Home · Blog · Voice messages in a chatbot: how the bot understands customers’ voice
Voice messages in a chatbot: how the bot understands customers’ voice
Yes, a CHBT bot understands voice messages: in Telegram, VK and MAX a voice message up to 30 seconds is transcribed into text and goes through the flow just like typed text — Keywords catch it, Data collection accepts it and the AI agent reads it. Transcription works on AI plans — Basic, Business and the trial — and costs 1 AI credit per message. Nothing needs to be turned on. On the Free plan voice messages are not transcribed: the bot receives an empty message, and the operator sees a “🎤 Voice message” mark in Conversations. Below — where exactly transcription works, its limits, how to set up a flow for voice and how to test it.
Where transcription works
| Channel | Voice messages |
|---|---|
| Telegram, private messages | Transcribed |
| Telegram, groups | No — in a group the bot answers only commands addressed to it and its own buttons |
| VK, community messages | Transcribed |
| MAX | Transcribed, but MAX does not always send the recording in a suitable format — then transcription fails |
| Website chat, email | No |
Transcription is needed only for messages that contain nothing but voice. Button taps are not transcribed — they are not voice.
Plan, price and limits
- Plan: Basic (650 ₽ a month per bot), Business and the trial — 3 days of the Business plan that turn on by themselves when the first channel is connected.
- Price: 1 AI credit per voice message — the same as an AI agent reply on the default model. If recognition fails, no credit is charged. If the recording is recognised but no words can be made out, 1 credit is charged.
- Length: up to 30 seconds. Longer — the history shows “🎤 Voice message — longer than 30 seconds, not transcribed”, no credit is charged.
- Size: up to 1 MB.
- Language: Russian. Recognition is done by SpeechKit.
- Credits ran out: voice messages are not transcribed, the history shows a “🎤 Voice message” mark. After the balance is topped up, transcription resumes.
What the operator sees in Conversations
| What happened | What the conversation history shows |
|---|---|
| Transcribed | “🎤 I would like to book for Saturday afternoon” |
| Longer than 30 seconds | “🎤 Voice message — longer than 30 seconds, not transcribed” |
| Words not made out (1 credit charged) | “🎤 Voice message — could not make out the words” |
| Free plan, no credits or a recognition failure (no credit charged) | “🎤 Voice message” |
If the mark has no text, reply to the customer yourself: “We could not listen to it — please type your message”. If the conversation is with an operator, the bot stays silent, and the voice message with its transcription simply appears in the chat.
How to set up a flow for voice
- Keywords. In speech words come in different forms. Use Contains a word mode and several comma-separated variants: “price, prices, cost”.
- Data collection — name. Works by voice without settings: the answer is saved to the card.
- Data collection — phone. Recognition may write the number in words or with spaces, and the Phone check will not accept it. In the question text write: “Please type the number in digits”.
- The AI agent receives the transcription as a regular question. A voice question with an agent reply on the default model costs 2 credits: 1 for transcription and 1 for the reply.
- Greeting. Suggest the format: “You can type or send voice messages up to 30 seconds”.
If there is no transcription
On the Free plan, for voice messages longer than 30 seconds and when recognition fails, the bot receives empty text. So that the customer is not left without an answer:
- In the greeting write: “Please type your messages — the bot does not understand voice yet” (for Free).
- Build the key flow steps on buttons — a tap does not depend on voice.
- Turn on Push for every customer message for yourself in Account → Notifications, so an operator notices the “🎤” mark in time and answers personally.
How to test
You cannot send a voice message in the Check flow test chat — it has only text and buttons. Test from a messenger:
- Make sure the plan includes AI and there are credits on the balance.
- Open the bot in Telegram, VK or MAX from your account and record a short phrase with a flow keyword.
- Expected result: the bot answered according to the flow, and Conversations show the message with a “🎤” mark and the transcription.
- Send a voice message longer than 30 seconds — the history shows a mark without transcription.
If transcription works in Telegram but not in MAX, MAX sent the recording in an unsuitable format.
What CHBT does with the recording itself
CHBT downloads the recording from the messenger’s servers only for recognition and sends it to the speech recognition service. The conversation history keeps text — the transcription or a mark with the reason. CHBT does not write the Telegram file link, which contains the bot token, to logs and does not store it.
A customer’s voice is also their data. If your data processing policy lists what the bot collects, mention voice messages that are converted to text. How to draw up a policy — in the article 152-FZ and a chatbot.
AI agentConversations and operatorsNotifications about new messagesPlans