Last updated: November 2026
Kepler has 2 voice features, both set up in Settings → Voice:
- Dictation adds a microphone to the agent prompt, so you can talk instead of typing. Transcription runs on your device, using a model Kepler downloads once. Kepler never uploads your audio.
- Read aloud reads an agent’s replies to you. See Read aloud.
Dictation has 2 transcription engines:
- Qwen3-ASR is the default. It reads what you’re writing, so it gets developer terms such as Jira, rebase, or pull request right more often than Whisper. It needs a graphics card with WebGPU.
- Whisper works on any machine. Use it where WebGPU is unavailable.
Dictation turns on once you download a model in Settings → Voice. The microphone in the prompt becomes available when the model is ready.
Choose what happens when you stop talking
To record, click Dictate in the prompt, speak, then click Stop and transcribe.
To choose what Kepler does with what you say, go to Settings → Voice → When you stop talking and pick one of these options:
| Option | What it does |
|---|---|
| Insert the text so you can edit it | Puts the text in the prompt for you to review. This is the default |
| Send the prompt | Sends the prompt as soon as you stop recording |
| Keep listening, transcribing each phrase as you pause | Keeps the microphone open and transcribes each phrase as you pause |
With Qwen3-ASR, Keep listening can correct earlier words in a paragraph as you keep talking. Once you edit that text, Kepler adds only new words and never overwrites what you typed.
Kepler treats short, whole utterances as commands rather than dictation: saying “send”, “cancel”, or “clear” acts on the prompt instead of adding text to it. Kepler dictates anything longer verbatim.
While a transcription is in progress, the microphone shows Transcribing…. If transcription falls behind, Kepler pauses the microphone until it catches up. If this happens often, pick a smaller model.
Teach dictation your words
Qwen3-ASR learns how you spell words it gets wrong. When you fix a single dictated word in the prompt before you send it, Kepler notes the fix. After you make the same fix in 2 sends, a notice suggests the word, and Review opens Settings. Dictation uses a suggested word only after you approve it.
Learned words, in Settings → Voice, lists the words in 2 groups:
- Suggested words: each shows what Qwen3-ASR heard, with Approve and Dismiss. A dismissed fix is never suggested again.
- Words: the words you approved and the words you added. Each has a remove button. Type in Add a word and click Add to add one yourself. Forget learned words… removes every learned, suggested, and added word on this device.
Kepler stores only the words, on this device only, and keeps them after you sign out.
Set up dictation in Settings
In Settings → Voice, the Dictation section’s Model list offers 3 models. Every model runs on this computer, and nothing you say leaves it.
| Model | What it is |
|---|---|
| Qwen3-ASR 1.7B | The most accurate, especially with developer terms and mixed languages. Needs a recent graphics card |
| Qwen3-ASR 0.6B | Better than Whisper with technical terms, and lighter on memory, for graphics cards with less memory |
| Whisper | Works on any machine, including without a compatible graphics card. Less accurate with technical terms |
Kepler checks the graphics card and marks the Qwen3-ASR size it suits as Recommended. Without WebGPU, only Whisper can run.
To set up dictation, follow these steps:
- Open Settings → Voice.
- For Whisper, pick a Quality on its row: Fast, Balanced, Accurate, or Most accurate. Kepler marks the best fit for this machine as suggested.
- Click Download on the model’s row.
Your first download turns dictation on. Downloads keep going if you leave the page, and Kepler notifies you when the model is ready. With more than 1 model downloaded, pick the one dictation uses.
Once a model is downloaded, these settings apply:
| Setting | What it does |
|---|---|
| Remove… | On a downloaded model’s row: deletes the model and frees its storage. Removing the model that dictates turns the microphone off until you pick another |
| Dictation language | With a Qwen3-ASR model: the language you speak. See Dictation language |
| When you stop talking | Chooses what happens to what you say. See Choose what happens when you stop talking |
Dictation language
With Qwen3-ASR, Dictation language is Auto-detect by default, or 1 of 30 languages, each listed by its own name. Choose a language to always dictate in it. Whisper ignores this setting.
