Voice dictation: recognition, polish, and history

Speech becomes text before it becomes a task#

Dictation turns speech into editable composer text. The assistant executes the message after you review and send it.

Prepare the microphone and recognition model#

1. Open Settings → Voice in the desktop app.
2. Turn on Enable voice input, allow the operating system’s microphone permission, and select the system default or your intended device under Microphone.
3. Click Manage speech recognition models. In the dialog, download a model for your main language, wait for it to become ready, then click Use and confirm it is marked Active.
4. Return to the composer and test a short sentence.

The initial model download needs network access; recognition runs locally. For Chinese, start with SenseVoice, or try Paraformer Large for Mandarin and English. These models transcribe after you stop recording, so an absence of live text is expected. Choose a streaming model from the list if you need real-time recognition.

Downloaded does not mean selected. Test a short sentence with numbers and proper names before switching models.

Choose whether to use AI polish#

Polish sends transcript text to the selected model and can incur model usage. It is separate from offline recognition. If it fails or times out, inspect the retained raw text before sending. Configure the connection in Providers.

Add correction rules and vocabulary

1. In Settings → Voice → Correction preferences, click Manage.
2. Write a short rule under Custom correction instructions, such as Keep product names and acronyms in English. Add punctuation without translating or expanding my words. You can use Fill from a template as a starting point; it fills editable rules and reference words rather than selecting a permanent industry mode. Review the text and click Save preferences.
3. Under My words, click Add a word. For a fictional product, enter Northstar as the correct spelling and North Star as a common misspelling, then click Add word. A word can have up to three comma-separated misspellings; manual words take priority.
4. Reopen the panel to check that both the instructions and the word were saved. Dictate a short sentence containing the name and inspect the result before sending. A saved preference is guidance for correction, not a guarantee that every occurrence will be recognized correctly.

These preferences apply when voice input and AI polish are enabled. You can maintain them while correction is off, but they do not replace the local recognition model.

Review what is learned from your edits

With AI polish enabled, clear terminology corrections you make to dictated text before sending can become examples for a later correction using the same model. Sending does not trigger an additional AI request. A word need not appear in Learned words immediately, and not every edit qualifies: ambiguous one-character changes, numbers, negation, and broad rewrites are excluded from automatic learning.

Use Correction preferences → Manage → Learned words to inspect, edit, disable, or delete a learned entry. Clear learned words also removes pending learning examples while retaining manual words and custom instructions; its confirmation describes that scope. If a preferred spelling is important, add it explicitly under My words instead of relying on learning.

Dictate your first task#

1. Open the intended project and conversation, then click the composer microphone. Defaults are Shift+Command+D on macOS and Ctrl+Shift+D on Windows/Linux. Customized bindings take precedence.
2. Speak a small task:

Read the attached error report.
Summarize reproduction steps and missing information.
Do not modify files yet.

3. Stop recording and wait for recognition and any enabled polish. The composer stays locked until the text is inserted once.
4. To skip the polish wait, press Escape and keep the raw transcript. Failures and timeouts also retain the raw text.
5. Check the text in the intended composer, correct it, and send.

Confirm the transcript is correct#

Check numbers, paths, proper names, and constraints such as “do not” or “read only.” Confirm the text entered the intended conversation rather than another split composer.

A transcript means input is ready, not that a task executed. After sending, verify the assistant's actual scope and result.

Inspect history and local recordings#

Open Settings → Voice → Dictation history and find the record by time:

• Click its copy icon to copy the currently used text.
• Choose More actions → View text versions to inspect Used text, AI corrected text, and Original transcript. Additional versions appear only when their text differs; select the version you need and click Copy transcript. The original recording is retained.
• Choose More actions → Download recording, then open the saved audio to verify it.
• Choose More actions → Delete recording when you no longer need that recording record.
• Use Open recordings folder to inspect local audio, or page through older records.

Local recognition does not mean recordings are never retained. Review audio and text before sharing. Do not assume Privacy → Clear Local Data also removes voice recordings.

More voice controls#

Turning off Enable voice input hides the composer microphone and disables the dictation shortcut. Turn it back on in Settings when needed. If a microphone is missing, allow OS access and check its connection. See the Voice settings reference for individual options.

Common problems#

• Cannot record: check OS microphone permission and the input device. Confirm Enable voice input is on.
• No transcription: confirm the model is downloaded and active, then test a short recording.
• Poor recognition: choose a suitable language model, reduce noise, then consider polish.
• Polish failure: check the provider and model; retain and review the raw text.
• A vocabulary change is missing: distinguish Save preferences from Add word / Save word, reopen the panel, and check AI polish is enabled. Learned words may only appear after a later correction.
• Waiting for polish: press Escape to retain the raw text, then check the polish model connection; do not restart the same recording repeatedly.

Next steps#

Check the Voice reference and attach task material through the tools menu. Review the complete request before sending it.