Skip to content
OpenChamber
↑↓Navigate ↵Open escClose

Voice Mode

Voice mode lets you dictate messages and have replies read back to you. Turn it on at Settings → OpenChamber → Voice.

Reading replies aloud (text-to-speech)

Pick how replies are spoken:

  • browser — your browser’s built-in voices, no setup
  • OpenAI — OpenAI’s voices; paste your API key and pick a voice
  • OpenAI-compatible — any service that speaks the OpenAI format; enter its URL and, if it needs one, an API key. If the service names its speech endpoint differently, put that path in Endpoint path; leaving it empty keeps /audio/speech
  • macOS say — the built-in say command, on Macs that have it

Once on, you get a play button on messages to hear them read aloud.

Dictating messages (speech-to-text)

Pick how your speech is transcribed:

  • Local runs a speech model on the computer where OpenChamber runs. The model downloads the first time you dictate, and your audio stays on that computer.
  • Server sends your audio to an OpenAI-compatible transcription service. Enter its URL and, if it needs one, an API key. If its transcription endpoint is not /audio/transcriptions, set Endpoint path. Leaving it empty keeps the OpenAI default.

Local models:

  • Parakeet v3 is the default. It knows 25 European languages, English and Ukrainian among them, and is the fastest. If you already dictated with Parakeet v2 before v3 became the default, you stay on v2 until you pick another model.
  • Parakeet v2 knows English only.
  • Whisper large-v3 turbo knows 99 languages. It is about nine times slower than Parakeet, so long dictations take a few seconds to appear.
  • Qwen3-ASR 0.6B knows 30 languages, including Chinese, Japanese, Korean and Arabic. It has no Ukrainian.
  • Whisper tiny is the smallest download, for slow computers.

The Whisper and Qwen3-ASR models detect the language on their own, and on a short phrase they can pick the wrong one. Enter your language code in Language below the model list, for example uk or en, and they stop guessing.

Turn on Fix misheard words to have your Small Model (Settings → Sessions → Small Model) go over the text after you stop. It replaces words the speech model misheard and drops filler sounds such as “uh”. The text then goes to the Small Model’s provider and appears a second or two later. Without a Small Model the text is inserted as heard.

API keys you enter for these services are kept by the OpenChamber server, not in your browser. Type a key and press Enter to save it; the field then shows that a key is saved, and the cross removes it.

A note for phones

For reading replies aloud on a phone, the OpenAI or OpenAI-compatible options are the most reliable — mobile browsers limit the built-in voices.