Speech input and read-aloud
Speech input turns your voice into text; read-aloud turns reply text into audio. Enable them separately and choose different sources. Your chat model does not need to support both.
Choose a source
Open Settings → Voice.
| Capability | Sources | Prepare |
|---|---|---|
| Speech input | Available transcription models, including capable plugins | An enabled model, microphone permission, and an input device |
| Read-aloud | System voices, or available speech-synthesis models and plugins | A usable voice and audio output device |
Speech models need the correct purpose, authentication, and protocol support. Enabling a chat model alone is not enough. A service's chat login does not necessarily provide a separate speech API. System languages and voices depend on your operating system and device.
Configure speech input
- Enable speech input and select a transcription model.
- Allow microphone access in the operating system. Use Test microphone and check the level meter.
- Use the short transcription test, finish recording, and inspect the text. Select a language if the source supports it.
- Use speech input in a task, then check the transcript before sending.
A microphone test confirms audio capture, not transcription availability. Test a short phrase before a long recording. Recognized text may remain after an interruption; check it before adding the same words again.
Configure read-aloud
- Enable speech output and choose a system voice or synthesis model.
- Choose a voice, enter a voice ID, or adjust the supported rate settings.
- Use Preview to check volume, language, and pronunciation.
- Use read-aloud on a reply with readable text. Cancel during preparation or stop playback.
If the voice list fails to load, enter an ID manually only when you know a valid one. If no system voices are available, check the system's speech components or choose a configured model source.
Where the data goes
Remote transcription sends recordings to the selected service; remote synthesis sends the text to be read. System read-aloud does not use your selected chat model; its capabilities come from the operating system. Corresponding local plugin inference runs on your computer, but preparing resources may use networking.
Transcription tests and previews make real calls to the selected source and may incur remote-service charges. Local speech does not make chat, web tools, or other features offline.
Recording or playback fails
- No microphone level: check system permission, the default input device, mute settings, and device availability.
- Recording works but no text appears: check model availability, language, credentials, credits, and networking; check the engine for local sources.
- No playback: test the output device and volume, then check the voice ID, model availability, and error.
- Interrupted connection: keep recognized text and retry manually after fixing the cause. Interruption does not establish that no charges occurred.
See Troubleshooting.