How to use AI audio in your business, from jingles to voice agents

Chapter 1. How do you make a jingle with Suno?

Suno generates complete tracks, music and sung vocals, from a written description. You need a Suno account. The free plan is enough to try it out; to use the track in advertising you need Pro, from $10 a month.

Write the brief in five points

“A nice song” expresses a taste, but it doesn't say how the track should sound. Before you open Suno, write down five pieces of information.

1. The genre and mood, such as upbeat bossa nova or elegant pop.

2. The tempo: slow, moderate or upbeat, or the beats per minute.

3. The voice, the language and the singing style, such as a bright, persuasive female voice singing in English.

4. The instruments you want to hear, such as nylon-string guitar, brushed drums and warm bass.

5. The structure and the ending, with a short intro, a memorable chorus and a clean final chord.

Step by step

1. In Suno, choose Custom mode.

2. Turn the brief into a prompt and paste it into the style field. Start from one of the two below.

3. Put the words to be sung in the lyrics field, kept apart from the instructions.

4. Generate and listen to the two versions Suno gives you.

5. Adjust the prompt and generate again. If the track is too slow, change only the tempo, for example from 90 to 110 BPM. If the vocals drown out the music, add “vocals only in the chorus”. One change at a time, so you know which one worked.

6. Download the file on a paid plan and save the prompt, the lyrics and your subscription receipt in one folder. If anyone challenges the license, you can show which plan you made it on and when.

This is the opening theme I made for a webinar for Confcommercio

The length you ask for in the prompt isn't guaranteed. If you need a 15-second jingle, trim it in the edit.

The spoken voice is a separate file

In an opening theme, the closing line, for example “Welcome to…”, isn't sung by Suno. It's a separate file made with an AI voice, which you lay over the music in the edit. That way you can change the words without redoing the song. You'll see how to make it in chapter 4.

What about Udio?

Suno and Udio are the two music generators people mention most. On October 29, 2025, Udio settled its lawsuit with Universal Music and announced a new platform with licensed music. In the meantime it switched off track downloads, so the songs stay inside the service.

By April 2026 Udio had also signed deals with Warner, Kobalt and Merlin, and the new platform was expected before the end of the year. For a business that needs to download the file and use it in an ad, Suno is the practical choice today. Before using Udio, check the terms in force that day.

Chapter 2. How do you add sound effects to a video?

An AI-generated video is often silent or has generic audio. If you see a mug being set down and hear nothing, you know right away it's fake. You need an ElevenLabs account, from $6 a month for commercial use, and a video editor like CapCut.

Decide what needs to be heard

Watch the scene and separate two kinds of sound. The foreground event is a specific noise tied to an action. The background ambience is a continuous sound, like the hum of a store. They're two different files, generated separately.

Step by step

1. Write the brief in two lines, with the object and the action in the first and the material and distance in the second. For example “mug set down” and “ceramic on wood, close-up”.

2. Turn it into a prompt and explicitly ask for no voice and no music, otherwise the generator may add them.

3. Paste it into ElevenLabs Sound Effects, not Text to Speech.

4. Listen to the options and pick the one whose length and weight match the action. A mug being set down needs a short tap, not a thud.

5. Download the file. It's a separate audio file. ElevenLabs doesn't touch the video.

6. In the editor, import the file, line up the start of the sound with the moment of the action and adjust the volume. Listen before and after.

You can write the brief in your own language. If the result doesn't convince you, try the same prompt in English.

For the video above, I removed the music and generated these sound effects, which I then edited in afterward.

The ASMR trick

There's a second route. Some video generators create picture and sound together, like Seedance 1.5 Pro, which is a native audio-video model. In that case you can ask for an “ASMR” video, meaning one made only of the small sounds of the action, with no voice and no music. With a reference-to-video model, which starts from a reference image or clip, a short prompt is enough.

Chapter 3. How do you create an original voice for your brand?

With a voice of your own, anyone who watches a video, follows a tutorial or calls your main line always hears the same person. ElevenLabs Voice Design creates it from a written description, without recording or cloning anyone. You need an ElevenLabs account; to use the voice in advertising, a $6-a-month plan.

How do you describe a voice?

“Professional” can mean anything. What works are traits you can recognize by ear. The official ElevenLabs guide lists six.

1. Language and variant, for example General American English.

2. Gender and age range.

3. Recording quality, for example Studio quality.

4. The persona, meaning the role in a few words, like interior design consultant, technician or narrator.

5. The emotion, in two or three adjectives such as warm, curious and reassuring.

6. Timbre and pacing, such as a soft timbre, clear diction and conversational pacing.

Two adjectives that contradict each other, like “energetic and calm”, make the result unpredictable. A short description like “a calm narrator” only works if you want a neutral voice.

The prompt in three lines

ElevenLabs recommends writing the description in three lines. The first says who's speaking: language, gender, age and quality. The second gives the character, with persona and emotion. The third says how it sounds, with timbre, pacing and delivery.

The language always goes in the first sentence. If it's missing, the voice can drift toward another accent.

