How does speech become a spoken reply?
There are two designs. The older one chains separate programs: speech recognition writes down what you said, a language model writes a reply, and a speech synthesiser reads it aloud. Each hand-over adds delay, and the written middle step throws away your tone of voice.
The newer design is a single speech-to-speech model that takes audio in and gives audio out, so it hears hesitation and emphasis and can answer sooner. OpenAI's Realtime API, released in October 2024, works this way, and Google's Gemini Live uses the same approach. Underneath, both are multimodal models: sound is simply another kind of input.
Why is knowing when you have finished the hard part?
Because people pause mid-thought. A voice AI listens for the end of your turn with a detector that notices when speech stops, called voice activity detection, and better systems add a model that judges whether the sentence sounded finished. Set it too eager and it cuts in while you are reaching for a word. Set it too patient and every reply starts with dead air, the delay covered under latency.
Then there is interruption. If you start talking while the AI is mid-answer, it has to stop, listen and drop the rest of what it was going to say. Background noise makes all of it harder, as anyone who has taken a call beside a running compressor knows.
What is a mobile mechanic's missed call worth? (example figures)
A mobile mechanic in Gladstone misses about 12 calls a week while under a car. Half are new jobs, about half of the new jobs that get answered turn into work, and the average job is $420.
| Sum | Result | |
|---|---|---|
| Missed calls a week | 12 | |
| New jobs among them | 12 × 0.5 | 6 |
| Jobs that would have been won | 6 × 0.5 | 3 |
| Work missed each week | 3 × $420 | $1,260 |
| A voice answering service at $150 a month | $150 ÷ $420 | Covered by one job a month |
Those numbers are why owners look at voice AI for the phone. The cautions are a caller who wants a person, a quote the AI should never give, and a booking made into a time that is already taken. A sound set-up says plainly that it is an AI, takes the details, books only into free slots and never names a price the mechanic has not set.
How is talking to an AI different from typing to one?
You say more. People talk through background they would never bother typing, change their minds halfway through a sentence and fill pauses with "um", and a voice AI has to follow all of it. A voice conversation also runs one way: you cannot scroll back up, so a good voice AI keeps answers short and checks before moving on. Typing leaves a record on the screen, while a call leaves only what the product chooses to keep, so find out whether that is the audio, a transcript or neither.
What is it like to talk to Cora?
CoraVision is voice AI in exactly this sense: you talk to Cora out loud and Cora answers out loud, so a session runs like a call. The onboarding call takes about 25 minutes and needs no preparation. During a session you start, Cora sees your screen as still images, so you can talk about whatever is in front of you, and those images are not saved to your account. Transcripts of your calls are kept with your account.
Questions
Does a voice AI keep a recording of what I say?
It depends on the product, and its privacy policy should say. Some keep the audio, some keep only a written transcript, and some keep nothing after the call. CoraVision keeps transcripts of calls and chats with your account.
Should a voice AI tell callers it is an AI?
It should, and good services say so in the first sentence. A caller who finds out halfway through feels tricked, and a business that lets customers believe a machine is a person risks misleading them, which the Australian Consumer Law prohibits.