Everything under Settings → Transcription controls how audio becomes text. All transcription runs on-device. Nothing is sent to the cloud.
Transcription model
Two models ship with talat:
- Parakeet TDT v2 (English). The default. Fastest and most accurate for English-only audio.
- Parakeet TDT v3 (multilingual). Pick this if your meetings are not in English, or mix English with another language. Slightly slower and slightly less accurate on pure English.
Changes take effect immediately or after the current recording ends.
Live preview
Show a real-time draft of what’s being said, before the final high-quality transcription runs. When off, speech activity dots appear instead. On lower-spec machines the toggle may be disabled, and the setting will tell you why.
For non-English or multilingual meetings, turn live preview off for now. The preview model sometimes tries to translate the speech into English, producing gibberish on screen until the final transcription overwrites it.
Speaker identification
Tell voices apart in system audio. Turn this on for calls with more than one person on the other side so you can name speakers and have talat recognise them in future meetings. For 1:1 calls it’s not needed.
On lower-spec machines the toggle may be disabled, and the setting will tell you why.
Speaker matching sensitivity
When speaker identification is on, this slider controls how closely a voice must match a known speaker to be identified as them. Lower values require a closer match. The default (0.55) suits most users. Raise it if talat is too eager to match, lower it if it keeps creating new speakers for the same voice.