Model updates
When a newer version of a Qwen3-ASR model ships, a notice offers Update, and the model’s row in Settings → Voice keeps offering Update. Dictation keeps using the installed model until you update.
Linux graphics cards
On Linux, Kepler turns WebGPU on so Qwen3-ASR can use the graphics card:
- Under Wayland, WebGPU is on by default.
- Under X11, start Kepler with
--enable-linux-webgputo turn it on. - Start Kepler with
--disable-linux-webgputo turn it off.
If Kepler’s first window hangs or never draws with WebGPU on, Kepler turns WebGPU off for later launches and restarts. Without WebGPU, use Whisper.
For every dictation and read aloud setting, see Voice settings.
Why the microphone is unavailable
Hover the microphone, and its tooltip explains why:
| Tooltip | What it means |
|---|---|
| Dictation needs a model. Open Settings to download one. | No model is downloaded yet. Clicking opens Settings |
| Dictation requires a secure (https) connection | You’re using Kepler over plain HTTP, where the browser will not grant microphone access |
| Dictation is not supported in this environment | This build or browser cannot run it |
| The selected voice model needs WebGPU, which is not available here. Open Settings to choose Whisper. | A Qwen3-ASR model is selected, but this computer or browser has no WebGPU. Clicking opens Settings |
| The graphics card is not responding, so dictation cannot start. Open Settings to check it again. | The graphics card did not answer. Clicking opens Settings |
| Loading voice model… / Transcribing… | Busy. The button is disabled for the moment |
With Qwen3-ASR, the microphone’s Dictate tooltip also names the model, for example Dictate · Qwen3-ASR 1.7B.
When dictation fails, Kepler names the cause, with Open Settings where Settings can fix it. For example:
- A voice model is being installed. Try again when it finishes.
- The voice model is no longer downloaded on this computer. Download it again in Settings.
- The voice engine stopped responding. Try again; the model reloads on its own.
- Dictation stopped unexpectedly.
Microphone permission on macOS
macOS asks for microphone access the first time you record.
If access is denied or restricted, turn it on in System Settings → Privacy & Security → Microphone, if your device policy allows it. Open System Settings, in Kepler’s message, takes you straight to that pane. You may need to restart Kepler for the change to take effect.
Read aloud
Read aloud speaks an agent’s replies to you. Kepler reads only the prose: code blocks, diffs, and tables become a short summary. In the chat, the sentence being read is highlighted.
To have a reply read aloud, use one of these controls:
- Read aloud, the speaker button under each chat reply. While that reply plays, the button reads Stop reading.
- Read aloud from here, in the right-click menu of a message or a terminal session. Kepler reads from the sentence under the pointer through the replies after it.
- Read last reply, on a terminal session’s status bar.
While anything is being read, Stop reading appears in the top bar, on every page.
The read aloud controls appear only on a computer with a voice to read with. Read aloud starts with your system’s voices. To change the voice, follow these steps:
- Open Settings → Voice, and go to Read aloud.
- Under Voice engine, choose System voices or Kokoro. Kokoro has natural English voices and needs a graphics card with WebGPU. Click Download to get it. Until Kokoro is downloaded, and for text that is not English, the system voice reads.
- Pick a Voice. Choosing one plays a short sample.
- Set the Speed.
Remote windows
Dictation works in a window attached to a remote environment. The agent runs on the remote machine; the microphone and the transcription stay local, on the computer in front of you.
Over a plain HTTP connection the microphone is unavailable. See Why the microphone is unavailable.
Voice settings belong to each computer:
- A desktop window, local or connected to a remote, uses this computer’s voice settings.
- A Remote Access browser uses the settings of the Kepler it’s connected to.
- Neither the settings nor the models sync between devices.
A remote window or a Remote Access browser may need to download the model once. A different browser, device, or private window downloads it again.
Learned words and the languages this device has dictated in stay on the device in front of you, and never go to the remote machine.
See Remote Environments.
