Voice Input

Click the microphone beside the composer to start dictation, then click again to finish. Choose the recognition provider under Settings → Providers → Capability settings → Speech recognition. Settings → Voice Input contains personalization and the dictionary, without a current-provider summary or controls for recognition mode, additional refinement or recognition language.

• Dictation keeps the selected recognition provider. Streaming sources display interim text while you speak and preserve the complete recording transcript at completion. An unavailable source reports recovery guidance rather than choosing a different account.
• Supported realtime sources receive dictionary, phrasing preferences and conversation context when recording starts. The same voice session returns the complete refined result. Normally that result is used directly. If it is slow, a text model refines the complete recognized text in parallel, and the first complete refinement wins. Text refinement and correction learning use the speech refinement pick under Default models → Auxiliary task models, including subscription and custom sources. An unset pick follows the dictating chat's main model; standalone dictation uses the remembered new-chat default. Each request reads the current setting, while in-flight work retains its captured model. A short delay never replaces pending refinement with raw text. Sources without native refinement use that text model; unavailable models and failed requests preserve the recognizer text.
• Personalization learns general phrasing preferences from corrections. Disabling it requires confirmation from its options menu; enabling it is direct. Disabled personalization uses generic phrasing while dictionary corrections remain available. There is no separate voice-memory list.
• After insertion, a short correction window observes the original composer. Further edits reset an approximately 15-second idle deadline; leaving the composer or sending also checks for corrections. Unchanged text, cleared drafts and wholesale rewrites are excluded. Learning requires a usable text model.
• Dictionary supports adding and editing terms, search, and All, Auto-added and Manually-added filters. A term can contain up to 100 UTF-16 units. Adding words does not evict existing terms. Checkboxes select words for deletion rather than enabling recognition; single and selected-word deletion require confirmation.
• Automatically added words show an Undo notice scoped to those new word ids.
• CSV import accepts one column with one term per line and no header. Files can contain up to 5 MiB and 3000 distinct terms per import. New terms are previewed before importing. Overlong terms are skipped; multiple nonempty cells on one row reject the file.
• Recordings are bounded to five minutes. Transcription history is stored locally for 30 days by default to help distinguish recognition output from refinement changes.