Step by step

1. Write the brief in two lines, with language, gender and age in the first and persona, emotion and pacing in the second. For example “General American English, adult woman” and “interior design consultant, reassuring, calm”.

2. In ElevenLabs, open Voices, then Create Voice, then Voice Design. There's only one field to fill in, Prompt, and that's where you paste the three-line prompt.

3. Before generating, open Settings. Set the Guidance Scale to around 30%. The higher it is, the more literally the model follows the prompt, but quality can drop. Leave Generate Preview Text on, or turn it off to write the test lines yourself.

4. Click Generate voice. The button shows how many credits it costs, and ElevenLabs gives you three voices.

5. You can't download the voices. Write a real line from your content and listen to it with each one, asking yourself whether every word comes through. Also check that it says your company name correctly, that it doesn't have a foreign accent and that it doesn't talk too fast.

6. If none of them works, change just one thing, a word in the prompt or the guidance, and generate again. If the voice drifts away from General American, raise the guidance.

7. Save the best voice. It takes up one voice slot on your plan, and you'll find it in Text to Speech.

ElevenLabs itself calls Voice Design experimental. Before using a voice in advertising, test it on a one-minute script with numbers, prices and proper names, which are the trickiest things to pronounce. With any tool that creates voices, two rules apply. Always state the language and variant, and judge voices by listening to them, not by reading the description.

For the same course, I made this closing line:

Why design a voice instead of cloning one

Today's speech synthesis models can imitate a person from very little audio. VALL-E, unveiled by Microsoft in 2023, reproduces a voice it has never heard before from a 3-second recording. NaturalSpeech 3 was trained on 200,000 hours of speech.

That's why it pays to start from a voice designed from scratch. If you use a real person's voice without their written consent, that person can ask you to pull the content and pay damages. And customers who recognize the voice feel cheated.

Chapter 4. How do you turn text into an audio file?

Once you've saved the voice, you have two ways to produce speech. Both are in ElevenLabs and both use the voice from chapter 3.

Step by step with Text to Speech

1. Prepare the text. If you're starting from notes or scattered facts, get help from an AI assistant with the prompt below.

2. In Text to Speech, select your saved voice.

3. Paste the message to be spoken, not the instructions for writing it.

4. Generate, then listen for the company name, the pauses and the closing call to action.

5. Download the file. To use it in advertising, it has to be generated on a paid plan.

Voices created with Voice Design v3 also work with the Eleven v3 model, which accepts tags in square brackets like [laughs] or [sighs] for laughter and sighs. One or two per script is plenty. Any more and they sound fake.

When to use Voice Changer

With Voice Changer, you set the pacing, pauses and emotion yourself with your own recording, and ElevenLabs only changes the voice. Use only your own recordings or ones you have permission to use, and listen back to the result, because names and numbers can come out mangled.

Once everything was edited together, this is the result

Chapter 5. How do you build a voice agent that answers the phone?

An audio file only says what you wrote. A voice agent listens to the question, prepares an answer and speaks it, then starts over with each new request. With ElevenLabs Agents it costs 8 cents a minute, plus the language model and the phone line.

The agent's answer depends on the instructions you give it, the information it can look up and the tools connected to it, like a calendar. You prepare them in that order.

What it needs to know

The agent only knows what you give it. You upload it as a knowledge base, a document or a page from your website. Three pieces of information are enough to start.

1. The services you actually offer.

2. Your opening hours.

3. The contact to pass people to when a question goes beyond that, for example the store's phone number.

Check them before you upload them. If an opening time is wrong, the agent will repeat it with total confidence. If a detail is missing, it has to say it doesn't know and give the contact, not make something up.

Saying and doing are different things

With the information, the agent can explain how a visit works. To check availability and actually book an appointment, it needs a connected calendar, with permissions and checks. Writing “book” in the instructions doesn't create the connection. Without a calendar, the agent points people to the right channel and doesn't invent a free slot.

The instructions

In the instructions, put the role, the sources to use, the limits and the handoff to a person. The opening greeting has to make it clear right away that this is an AI assistant.

Three test calls

Before you put it in front of customers, call it three times with test data.

1. An expected question, for example how to set up a visit.

2. A change of request halfway through the call.

3. A question about something you haven't told it.

As you listen, check that it introduces itself as an AI assistant, that it understands the request and that it answers only with the information you gave it. The third call is the most useful. It tells you whether the agent knows its own limits or makes things up.

The first scope

Start with an agent that handles a single request. Write two lines. In the first, what it can handle, for example explaining how to book a visit. In the second, when it has to hand off to a person, for example for a complaint about an order.

Take a typical case. A customer asks whether their delivery will arrive tomorrow, but the agent can't see the warehouse. The right behavior is to say it can't check, give a verified contact and transfer the call only if transfer is actually set up. Promising “it'll arrive tomorrow” is the mistake to avoid.

Chapter 6. How do you turn a website page into an avatar video?

You don't have to write from scratch. Start from text you already have, like your “About us” page, a landing page or a brochure, and you get a video with a virtual presenter. You need a Google account. The free one gives you a few avatar videos a month and only preset presenters, while Workspace Business Starter includes 10.

