---
title: "Voice and audio"
description: "How a voice agent handles a call from start to finish; how to choose and tune its voice; who places outbound calls; where recordings and transcripts are, and what an agent does with WhatsApp audio."
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.getarbol.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Voice and audio

Read this page before you put a voice agent on your clinic's phone, and whenever a call doesn't go the way you expect. At the end you'll find what messaging agents do with the audio, photos and documents that arrive on WhatsApp and on your website's chat.

## Before you start

- To change the voice and the call settings you need a role that can **Create and tune agents**, such as **Agent design**, **Administration** or **Owner**.
- To listen to recordings and read transcripts you need a role that can **See conversations and listen to calls**.
- To make a test call or call someone back you need a role that can **Handle conversations and reply on WhatsApp**, such as **Operations** or **Owner**.

## How it handles a call

With its welcome message.

Waits for the person to speak and lets them interrupt.

Your knowledge base and your systems.

Says what it will do before doing it.

When the person needs nothing else.

Transcript, recording and summary.

### When it answers

- The agent answers and says its **Welcome message**, the greeting under **Call experience › Greeting**. Every call starts with it: that's why a voice agent without a greeting isn't ready.
- It knows the date and time in its time zone. When it offers or confirms an appointment, it says the calendar date as well as the time.
- It applies the [bulletins](/en/raiz/campaigns/) in force when the call starts.
- It serves in a single language, the one you chose under **Language and time zone**: it speaks and understands that language. It doesn't switch languages during the call.

> **Greeting people by name**
>
> When the caller's number is in **People**, Raíz passes their name to the agent. For the agent to use it, write something like `If you know the person's name ({{contact_given_name}}), greet them by name.` in its instructions. For an unknown number that value arrives empty, so don't put it in the **Welcome message**.

### During the conversation

- **Turns.** The agent answers when the person finishes speaking. If the person stays silent, it speaks to them again after a few seconds: 7 by default, or whatever **Wait after a pause** says.
- **Interruptions.** The person can interrupt it: the agent stops talking, listens and answers the new point. In the recording, each interruption is marked.
- **Asking for a moment.** If the person says they're going to look for a document or need to think, the agent waits silently without hanging up.
- **Numbers on the keypad.** The person can type a long number, such as their ID, on the phone's keypad. The agent receives it when they press **#** or stop typing for a few seconds.
- **Lookups.** When it searches your knowledge base or checks your systems, for example availability, it does so right away.
- **Actions that change something.** When it's about to book, cancel or change something in your systems, the agent first finishes saying what it's going to do, for example "I'm booking you for Tuesday the 12th at 9:00", and only then does it. That way the person can correct it in time. Once the action starts, it isn't interrupted until it's done.

### When it ends

- The agent hangs up only after the person has said they need nothing else, and never while there's something they still need to hear, such as the date or place of their appointment.
- The person can hang up whenever they want.
- The call also ends when it reaches the **Maximum duration** (10 minutes by default) or, if you turned on **End after silence**, when the person says nothing for that long.

### Handing the call to a person

A voice agent doesn't transfer the call to your team or to another number. What you write under **Tools and knowledge › Escalation and support** tells it when a case needs your team, and its rules keep it from saying it transferred the call. Write in its instructions what it should offer in that case, for example the clinic's WhatsApp or its walk-in hours, and review those calls in **Conversations**.

## Choose and hear the voice

1. In the menu, click **Workers** and open the voice agent.

2. Under **Call experience**, click **Voice**. The voice it uses now shows first and selected. If that voice is no longer in the catalog, you'll see it as **Current voice**.

3. Type in **Search voices…** or browse the list. Voices recorded in the agent's language come first, then those that also speak it, and the rest at the end.

4. Click the play button next to a voice to hear its sample. It plays in the agent's language when the voice has one. The sample is the original voice: it doesn't reflect your **Voice settings**.

5. Select the voice you prefer.

6. If you like, open **Voice settings** (it appears once a voice is selected) and move the controls:

   | Control | What it changes | Range | If you don't touch it |
   | - | - | - | - |
   | **Stability** | How even the voice sounds. Higher is steadier; lower is more expressive and varied. | 0% to 100% | 50% |
   | **Speed** | How fast it speaks. | 0.7x to 1.2x | 1.0x |
   | **Clarity** | How closely it matches the voice's original sample. | 0% to 100% | 80% |

7. Click **Save**.

Changes take a few moments to reach calls. Make a [test call](/en/raiz/test-an-agent/) to hear it the way your patients will. If you change the agent's language, check that its voice speaks the new language.

