Voice AI

Speaking is faster than typing and works when your hands are busy, which is why voice keeps returning as a business proposition. What has changed is that the middle stage can now hold a real conversation rather than matching phrases, so the interesting uses moved from commands to genuine dialogue. What has not changed is that a chain of three systems accumulates delay and errors at every joint, and those two problems are what separate a demonstration that impresses from a deployment your customers do not complain about.

The first stage turns sound into text, and its errors are the ones nobody sees. A misheard name or figure is passed downstream as though it had been typed, and everything after it reasons confidently about the wrong word. Accents, background noise, technical vocabulary and proper nouns are where this stage is weakest, and unfortunately those are also where most of your business conversations live.

The middle stage is an ordinary language model working on the transcript. It does not hear tone, hesitation or emphasis unless the system was built to pass those along, which most are not. That is why a voice assistant can miss sarcasm, urgency or uncertainty that any person on the phone would catch immediately.

The third stage turns the reply into speech, and the quality of that voice shapes expectations more than anything else in the chain. A natural-sounding voice makes people assume more competence than is there, which is useful for adoption and dangerous for anything where a misunderstanding matters.

Delay accumulates at every joint, and it is the thing that makes a conversation feel wrong rather than slow. People tolerate a short gap and start talking over anything longer, so systems have to begin responding before they have finished thinking. That is why many of them start with a filler phrase, which is a timing device rather than politeness.

Handling interruption is the hardest part and the clearest test. A person who is cut off stops, listens and adjusts. A system has to notice speech during its own output, decide whether it was an interruption or a cough, stop cleanly and carry over what it had already committed to. Products differ enormously here, and it is worth testing yourself rather than reading about.

Some newer arrangements handle speech directly rather than converting to text and back. That removes joints and preserves tone, and it changes what can be inspected, because there is no transcript sitting in the middle to check, log or correct. Which arrangement a product uses is a fair question, with real consequences for anything you have to audit later.

• Hold a spoken conversation with an assistant and deliberately interrupt it mid-sentence, which tests the part of the chain most demonstrations avoid.

  ChatGPT

• Dictate a paragraph containing three proper nouns and a figure, then read the transcript rather than the reply, which shows you what the later stages were actually given.

  Google Gemini

• Wire a transcription step and a language step together in an automation yourself, where the handoff between them is something you can see and log.

  n8n

Why does it pause before answering?

Because three things happen in sequence and each takes time: your speech has to end and be transcribed, a reply has to be generated, and that reply has to be turned into audio. Products disguise the gap with a short filler phrase, which is a timing device rather than conversational warmth.

It keeps mishearing our product names. What can be done?

Quite a lot, and it belongs to the first stage rather than the model. Most transcription systems accept a list of expected terms, names and jargon that biases recognition towards them. Supplying that list is usually the single most effective fix, and it is often not switched on by default.

Can it interrupt, and be interrupted?

That varies more than any other feature and is the best thing to test. Speak over it mid-sentence and see whether it stops cleanly, whether it understood what you said while talking, and whether it carries on sensibly afterwards. Products that handle this well feel entirely different from ones that do not.

Is the audio kept, and who can hear it?

That depends on the provider and the plan, and voice raises the question more sharply than text because a recording carries a person's identity in a way a transcript does not. Ask what is stored, for how long, whether transcripts and audio are treated differently, and what a caller is told.

• It is nearly always three systems in a row, and most problems belong to one joint rather than to the whole.
• Mishearings at the first stage are inherited with confidence by everything after them.
• The middle stage usually receives words with no tone attached, so emphasis and hesitation are lost.
• Delay accumulates at every joint, which is why filler phrases exist.
• Interruption handling varies enormously and is the most revealing thing to test.

• ChatGPT

  Spoken conversation you can interrupt deliberately to test the hardest part.

• Google Gemini

  Dictation where the transcript is visible, so you can see what later stages received.

• n8n

  Join the stages yourself, where each handoff can be inspected and logged.