Voice Input & Transcription

Voice input records from the default microphone and places a transcript in the message composer. Review the text before sending it.

Local Whisper#

For local transcription, Sythoria:

1. records audio with the system input device;
2. mixes the recording to one channel;
3. resamples it to 16 kHz; and
4. sends the samples to whisper-rs for transcription.

The local path does not use Silero voice-activity detection. It also does not create a 16-bit WAV file before local inference.

Settings offers Tiny, Base, Small, and Large v3 Turbo Whisper models. English-only and multilingual choices are available for the smaller sizes. Preset downloads are checked against the values built into the app. A custom model is accepted as an unverified local file, so obtain it from a source you trust.

Recordings are limited to five minutes.

Cloud transcription#

The cloud path converts the captured audio to a 16 kHz WAV file and uploads it to the configured speech-to-text endpoint. The default fields point to Groq's OpenAI-compatible transcription endpoint and whisper-large-v3, but both can be changed.

The cloud service receives the recording, its credential, and normal request metadata. Check its pricing, retention, and region before using private speech.

Transcript refinement#

Voice settings can send the transcript to one of your configured language models for cleanup. This option is enabled by default in v0.4.4. If the refinement model is hosted, the transcript leaves the device even when Whisper transcription itself was local.

Turn off LLM polish when the transcript must stay on the local transcription path.

Troubleshooting#

Recordings and transcripts can contain names, notifications, or background conversations. Remove private text from the composer before sending it to a model.