How to make an AI voice bot sound human and answer without pauses

The first seconds of a call decide whether the client hangs up. How do you make an AI voice bot sound human? Above all, it must not go silent before answering, interrupt or sound like an automated announcement: otherwise the client leaves, and you still pay for the minutes. The voice bot answers in about 1.5 seconds — in a lively voice with emotions, pauses and short “Uh-huh”s. Many callers never realize they were talking to a bot.

The speed holds even at peak hours. With many voice services, the reply stretches to 4–5 seconds in the middle of the day, and the robot gives itself away. This bot has no such slowdown. How fast, lively and cheap your particular bot sounds depends on the settings on the “Settings” → “General” tab of the bot card. Watch the two-and-a-half-minute video, and below you'll find the step-by-step setup.

Why businesses need it

• A clinic or salon picks a lively voice from the catalog or its own voice, plus office background noise. The client listens to the invitation to the end.
• An online store with short calls — a payment reminder, an abandoned cart — takes the “Fast” model and reply caching. Replies are short, and the minute is cheaper.
• A sales team with a complex product picks the “Smart, fast and pricier” model: it handles a long instruction better. Choose it only if the standard model gets confused.
• Any calling campaign saves on answering machines: the bot drops the call on silence and on spam phrases. Fewer minutes are paid for nothing.

How an AI voice bot prepares a reply

The client speaks — the system recognizes the speech and waits for a pause. When the pause is longer than the set value, the bot model prepares a reply. The bot model is the AI that writes the replies. Speech synthesis voices the reply in the chosen voice. While the reply is being prepared, the bot can say an interjection — a short phrase like “Uh-huh” or “One second, let me check” that fills the pause. The cache is a memory of ready-made replies: it serves repeated replies instantly, without generating them again.

How to set up the bot's voice and reply speed

1. 1
   Choose the model

   In the “Bot model” field (how the models differ is explained below). In the “Change the bot model?” window, click “Change model”.
2. 2
   Choose the voice

   Click the “Voice style” field, find a voice and play the demo with the ▶ button.
3. 3
   Listen to your own phrase

   Type the text into “Hear how the robot sounds” and click “Voice it”. You can download the finished recording.
4. 4
   Tune the pause

   Expand “Speech recognition settings” and move the “Silence (ms) after the customer stops talking…” slider.
5. 5
   Add interjections

   Expand “Interjections” and turn on “Use interjections”. In “Short interjections”, enter “Uh-huh” and “So”, and in “Long interjections” — “One second, let me check”.
6. 6
   Save and test

   Click “Save” and call the bot with a test call.

Short interjections play before every reply, in turn, one per reply — add several so the bot doesn't repeat itself. Long ones play only when the AI takes more than two seconds to prepare a reply. So write a waiting phrase a person would say: “Okay, one second, please.”

Which model and voice to choose for the bot

“Standard” thinks the reply over and is the default. “Fast” answers fastest of all. “Smart, fast and pricier” understands complex instructions better. The price is shown right in the list — per token. A token is a small piece of text that AI processes. “Standard” costs $0.00000031, “Fast” — $0.00000034, “Smart, fast and pricier” — $0.00000153. Per minute of conversation, the first two come to $0.04–0.06 on average, and “Smart” about five times more.

A full minute of conversation costs $0.12 on average: a long instruction, large loading webhooks and wordy replies for speech synthesis make it pricier, and a voice bot minute package makes it cheaper. Telephony is already included in these $0.12, and the cost of renting numbers for calling is covered in “AI outbound calling bot”. After changing the model, check the bot in “Training AI”: models understand instructions differently. How to do this is explained in “AI voice bot training”.

The voices in “Voice style” are split into groups: “My voices”, “Recommended voices” and “All voices”. The “Add your own voice” button opens the “Your own voice” section in a new tab — there is a separate article about it, “Voice cloning for an AI voice bot”. Some voices have “Additional settings”: “Stability” — from emotional to even sound — and “Similarity”.

How to fix the intonation of an AI voice bot

A flat, rushed delivery gives a robot away just like a manager reading from a script: the client figures it out and hangs up. Pauses, soft connecting words and a natural-sounding question keep the client in the conversation. Intonation is tuned in three stages. Change one thing at a time and listen to the result with the “Voice it” button or a test call.

1. 1
   The voice and its settings

   Start with the voice: different voices read the same text differently, and the “Recommended voices” group is a good place to begin. If the voice has “Additional settings”, try “Stability”: 0 is more emotional, 100 is more even, 50 is in between. The lower the value, the more the same phrase sounds different each time. “Similarity” is 75 by default and is usually left as is.
2. 2
   Periods, commas and pauses

   The bot reads a long text without punctuation in a monotone. Put a period or a comma where a person would pause: before a question, before “or”, even if grammar doesn't call for it. You can also insert a pause with the “{...}” button: “Add mini pause”, “Add middle pause” or “Add 1sec pause”.
3. 3
   Different words

   If a word still sounds odd, rephrase the sentence. A lone “Okay” sometimes sounds flat, while “Yes, okay” or “I can send it, okay?” sounds lively. Add to “Speaking style”: “In long replies, use short connecting words like ‘so’, ‘well’, ‘right?’” — they soften a memorized script and strengthen the question. Write abbreviations and numbers the way they should be pronounced: “S-M-S”, “twenty-five”.

How to set word stress and emotion tags is explained in “AI voice bot script”. If the bot speaks too slowly, speed it up in the “Speech pace” list on the “Instruction” sub-tab.

Why is my voice bot slow to respond, and how to speed it up

• “Model thinking setting” — from 0 to 128, 16 by default. Lower means faster. 128 is set if the bot mixes up dates, but then pauses are possible. The field exists only for the “Standard” model.
• “AI caching” is on by default. The first 10–20 calls cost more, and after about 50–100 dialogs the costs drop to a minimum. When you edit the prompt (the bot's instruction), the cache resets.
• “Silence (ms) after the customer stops talking…” — the silence before a reply, one value for the whole conversation: from 500 to 3,000 ms, 700 by default. Too short, and the bot interrupts the client; too long, and every reply comes later.
• “Minimum customer words before the AI can be interrupted” — 1, 2, 3 or 100, 1 by default: any word from the client, even “uh-huh” or noise, interrupts the bot. If the bot cuts off its own sentence and starts it again, set 2 or 3; any number above 3 is replaced with 100.
• “Silence sensitivity” — from 1 to 4, 3 by default. Values 1–2 suit clients who speak quietly.
• “Region (AI time zone)” — so the bot knows the current time and sets the stage date in the CRM correctly.

If the bot interrupts, stays silent or repeats the same phrases, find your complaint in the table in “AI voice bot troubleshooting”: for each one it says which setting to change and where.

Language, answering machines and background noise

• “Which language to recognise” — Russian, English, Portuguese, Turkish or Kazakh. “What recognises the speech” — “Yandex” (the main service) or “Google” as a backup if Yandex understands the language poorly or is unavailable.
• “Hang up if the customer stays silent in the first answer” — 14 seconds by default. It cuts off answering machines that pick up and stay silent.
• “Spam phrases” — answering machine phrases like “leave a message after the tone”. Once a phrase is heard the set number of times, the call ends.
• “Detect an answering machine in the first seconds of the call” — uncheck it only if calls drop in the first seconds. This happens when a number plays its own announcement before connecting.
• “Background sounds” — an mp3 file up to 4 MB, such as office noise, and its volume relative to the robot's voice.

Read also: Voice cloning for AI voice bots: make the bot speak in your voice, How to train an AI voice bot to answer like your best sales rep, AI outbound calling bot: call your contact list without wasting budget