Voice interaction

Voice in Mosael is three separate things. You can use just one of them:

	What it does	Where
Speak instead of typing	Puts what you said into the composer for you to review before sending	Next to the agent composer
Read this aloud	Speaks one reply	Under each reply
Hands-free	Always listening: sends each sentence when you finish it, then reads the answer back	The dock in the bottom-right corner

Speak instead of typing#

Click the microphone to start talking and click again (Stop and transcribe) to have the recognized text appended to the composer — appended, not replaced, because you may have typed half a sentence before switching to speech. It is not push-to-talk: the composer is where you think out loud, and holding a button forces you to have the whole sentence ready. One take runs up to 120 seconds; at the limit it wraps up and hands in what you have said so far.

Recognition follows whichever engine is set under "Settings → Capability providers → Transcription" (local FunASR, sherpa-onnx or WhisperX, or a transcription plugin) by default; you can also pick a separate one under "Settings → Agent → Spoken replies" · "Speech input engine" (the default is "Follow transcription", and only ready engines are listed). A missing transcription runtime, a denied microphone or silence each get a plain explanation.

It only fills the box; it does not send. Fixing a misheard word right there is the point: recognition will get things wrong sometimes, and sending straight through means the model has to guess what you actually meant.

Read this aloud#

Every reply has a play button underneath. Only one line plays at a time — start another and the first one stops immediately.

Which voice it uses is a setting; see "Dubbing and conversation are two settings" below.

Hands-free#

The dock in the bottom-right corner; click it to start. It has four states, told apart by the colour and motion of one set of sound bars:

• Listening — waiting for you to speak, breathing slowly.
• Hearing you — the bars follow your actual volume. The moment they cross half height and turn primary is the moment it decides you are speaking, so "how loud do I need to be" is not something you have to guess.
• Thinking — a spinner.
• Speaking — a ripple travelling outward.

Drag the dock anywhere; it remembers where you put it. Hovering shows the current state and the last thing it heard you say — so when it mishears, you see which word went wrong right then, instead of working it out from an answer to the wrong question.

You can interrupt it mid-sentence. Speak while it is reading and playback stops immediately, recording what you are saying. Otherwise you would have to sit through the rest of a wrong answer before correcting it, which is the single most irritating thing a voice assistant does.

Read-only tools can run from voice; anything that changes something still needs on-screen confirmation. Creating a project or editing a workflow raises a confirmation card. Accepting by saying "yes" would mean one misheard word actually changes your work — and you are probably not looking at the screen.

Failures speak up. In voice mode a silent failure reads as "it didn't hear me", so you say it again and fail again.

The dock is off by default; turn it on in Settings, under Spoken replies.

Dubbing and conversation are two settings#

The conversation voice is kept separate from the dubbing default, deliberately:

• Dubbing wants quality: a local zero-shot engine, a cloned voice, a fifteen-minute first load if that is what it takes — that audio is going into the finished video.
• Conversation wants latency: waiting a minute after you finish a sentence is not a conversation.

One default serving both is necessarily wrong for one of them. The Spoken replies section governs the conversation voice only; the dubbing default lives with dubbing. Read aloud on a passage in a note uses the Spoken replies voice too (the free Edge voice if you never chose one); see Notes → the selection toolbar.

Both draw their engine and voice lists from the same endpoints — this is another place to choose, not a second catalogue.

Conversation can use your own voice too. With the engine Alibaba Bailian (CosyVoice), the voice dropdown gains a My cloned voices group under the stock voices: what speaks is that library voice's copy cloned into your Bailian account, billed per character like a stock voice, with no local engine to install or wait for. The first time you pick a voice you're asked whether its reference clip may be uploaded — only once you agree is it cloned and the choice saved; if you don't, nothing is uploaded and the voice field is cleared so you can pick again. Voices nobody has declared ownership of aren't listed. Read aloud and Save as audio in notes speak with this voice. For cloning, copies and deletion, see Editing → Your own voice through Bailian.