Set up voice and push to talk

warpdrv has two voice capabilities, both powered by a Whisper speech-to-text model:

• Dictation — speak, and it’s automatically transcribed to text.
• Voice chat mode — also uses Whisper, but the moment you stop speaking it automatically triggers inference and sends your text to the model. As the reply starts to stream in, it reads the response aloud using the Kokoro voice you’ve selected in Settings.

To use either feature you need a running Whisper server, which requires two things: a Whisper model (the .bin file) and a Whisper backend (the whisper-server program). Set them up once and you’re done.

Step 1: Download a Whisper model

A Whisper model is a .bin file — different from the GGUF chat models you use for inference.

1. Open the Hub page.
2. In the search box, type ggml. The results will be Whisper models in ggml format (e.g. ggml-org/whisper-ggml).
3. Pick a model size:
   • ggml-medium — a good balance of accuracy and speed (~1.5 GB). Recommended for most setups.
   • ggml-large-v3-turbo — the best quality, but needs more VRAM (~3 GB).
4. Click Download and wait for it to finish. The model is saved into your model folder and indexed automatically.

Step 2: Get a Whisper backend

The backend is the whisper-server binary that runs the model. You can get it two ways:

Option A: Auto-install (recommended)

The easiest way is through onboarding:

1. Go to Settings and click Re-run Onboarding (or just finish onboarding if you’re still going through it).
2. On the Backends step, you’ll see a section for whisper.cpp backends below the llama.cpp ones.
3. Check the box next to a whisper backend (pick one matching your hardware — CUDA for NVIDIA, ROCm for AMD, CPU if no GPU).
4. Click Install & continue. warpdrv downloads and registers it automatically.

Option B: Add it manually

If you’d rather do it by hand:

1. Download the whisper-server binary for your OS from the whisper.cpp releases page.
2. In warpdrv, go to the Backends page and expand the Whisper.cpp Backends section.
3. Click Add.
4. Give the backend a name (e.g. whisper-cuda), click the file-picker to browse to the whisper-server binary you downloaded, and click Save.

Step 3: Launch the Whisper server

1. Go to the Servers page.
2. Click Launch Whisper.
3. In the dialog, pick the Whisper model you downloaded and the backend you registered.
4. Click Launch.

Tip: Enable the auto-launch toggle in the dialog so the Whisper server starts every time the app opens — that way dictation and voice chat are always available without manual setup.

Dictation & push-to-talk

• Click the mic button in the chat composer to dictate. Your speech is transcribed and inserted into the input.
• Use the push-to-talk (PTT) button for hands-free control: hold to record, release to transcribe.

The PTT key is configurable in Settings (default is Insert). We recommend setting it to a modifier combination like Ctrl+Shift or Alt+Shift so it doesn’t clash with normal typing. Settings also offers a Hold mode and a Global PTT key that works even when warpdrv isn’t the focused window.

Text-to-speech (Kokoro)

Kokoro TTS works out of the box — no setup required.

• Press the small TTS button at the end of any message to hear it read aloud.
• Or select a message (or part of the text) and use the TTS button in the small pop-up that appears.

The same selection pop-up also lets you dictate or annotate the selected text.

Annotating with voice

When you select text and start an annotation via the selection pop-up, you can keep using voice:

• Dictation mode — your spoken words are transcribed straight into the annotation/comment input box.
• Voice chat mode — as soon as you stop speaking, the annotation is saved as a new note. Inference is not triggered; it waits so you can keep annotating as many times as you like.

Next steps

• Use the model router — expose your models to external apps.