## Tune the call's pacing

Under **Call experience › Conversation pacing**:

| Setting | What it does | Values |
| --- | --- | --- |
| **Listening › Important words** | Names and terms it must recognize precisely: your clinic, your locations, your services, your professionals' surnames. | Up to 50 words or phrases, up to 64 characters each. |
| **Wait after a pause** | Seconds of silence before it speaks to the person again. | From 1 to 30. Empty: 7. |
| **End after silence** | Hangs up when the person says nothing for that long (**Hang up after**). | From 10 to 600 seconds. Off by default: it never hangs up on silence. |
| **Maximum duration** | Total limit of each call. | From 1 to 120 minutes. Empty: 10. |
| **Record calls** | Keeps the audio of each call with its transcript. | On by default. |
| **Hide numbers entered on the keypad** | Numbers the person types on the keypad are stored as `<REDACTED>` in the transcript. The agent still receives them during the call. | Off by default. Only affects new calls. |

Under **Call experience › Intelligence** you choose the **Language model** that thinks during the call. The default is **Arbolito Agile**, the quickest: on a call, every second of waiting shows. More capable models can take longer to answer.

For the rest of the agent's settings, see [Agent settings](/en/raiz/agent-settings/).

## Outbound calls

A voice agent also places calls. These are the ways an outbound call starts:

| Who or what starts it | How | What for |
| --- | --- | --- |
| You, from the agent's page | **Test call** (see [Test an agent](/en/raiz/test-an-agent/)) | Hearing the agent on a real call to your phone. |
| You, from a conversation | **Actions › Call back** on a call | Having the agent call that person again. |
| A campaign | From your systems or with the Arbol team; you follow its progress in **Reports › Campaigns** (see [Bulletins, templates and campaigns](/en/raiz/campaigns/)) | Calling a list of people within the allowed hours. |
| Your other systems, such as Agenda | With a call-type **Contact profile** (see [Bulletins, templates and campaigns](/en/raiz/campaigns/)) | For example, reminding someone of an appointment or confirming it. |
| Arbi, the assistant | You ask it and approve the call before it goes out | A one-off call. |

All of them are real calls: the agent's tools really run. No agent calls a blocked number (see [People and the block list](/en/raiz/contacts/)).

**Which number it calls from.** The agent calls from one of its numbers. If it has several, you choose the **Source number** on a test call. If it has no number of its own and your organization has the shared outbound line (in **Settings › Channels** it shows as **Global outbound**, "Available to every voice agent"), the call goes out from that line, and the person sees that line's number, not one of your clinic's. The shared line is only for calling out.

Outbound calls appear in **Conversations** with **Direction › Outbound**. What happens when nobody answers: [When the patient doesn't answer](/en/raiz/no-answer/).

## Recordings and transcripts

Every call is kept in **Conversations**, with the **Voice** channel. An outbound call appears as soon as it starts; an inbound one, when it ends.

On the call's page:

- **The transcript**: every turn, from the person and from the agent. Interrupted turns say **Interruption**, and the tools the agent used appear as cards with their result.
- **The recording**, if the agent records. The bar lets you play, go **Back 10 seconds** or **Forward 10 seconds** and change the **Playback speed**. The amber marks on the bar are interruptions. Click a turn's time (**Play from here**) to hear it, and while it plays, the transcript highlights the turn.
- **The response time** of each agent turn: how long the person waited. Hover over it to see the detail: **Understanding speech**, **Thinking of the answer**, **Generating speech** and **Perceived total**. Turns over 3 seconds show in amber.
- **The summary and the outcome**, which Raíz generates when the call ends and can take a few minutes (see [Follow conversations](/en/raiz/conversations/)).

The recording appears a few moments after the call ends.

### Who can listen to them

Only people in your organization with a role that can **See conversations and listen to calls**. A conversation's PDF report includes the transcript, not the audio. Email notices include the title, the summary and the evaluations, never the transcript.

### Stop recording

1. Open the agent and, under **Call experience**, click **Conversation pacing**.

2. Turn off **Record calls**.

3. Click **Save**.

From the next calls on, the audio isn't kept. The transcript still is, and earlier recordings stay in their conversations.

> **Recordings are patient data**
>
> Listen to them and share them only with whoever needs to know. If your institution must tell callers that the call is recorded, include it in the **Welcome message**.

