Get a speech API key

Cloud speech sends completed speech phrases to a provider you choose, using your own API key. Whisper stays the local option. You do not need a cloud key for local speech or live drafts.

Groq

1. Sign in to the Groq console.
2. Open API Keys and create a key for Astral Overlay.
3. Copy it on the Windows desktop.
4. In Settings > Speech, set Voice engine to Cloud and Cloud service to Groq.
5. Press Paste key, then try the keyboard microphone.

Groq is a practical starting point for frequent dictation. Check your console’s Limits page and current limits; free allowances and paid usage can change. One finished phrase makes one request, so both phrase count and audio-duration limits matter.

Google Gemini

1. Open Google AI Studio’s API keys page.
2. Create a key in your chosen project, following Google’s key guide.
3. Check that the project can use the transcription model and has enough active rate limits.
4. Copy the key, choose Google Gemini under Cloud service, then press Paste key.

Small Gemini allowances can run out after only a short dictation session. Adding billing does not guarantee a useful speech allowance; check the model’s actual limits in AI Studio before relying on it. A YouTube-only key cannot be reused for Gemini.

Keep keys private

Paste from the Windows clipboard; there is no speech-key typing field. Each service remembers its own key. Forget key removes the selected service’s saved key. A key must come from that provider, not be a project ID or account password.

Keys are saved as plain text in settings.json. Do not share that file without removing them. Revoke a key in the provider console if it is exposed. See saved data.

Audio and provider privacy

Audio goes directly to the selected provider only when Cloud is selected, a usable key is present and a speech phrase is detected. Silence sends nothing; loud background noise can count as speech. Live drafts use local Whisper.

Groq’s data policy allows temporary input/output logs for reliability or abuse investigations, up to 30 days. Its Data Controls offer Zero Data Retention. Review your organisation’s settings.

Google’s terms distinguish paid and unpaid use, with regional exceptions. Unpaid content can be used to improve products and reviewed by people; paid content is not used for that improvement but can be logged for safety. Check the terms for your account and region.

If cloud speech fails

A missing or refused key, rate limit, network problem or unusable answer falls back to the downloaded local Whisper model. Keep that model installed as a safety net. The status in Settings > Speech explains the problem; see speech to text.