Voice messages

Customers can say their question instead of typing it. The platform turns a customer's voice message into text before anyone reads it, so the AI assistant answers it like a typed question and your agents see what was said without listening.

Where voice messages work

• Telegram and WhatsApp: customers send voice notes as usual.
• Website chat and mobile app chat: the microphone button next to the message field records a voice message.
• Email: an audio file attached to an incoming email is turned into text too.

Only messages from customers are transcribed, and only audio. Video messages, such as Telegram round videos, are not.

What the customer gets back

The assistant reads the text of the voice message and answers in text, as if the question had been typed. A conversation can also start with a voice message. The assistant never answers with voice, and the platform makes no phone calls; a customer who wants to talk can ask to be called back (see Call me back below). In Telegram an agent can also reply with a recording of their own: see Telegram.

The microphone button in the chat

In the website chat and in the mobile app chat, the microphone button appears next to the message field when all three are true:

1. Speech recognition is set up. A new platform uses ours, see below.
2. File upload is on in the live chat inbox, tab Conversations, section Features.
3. Voice messages is on in the Designer tab of the inbox, set separately on the Website and App tabs.

On a website the microphone is in the chat of the live chat inbox your platform came with and of every live chat inbox you add later, since they get the same chat. Only an inbox you switch to the platform's standard chat window, by clearing Chat widget address and Designer address, has no microphone: see Live chat widget.

The first tap asks the browser or the phone for permission to use the microphone. If the visitor refuses, the button stays, and a hint under the input field suggests allowing the microphone or typing instead. Your website code needs no change: the chat window is already allowed to use the microphone. Only a site whose own security headers forbid the microphone in embedded frames blocks recording, and the visitor then sees the same hint.

In the mobile app

The chat package asks the phone for microphone permission at the first tap on the button. The app developer adds one line for Android and checks one text for iOS:

• Android, in AndroidManifest.xml:

<uses-permission android:name="android.permission.RECORD_AUDIO"/>

• iOS, in Info.plist: the NSMicrophoneUsageDescription text should mention voice messages.

Without the Android line the chat works as before, but recording does not start and the hint suggests typing. Voice messages need version 0.3.0 of the chat package or newer. See Mobile app chat.

What agents see

In the conversation the message shows the text with the mark Voice message transcript and the recording, which can be played right there; its link opens in the browser's player instead of downloading a file. If recognition failed, the mark reads Could not transcribe the voice message, and the agent listens to the recording. When speech recognition is not set up at all, the message shows the recording without a mark. The transcript is ordinary message text, so conversation search and the AI helpers for agents work with it.

Speech recognition provider

A separate model turns speech into text. On the trial the platform uses ours, free of charge; as with our answering model, we do not promise that it is always available. After payment the cloud AI of your plan transcribes voice messages from its budget, and ours takes over when the budget runs out.

To use your own provider, open AI > Providers and press Add provider in the Speech recognition list. Ready presets are OpenAI gpt-4o-mini-transcribe and Groq whisper-large-v3-turbo, and any service with an /audio/transcriptions endpoint works. Such providers charge less than a cent per minute of speech. Providers in the list are tried from top to bottom, and Check sends a second of silence to test the connection. A model on your own hardware is a separate service: see AI providers and your own model.

You do not set a language: the model hears it, and a customer may speak a language other than the one of the conversation.

If recognition fails

A voice message is never lost. When recognition is not set up, does not answer in time or hears only silence, the message arrives without text: the agent listens to the recording, and the assistant asks the customer to type the question.

Limits

• A recording in the website chat or the app lasts up to 2 minutes. At the limit it stops and is sent, so nothing said is lost.
• A voice message is transcribed if it is up to 25 MB, and the platform waits up to 45 seconds for its text. A larger or slower one arrives without text.
• In the website chat and the app a recording also follows the upload limits of the platform in Admin > General: Max allowed file upload size and Allowed file upload extensions.
• Telegram lets bots receive files of up to 20 MB.

Call me back

A customer who prefers to talk can ask for a call. The assistant asks for the phone number with the country code if it has none, notes a convenient time if the customer names one, then hands the conversation to your agents with a note such as "Asks to be called back: +442079460123. Convenient time: after 3 pm." The customer is told in their language: "I have passed on your request: you will get a call back." An agent then calls from their own phone, because the platform itself makes no calls.

• A phone number already known from the contact card, for example from sign-in by phone in your app or from WhatsApp, is not asked again.
• The number must have 7 to 15 digits, and 00 at the start counts as +.
• The assistant arranges a call back only while Offer handoff to a human is on in AI > Assistants, and never in an AI chat without agents, where nobody would call.