Voice control

Drive OpenTraderWorld with your voice, or dictate into any text field. Push to talk only: the microphone opens when you press and closes when you release. Nothing listens in between.

It is off by default. Turn it on in Settings → Voice → Speech & shortcut.

The microphone needs HTTPS or localhost

Browsers only give a web page the microphone (and their built-in speech recognition) on a secure origin: a page served over HTTPS, or opened on localhost. This is a browser rule, not an OpenTraderWorld one, and no setting in the app can lift it.

How you open OTW	Voice works?
On the machine that runs it, http://127.0.0.1:5454 or http://localhost:5454	Yes
LAN + HTTPS or Public mode	Yes
Local network (LAN) mode over plain HTTP, from another device	No, the browser hides the microphone

To use voice from a phone or another computer on your network, switch Settings → Network to LAN + HTTPS. On plain HTTP the voice settings page shows a warning and the microphone button explains why it cannot start.

Two shortcuts, two modes

	Default	What happens to what you say
Command	Alt+V (⌥V on a Mac)	Becomes a plan of actions, even when a text field has focus.
Dictation	Alt+Shift+V (⌥⇧V)	Is typed, word for word, into the text field that has focus. Never read as a command.

Hold the shortcut while you speak, release to finish, Esc to cancel. The microphone in the top bar always runs commands: click to start, click again to stop, or hold it like a walkie-talkie.

Both shortcuts can be changed in Settings → Voice → Speech & shortcut. Each needs Ctrl, Alt or ⌘ (or a function key F1 to F12), so it never gets in the way of normal typing, and the two must differ.

Speech engine

The engine turns your recording into text. Pick one in Settings → Voice → Speech & shortcut.

Engine	Setup	Where the audio goes
This browser	none	The browser vendor's speech service (Google for Chrome, Microsoft for Edge, Apple for Safari). Not available in Firefox.
Self-hosted Whisper	the bundled service, see Example: self-hosted Whisper	Stays on your machine
whisper.cpp server	your own server, its /inference endpoint	Stays on your server
OpenAI / Groq	an API key (pasted or plugged from the Vault)	Sent to that provider
Other	any server exposing the OpenAI /audio/transcriptions API	That server

With a server engine the browser records, converts the audio to a small WAV file and OTW relays it to the engine. Test sends half a second of silence to check the URL, key and model. Keys are encrypted at rest and never sent back to the browser.

Example: self-hosted Whisper, from scratch

The bundled Whisper service is optional and not started by a plain up. Steps 1 to 3 are done once: the model is kept in the whisper-cache volume across restarts.

1. Start the service, from the repository root:

docker compose -f deploy/docker-compose.yml --env-file deploy/.env --env-file deploy/network.env --profile voice up -d whisper
docker logs -f opentraderworld-whisper-1

Wait for Uvicorn running on http://0.0.0.0:8000, then Ctrl+C. The port is only reachable by OTW inside Docker, not from your browser, and that is intended.

2. Download the model (about 500 MB, a few minutes):

docker exec opentraderworld-whisper-1 python -c "import urllib.request;print(urllib.request.urlopen(urllib.request.Request('http://localhost:8000/v1/models/Systran/faster-whisper-small',method='POST'),timeout=1800).read())"

3. Check it is installed: the answer must list Systran/faster-whisper-small.

docker exec opentraderworld-whisper-1 python -c "import urllib.request;print(urllib.request.urlopen('http://localhost:8000/v1/models').read())"

4. Plug it into OTW. Open the app on http://localhost:5454 (or over HTTPS, see above), then Settings → Voice → Speech & shortcut:

1. Add engine, preset Self-hosted Whisper: it fills in http://whisper:8000/v1 and Systran/faster-whisper-small.
2. Save and test: the engine shows Works · … ms.
3. Select the engine, then turn the switch at the top On. Optionally set the Language.

5. Try it: hold ⌥V / Alt+V, say "open journal", release, press Enter. For dictation, click into a text field and hold ⌥⇧V / Alt+Shift+V.

The first transcription after a restart is slower while the model loads into memory.

Voice commands

Settings → Voice → Voice commands: a command is a phrase (plus other ways of saying it) and an ordered list of steps:

• Open a page: a module, the dashboard or Settings.
• Ask the agent: a prompt sent to the floating assistant, which acts with its own tools.
• Run a workflow: start an Automator workflow.
• Theme, Hide figures: the same as the buttons in the top bar.
• Speak: read a sentence aloud.

A phrase fires only when it is said on its own, never as a word inside a sentence: dictating "a blue turtle" does not run your turtle command.

Built in, no setup

• "open <page>" ("ouvre <page>", "öffne <Seite>", "abre <página>", "apri <pagina>", "打开 <页面>") opens any installed module, the dashboard or Settings. A name that matches several pages is refused with the list, never guessed.
• With Hand the rest to the agent on, anything no command matches goes to the assistant as one request.

Chaining

Say several things in one go, joined by and or then (et, puis in French, and so on, following the Language setting). A comma in the transcript splits too:

"turtle, then open settings and compare AAPL and MSFT"

runs your turtle command, opens Settings, then asks the agent to "compare AAPL and MSFT". Leftover pieces next to each other are kept together, so the agent gets the whole request.

Confirmation

Every plan is shown before it runs: what was heard, each step and where it came from. Enter runs it, Esc cancels. Steps run in order and the plan stops at the first failure, naming the step.

A command can be marked Run without confirmation. It is off by default, and it only applies when that phrase is said alone: chained with anything else, the plan is still shown first. Agent steps keep the assistant's own confirmation for every write.

Good to know

• The assistant is hidden on the Agent page, so a plan with an agent step cannot run from there.
• Read the outcome aloud uses the browser's own voices: no setup, no audio leaves the browser.
• The public demo shows the voice pages read-only: it does not save settings or send audio.
• Your own reverse proxy in front of OTW must not block the microphone: its Permissions-Policy header needs microphone=(self). With microphone=() the browser refuses at once, without asking.