## Audio, photos and documents on WhatsApp

Messaging agents don't speak: they reply in writing. But they understand what the person sends:

| What the person sends | What the agent does | What your team sees in the conversation |
| --- | --- | --- |
| **Voice note** or audio | Raíz transcribes it and the agent answers what was said. | The audio, to play, with its transcript as text. |
| Photo (JPEG, PNG, WebP or GIF) | Reads it, text included: a medical order, an ID card, a photographed document. | The image, which opens with a click. |
| PDF | Reads it, even a scanned document. If it's password-protected, it can't open it: it asks the person for a photo or the same PDF without a password, never for the password. | The document, which opens with a click. |
| Video or sticker | Doesn't see it: it only knows it arrived. | The file. |
| Location | Receives the place's name or address. | **Shared location**. |
| Contact card | Only knows a card was shared. | **Contact card**. |

Audio is transcribed even when the number is outside business hours, so your team can read it. A file over 25 MB isn't kept, and audio over 10 MB is kept without a transcript.

The agent can also send files that your integrations generate, such as an appointment voucher, when the tool provides them: PDF, JPEG or PNG.

### On your website's chat

Under **Sites and forms › Settings › Conversation** you can turn on **Voice messages** (the visitor records audio, Raíz transcribes it and the agent answers what was said), **Offer speech input** and **Offer spoken replies**. Voice messages need the visitor to allow the microphone in their browser. See [Chat on your website](/en/raiz/web-chat/).

## Troubleshooting

After a few seconds of silence, the agent speaks again: by default, after 7.

**What to do:**

- Raise **Wait after a pause**, for example to 10 or 12 seconds.
- For long numbers, such as an ID, write in its instructions that it should ask the person to type it on the keypad and finish with #.
- If the agent answers before the person finishes a sentence, listen to the recording and send the conversation's link to **Support**: that moment can't be tuned from the Dashboard.

Open the call and look at the response time of the turns in amber.

**What to do:**

- If **Thinking of the answer** takes the longest, set the **Language model** back to **Default (Arbolito Agile)** under **Intelligence**.
- If the wait matches a tool, check on its card how long it took: your system is answering slowly. Tell whoever manages your integrations.
- Check that its instructions ask it to announce a lookup that takes a few seconds. The instructions Raíz writes when you create the agent already say so.

It depends on the voice, its controls and what the agent says.

**What to do:**

- Try other voices under **Voice** and pick one recorded in the agent's language.
- Under **Voice settings**, adjust the **Speed** and try lowering the **Stability** a little for a more expressive voice.
- Check that its instructions ask it to speak in short sentences, without lists or links, and to say numbers and dates naturally.
- Make a test call after each change.

It can be noise, a bad connection or words the agent doesn't know.

**What to do:**

- Listen to the recording and compare it with the transcript to see what it understood.
- Add your clinic's name, your locations, your services and your professionals' surnames to **Important words**.
- Check that the agent's **Language** is the one your callers speak.
- Ask it in its instructions to read back important details before using them and to ask people to type numbers on the keypad.

In **Conversations**, the call's status and duration tell you why it ended.

**What to do:**

- **Timed out**: it hung up because of **End after silence**. Turn it off or give it more time.
- If it lasted exactly the **Maximum duration** (10 minutes if you didn't change it), raise it.
- If the agent said goodbye in the transcript, check whether the person said something that sounded like a goodbye. If it hung up too early, send the conversation's link to **Support**.
- **Errored**: the call didn't connect or the line failed. See [When the patient doesn't answer](/en/raiz/no-answer/).

Check, in this order:

- In **Workers**, the agent is **Active** and doesn't say **Not published**.
- In **Settings › Channels**, the number's row shows the agent's name and the **Active** status. If it says **Needs attention**, use **Check connection** and, if it stays the same, tell the Arbol team.
- Make a test call from the agent's page. If the agent calls you fine but doesn't answer inbound calls, tell the Arbol team through **Support**.

The agent wasn't recording or the recording hasn't arrived yet.

**What to do:**

- Check **Record calls** under **Conversation pacing**. It only applies to calls after the change.
- Wait a few minutes after the call and open it again.
- If the call was very long, its recording may not have been kept: the transcript is there.

On a call, the agent can't transfer.

**What to do:** find those calls in **Conversations**, read the summary and contact the person. For next time, write in the instructions what the agent should offer when someone asks for a person.

The audio may have arrived without a transcript, for example because it was very long, or with a lot of noise.

**What to do:** open the conversation and listen to it. If needed, reply to the person yourself (see [Handle a WhatsApp conversation](/en/raiz/take-over-a-conversation/)).

Source: https://docs.getarbol.com/en/raiz/voice/index.mdx