## Overview

Audio transcription turns conversation audio into text and shows that text under the player, inside
the message bubble itself. Agents read the message without pressing play — useful in a noisy room,
when handling several conversations at once, and for finding an old voice note later by what was said
in it.

It works on **any channel that can receive audio**: WhatsApp (Cloud and Web), Telegram, SMS, Line,
Messenger, Instagram, email, the website widget, the API channel and call recordings. The setup is the
same for all of them.

Three decisions, in this order:

- **On or off** on the inbox. Every new inbox starts off.
- **Which audio it covers**: what the contact sent, what the agent recorded, or both.
- **An exception per conversation**, when one conversation needs to behave differently from the rest of
  the inbox.

## Prerequisites

- An **OpenAI** or **Groq** API key available to the account. Configure it in
  **Settings → Integrations**, or ask the platform operator to configure it on the server.
- Administrator access to configure the inbox.
- Access to the conversation to create a one-off exception.

The screen tells you whether a key exists: a **Key configured** or **No key** badge appears at the top
of the tab. Without a key, transcription cannot produce any text at all — so it is worth checking that
badge before turning the feature on.

## Step by step

### Turn it on for the inbox

1. Open **Settings → Inboxes** and select the inbox.
2. Open the **Audio transcription** tab.
3. Check the key badge at the top. If it reads **No key**, use the link to configure the integration
   before continuing.
4. Turn **Audio transcription** on.
5. Choose the scope with the two switches below:
   - **Transcribe audio received from contacts** — on by default.
   - **Transcribe audio recorded by agents** — off on new inboxes; inboxes that already existed
     come with it on, preserving what they were already doing.

Each switch saves on its own, the moment you change it.

### Adjust a single conversation

1. Open the conversation and expand **Audio transcription** in the right sidebar.
2. The panel shows the current policy and which audio is being transcribed there.
3. Use **Adjust for this conversation** and pick the policy:
   - **Follow the inbox** — the default.
   - **Always transcribe** — applies even when the inbox has transcription switched off.
   - **Never transcribe** — turns it off for this conversation only.
4. With **Always transcribe** you can still set each scope separately, or leave it inheriting the
   inbox.
5. Click **Save**.

## Settings and options

| Where | Option | Default | What it does |
|---|---|---|---|
| Inbox | Audio transcription | Off | Master switch for the inbox. |
| Inbox | Audio received from contacts | On | Transcribes what the contact sent. |
| Inbox | Audio recorded by agents | Off on new inboxes | Transcribes what the agent recorded. Inboxes that already existed keep what they were doing: on. |
| Conversation | Policy | Follow the inbox | A one-off exception that leaves the inbox untouched. |

A **new** inbox starts with agent audio off on purpose: it spends AI producing text the agent just
spoke. An inbox that **already existed** keeps transcribing agent audio, because that is what it was
already doing — the update does not switch anything off on its own. If you do not need that text, turn
the switch off and the spend stops.

## Use cases

- **Audio-heavy support**: the team reads the message in seconds and replies in text.
- **Searching the history**: transcribed text is searchable, so an old voice note becomes findable by
  what was said in it.
- **A sensitive conversation**: one conversation can stay on **Never transcribe** even with the inbox
  switched on.
- **A controlled pilot**: enable transcription only on the inbox you want to evaluate before rolling it
  out.

## Tips, limits and best practices

- The per-file limit is **25 MB**. Larger audio is not transcribed.
- Supported formats: `flac`, `m4a`, `mp3`, `mp4`, `mpeg`, `mpga`, `oga`, `ogg`, `wav` and `webm`.
  Older formats such as `amr` and `3gp` are not transcribed — the bubble says so instead of staying
  empty.
- The language is detected automatically. When the conversation already has a known language, it is
  used as a hint to improve the result.
- The transcription appears on its own, with no page reload.
- Transcribed text can also be translated — see the automatic translation article at the end.

## Troubleshooting

**The audio bubble shows no text at all.**
Check, in this order: the inbox has **Audio transcription** on; the scope matching the direction of
that audio is on; the conversation is not set to **Never transcribe**; and the badge at the top of the
tab reads **Key configured**.

**The bubble shows a notice instead of the text.**
The notice says exactly what happened:

| Notice | What to do |
|---|---|
| This audio format cannot be transcribed | The audio arrived in a container the providers do not accept. Ask for the audio again or convert the file. |
| Needs an OpenAI or Groq key | Configure the integration and use **Try again**. |
| The provider did not answer | A temporary failure. Use **Try again**. |
| Audio above the 25 MB limit | Not transcribable. Ask for a shorter recording. |
| Empty audio file | The file arrived with no content. Ask for it to be sent again. |

**The "Try again" button is not there.**
It only appears for failures a second attempt could actually fix. An unsupported format or an
oversized file would fail again in exactly the same way.

**We are paying for AI on audio we do not need.**
Switch **Transcribe audio recorded by agents** off on the inbox. It is the most common source of spend
that produces no value.

## See also

- [Inbox settings](/hc/ajuda/articles/inboxes-channels-configuracoes-de-inbox-en)
- [Real-time automatic translation](/hc/ajuda/articles/conversations-traducao-automatica-em-tempo-real-en)