ElevenLabs for beginners: turn a short script into reviewed narration

The audio player contains the verified result: 29.1 seconds of narration from an approved 61-word Swedish script. The spoken date, time, "AI," and "granskningskvitto" passed full-file decoding and machine readback, so the first take was kept without inventing a correction. The complete file was decoded and saved as elevenlabs-demo-sv-v1.mp3.

This is Hammer's first ElevenLabs guide. It covers prerecorded narration, not voice cloning or an AI agent that responds in real time.

Who this guide is for

Use this workflow when you already have a short approved script for a course introduction, internal explainer, product walkthrough, or tutorial clip. ElevenLabs is not writing the message in this guide. It turns finished text into speech, and the review identifies the exact audio file that was checked.

The verified run below used the Swedish workshop script. The English narration card is a native practice adaptation, not a second verified generation.

What you'll do in 10–20 minutes

• prepare a 60–90-word script for speech
• open the correct Text to Speech view in ElevenLabs
• choose a voice suited to the language and accent
• review names, numbers, pacing, pauses, and the ending
• download the correct take and save a voice receipt

How ElevenLabs Text to Speech works

Text to Speech, or TTS, is software that turns written text into generated spoken audio. The result is a prerecorded file. It does not listen to anyone or answer follow-up questions.

Open ElevenLabs Text to Speech. Sign in or create an account if needed. Find the text input, voice selector, model settings, and the Generate or Generate Speech button. ElevenLabs still uses both Text to Speech and Speech Synthesis in its material, so exact labels and panel positions may change.

Source: ElevenLabs Text to Speech product guide.

Before you generate

Approve the script first. The fictional workshop introduction contains no customer names, student information, passwords, or contract details. Confirm that you have the right to use the script, names, and finished audio.

Commercial use requires a check of the current plan, licence, and whether the selected feature is marked Beta. ElevenLabs says free or signed-out output has no commercial licence. Output created under a paid plan may be used commercially when you hold the necessary rights, follow the terms, and are not using a Beta Service. Swedish and other EEA users should also check the ElevenLabs EEA terms.

Source: Can I publish the content I generate on the platform?

If model-training use matters to your organisation, review the setting under your profile icon, Terms and privacy → Data use. Disable Improve the models for everyone before submitting new material when that matches your policy. This setting does not prove that scripts are never stored or processed, so keep using approved material.

Source: Is my data used to improve ElevenLabs AI models?

Create and review the narration step by step

1. Make the script ready for speech

Write a short narration card before opening the tool:

Narration card

Purpose: [what the listener should understand or do]
Audience: [one concrete listener]
Language/accent: [target]
Length: 60–90 words
Script version: [name + date]
Must be spoken exactly: [names, dates, numbers, acronyms]
Do not change: [facts, commitments, source wording]
Output use: [internal, noncommercial, commercial/public]

Write dates, clock times, and numbers the way they should be spoken. Break long sentences where you want a pause. Expand acronyms when their pronunciation could be ambiguous. Proofread every word because a typo can become an audible error.

A native English practice version of the workshop script could begin: "Welcome to the fictional workshop on AI in everyday work. We meet on Wednesday, August twelfth, at nine in the morning." It keeps the facts but uses English date and time phrasing.

2. Choose one voice and start conservatively

Preview only a few voices that fit the target language. A voice designed for Swedish is a better starting point for Swedish text than a voice that carries a strong accent from another language.

Choose a model that supports the language. Start with the defaults and keep speed near 1.0. Adjust only the control connected to a problem you can actually hear. Not every model exposes the same controls.

Hammer's verified run used Adam medelålders ung man, Eleven Multilingual v2, stability 0.72, similarity 0.62, and style 0. The model supported Swedish and was not labelled Beta in the API metadata used for the run. These values identify this take; they are not a universal recipe.

3. Check the cost label and generate

Read the full script once more. Check the credit amount shown in the current interface, then generate one take. ElevenLabs says credits are deducted when you press Generate, not when you download the file. There is no free preview with your own text.

Source: Is there a way to preview audio without losing quota before downloading?

4. Listen twice

First listen without reading. Then open the script and check:

• names, product names, and specialist terms
• dates, clock times, counts, and durations
• pacing, pauses, and sentence endings
• clarity, audible artefacts, and the final line

An automatic transcript can flag possible mismatches, but it cannot prove pronunciation. In Hammer's run, Whisper returned the complete message and normalised the spoken Swedish date, time, and duration into numerals. That was a reason to replay those moments, not a reason to change correct speech.

5. Correct only what needs correction

If you hear a problem, spell out a date, split a sentence, change punctuation, or choose a voice better suited to the language. Change one thing at a time and compare the new take with the first.

Hammer's take did not require a correction. The first generation passed the full-file decode, transcript, and signal checks, so it was kept. A truthful clean pass is more useful than manufacturing a before-and-after example for the sake of the tutorial.

6. Replay and download the correct file

Replay the chosen take from beginning to end immediately before downloading. Confirm that you have not selected an older History item. Download MP3 or WAV, depending on what the account shows and what your production needs.

ElevenLabs documents downloads immediately after generation or later through Text to Speech History. Hammer saved a mono MP3 at 44.1 kHz for this test: 29.1 seconds and 466,068 bytes.

Source: How do I download generated files from Text to Speech?

7. Save a voice receipt

A voice receipt is a short record of the script, voice, and audio file that were actually reviewed.

Voice receipt

Script: workshop intro, sv-v1
Language: Swedish
Voice: Adam medelålders ung man
Model: Eleven Multilingual v2
Take 1: full file and transcript check
Correction: no correction needed
Final file: elevenlabs-demo-sv-v1.mp3
Technical check: MP3, 44.1 kHz, mono, 29.1 seconds
Final review: names, numbers, pacing, pauses, and ending checked

Common mistakes

• pasting an unfinished script and trying to fix writing and speech at the same time
• choosing a voice that does not fit the target language
• changing the voice, model, speed, and punctuation together
• accepting a natural-sounding take without checking names, dates, and the ending
• downloading an older generation without replaying it

Place the final narration under the real slide, tutorial clip, or course introduction and listen once in context. If the same voice will appear regularly in customer-facing content, continue with Hammer's guide to blind-testing AI voice and brand fit.

Want to turn scripts, narration, and final review into a repeatable workflow? Skill Forge helps your team define the routine. Tool Forge is the next step when the handoff is worth automating. Subscribe for the next practical guide.