| Tiny | Pro |
| Base | Pro |
| Small (244M) | Free — the default |
| Medium | Pro |
| Large | Pro |
| Large V3 Turbo | Pro |
| Large V3 Turbo Q5 / Q8 | Pro |
As a rule, smaller models are faster and lighter; larger ones are more accurate. See Models and providers for the full trade-off.
Acceleration. A Use GPU for dictation toggle (on by default) appears only where hardware acceleration is available. On macOS you'll also see an optional Apple Neural Engine accelerator — a small extra download that's used automatically once installed. See Platform notes below.
Text — local model. The On-Device tab also offers a downloadable local text model (GGUF format) so refinement can run fully offline. It has one tunable cloud models don't:
| Setting | Default | Range | What it does |
|---|---|---|---|
| Max context size | 16,384 | 512 – 131,072 | How much text (in tokens) the local model can consider at once. Larger uses more memory. |
On-device models can be RAM-hungry. By default InkSpoke unloads idle models after 10 minutes and reloads them next time you dictate — the Model Memory strategy on General and hotkeys.
Providers (BYOK, Pro)
PRO (requires Pro or Perpetual) — Bring Your Own Key. Connect any OpenAI-compatible endpoint (a commercial API, or a local server like Ollama or LM Studio) with your own key. Each provider you add can expose several models that then appear in the Global Defaults and workspace pickers alongside Platform and On-Device options.
┌───────────────────────────────────────────────────────────┐
│ AI Models › Providers [PRO] [ + Add ] │
├───────────────────────────────────────────────────────────┤
│ My OpenAI api.openai.com/v1 3 models │
│ Local LM Studio localhost:1234/v1 1 model │
│ │
│ ── Editing a provider ────────────────────────────── │
│ Quick Setup [ preset ▾ ] [ Apply ] │
│ Name [ My OpenAI ] │
│ API Base URL [ https://api.openai.com/v1 ] │
│ API Key [ •••••••••••• ] (stored in keychain) │
│ Timeout [ 30 ] s │
│ Models id · type · token limit · temperature │
│ [ + add row ] │
│ [ Test Connection ] [ Cancel ] [ Save ] │
└───────────────────────────────────────────────────────────┘
Add Provider is disabled unless you're on Pro — but any providers you already added stay viewable and deletable even if your plan lapses; you just can't add new ones until you upgrade again. Each provider card shows its name, base URL, and model count.
Editing a provider gives you these fields:
| Field | Notes |
|---|---|
| Quick Setup | Optional preset picker + Apply to pre-fill a common provider. |
| Name | Your label for the provider. |
| API Base URL | The OpenAI-compatible endpoint. Built-in presets already include the /v1 suffix. |
| API Key | Masked in the UI and stored in your OS keychain — never in InkSpoke's settings file. |
| Timeout | 5–120 s. Defaults to 10 s for text models, 30 s for speech. |
| Temperature | Sampling temperature for the provider. |
| Models table | One row per model: id, type (speech or text), token limit (50k–1M, in 5k steps), and temperature. Add or remove rows as needed. |
| Test Connection | Verifies your key and URL work before you save. |
Providers and their models are stored in InkSpoke's local database, with keys in the keychain — none of it lands in settings.json. You can manage the same personal keys from the web account.
API keys you paste here are held in your operating system's secure credential store on this device. Removing a provider deletes its stored key.
Platform notes
On-device speech models use whatever acceleration your machine offers. The Use GPU for dictation toggle and the Apple Neural Engine card appear only where they apply:
| Platform | On-device acceleration |
|---|---|
| Windows | CUDA GPU when present, otherwise CPU. |
| macOS | Metal GPU, plus an optional Apple Neural Engine download used automatically once installed. |
| Linux | CPU only (no GPU toggle shown). |
The Sound Effects card is shown only on platforms that ship capture chimes.
Next steps
- Models and providers — the concepts: two roles, three sources, and how workspaces override them.
- General and hotkeys — the AI Refinement master switch, Model Memory, and injection settings.
- On-device vs. cloud and privacy — decide where your words get processed.
- Account, sync, and updates — unlock Pro to use the models above.