Back to blog
Talking to LIA when typing is a pain
Features4 min read

Talking to LIA when typing is a pain

Why voice

There are moments when typing is a pain: hands covered in flour, in the car, while walking. Speaking is more natural. But voice raises two questions a keyboard doesn't: where does the recording of what you say go, and how much does each sentence read aloud cost?

What I wanted

I wanted voice to be another way of using the same assistant, with the same memory, the same confirmations and the same limits as the written chat. And I wanted your voice to stay on your server by default.

Dictate and listen

The simplest way: hold the send button, speak, let go. LIA transcribes and puts the text in the input box, where you can reread it before sending. You can also turn on hands-free mode from the chat header: one tap starts listening, which stops by itself after a short silence. With the interface in French, you can say “Dis LIA” to talk to her, or “LIA, stop” to cut her voice while she's reading. These phrases are recognized in your browser, without sending any sound.

Transcription happens on your server by default, with a local speech recognition model: it's free, and the audio doesn't leave. If you want more accuracy, in a noisy environment for example, you can choose in the settings the remote engine your administrator has configured. The settings name it, so you know who receives your voice, and the cost of each transcription shows on your bubble.

The other way round, if you turn it on, LIA reads her answers aloud with the voice you choose for her. The engine is configured by the administrator; there's a free one, and for the others a small badge shows what the reading cost. Reading starts as soon as the first sentence is ready, while the rest is still being written. And LIA knows that the period in “3.5 degrees” doesn't end a sentence.

A real conversation, in real time

Live mode goes further: you speak, LIA answers right away, and you can interrupt her. It runs on a real-time voice model you connect with your own key (Gemini, OpenAI or an ElevenLabs agent). I made a simple choice: the voice holds the conversation, and LIA does the work. When you ask for information or an action, the voice passes the request to LIA, who handles it like a chat message, with the same confirmations. The answer shows up in the thread while you talk, and if LIA has a question for you, it's the voice that asks it.

After a silence, the session goes to sleep without stopping: the connection to the provider closes and nothing is billed. One tap wakes it up, with LIA's current context. You're the one who ends the session, unless it sleeps for too long. Either way, a card sums up its duration, your requests and what LIA spent.

Where the model allows it, a Live direct session lets the voice read your data itself (calendar, e-mails, contacts, weather) in the domains you allow, without acting on anything. Your words are then relayed to the chat as a message from you. It answers faster, but it costs more at the provider.

In both modes, and when LIA calls you, the voice speaks in the personality you chose for her: her words, her manner and her humour follow it, the facts never do. When she calls someone else on your behalf, she stays cordial.

When LIA calls you

If you've set up telephony, you can also declare your number: LIA calls you once and reads you a code that you type back in. From then on, “call me at 8 a.m. to go over my day” makes your phone ring, including from a routine. What you ask during the call comes back to the chat as a message from you, with the same confirmations. The details are in the phone FAQ.

What goes out, and what it costs

  • In Live mode, your voice goes from your browser straight to the provider you chose, which bills you on your key. LIA shows you an estimate when the rate is known, without recording it, since it's your key.
  • What LIA herself spends to handle your requests counts toward your usage, like a written message. Remote transcription and paid reading count too, with no separate quota.
  • Hands-free mode and reading aloud are off at first: you're the one who turns them on.
  • The phrases “Dis LIA” and “LIA, stop” are in beta and only exist in French for now. In other languages, just tap the button.

Read more: Your data stays yours: what LIA keeps, shows and erases; Reminders and routines: let LIA take care of the timing.

#voice#speech-to-text#tts#wake-word#hands-free