Step by step

1. Copy the text of the page and decide who the video is for: new customers, an event or social media.

2. In a new Google Slides presentation, open Gemini, the star icon at the top right, and paste the prompt below along with the text.

3. Check names, prices and promises in the slides and the notes.

4. From the File menu, choose Convert to video. Google Vids opens with your slides turned into scenes.

5. In Vids, in the right-hand column, click Avatar, choose the presenter, give it the text to say, for example the slide notes, and generate. Each avatar speaks for at most 60 seconds.

6. Listen for names and pronunciation, then export. State in the video or in the description that the presenter is an AI avatar.

If Gemini isn't in your account, copy the text and notes into Slides by hand and carry on the same way.

Chapter 7. Can you use this content in advertising?

Before you publish, ask yourself three separate questions. Does the service allow me this use? Have I respected other people's rights and data? Is the result protected by a copyright I can enforce? A subscription only answers the first one.

Commercial use

• With Suno, commercial use covers tracks created while you're subscribed to Pro or Premier and downloaded from the platform. Some features, like certain Remixes, stay non-commercial even for paying users.

• With ElevenLabs, files generated during a paid subscription can be used commercially even after you cancel. Files made on the free plan stay non-commercial even if you subscribe later, and Beta features are excluded.

• For video, the model name isn't enough. The license depends on the service you used it through, so write down the service, plan and date.

• Editing in CapCut doesn't change the licenses of your files. If you add music from its library, that music's license applies, and for commercial sounds it's limited to CapCut and TikTok contexts.

• Google doesn't claim ownership of what you generate in Workspace, but it doesn't guarantee exclusivity either. Similar results can be produced for other customers.

Rights

A commercial license isn't the same thing as copyright. Suno, for example, doesn't guarantee that the output is protected by copyright. And a subscription gives you no rights to other people's text, images, voices or trademarks. If you use them without permission, generating them with AI doesn't make them yours.

Transparency and privacy

From August 2, 2026, Article 50 of the AI Act applies, the one on transparency. For a business that uses these tools, that means three obligations.

1. A phone agent has to say right away that it's an AI assistant.

2. Content that realistically imitates real people, places or events, a deepfake, must be disclosed as AI-generated the first time the public comes across it.

3. A creative or satirical ad has to be disclosed too, but a line in the end credits or the description is enough.

Not every synthetic voice or song is a deepfake. An invented voice isn't; a fake statement from a real person is.

The extension to December 2 only applies to the providers of the tools, for machine-readable marking. Anyone who uses the tools to publish, like a business, follows the transparency rules from August 2, 2026.

For a phone agent, privacy matters too, and saying “I'm an AI” isn't enough. Before using it with real customers, define the purpose and the legal basis, prepare the privacy notice, collect only the data you need and decide how long you keep recordings and transcripts. Those are two separate things, and turning off audio recording doesn't delete the transcripts. Also check the data processing agreement with the provider, and use only made-up data while you test.

Chapter 8. How can you tell if audio was generated with AI?

By ear, less and less. That's why providers hide a digital watermark in the audio, a signal you can't hear but software can detect. It's the machine-readable marking the AI Act asks for.

• SynthID from Google DeepMind marks audio created with Lyria and NotebookLM podcasts, and it holds up against noise and MP3 compression. In the Gemini app you can check an audio file of up to 1 hour and 100 MB, but only if it was made with Google tools.

• ElevenLabs has an Audio Detector that looks for its own watermark first and then falls back on the older AI Speech Classifier. That classifier doesn't reliably recognize the Eleven v3 model.

• Meta's AudioSeal is an open-source watermark, presented at ICML 2024, that pinpoints the marked parts of a file. WavMark embeds up to 32 bits of information in one second of audio.

• C2PA, the Content Credentials standard, isn't a watermark but a certificate of origin stored in the metadata. Since version 2.4 of the specification, it also applies to MP3 files.

Spotting audio without a watermark is still hard. The international ASVspoof 5 challenge gathered speech from about 2,000 people and 32 different attack systems to train detectors. So keep the original files with their generation data and disclose your use of AI, instead of counting on nobody noticing.

What should you check before publishing?

1. A list of the pieces in the final file and where each one came from.

2. For each piece: service, plan, date, ID and proof of license or download.

3. The channel you'll use it on, whether website, social media, advertising or a client, with any limits on territory and duration.

4. No one else's voice or identity, no invented testimonials.

5. No library material added without checking its license.

6. If the content imitates real people, places or events, a statement that it was made with AI.

7. For the agent, before real use: privacy notice, legal basis, provider contract and retention periods, all checked.

8. A human review of names, trademarks, recognizable people and the final result.

Where should you start?

You don't need five subscriptions and five projects at once. Pick one chapter and tie it to something you already do by finishing this sentence: “After [usual task], I'll try [prompt] in [tool] to create [content].”

For example, after preparing your next company post, try the voice prompt in Voice Design to create the voice for your series. Or, after shooting your next product video, try the sound effects prompt in Sound Effects for the main action. Pick yours and try it on the next piece of content you need to publish.