What is ElevenLabs and what is it good for at work?
ElevenLabs makes AI audio. Its homepage lists Text to Speech across 90 or more languages, voice cloning and voice design from a prompt, a library of thousands of voices, dubbing that translates audio while keeping the speaker's delivery, Speech to Text with speaker diarisation, sound effects, music and ElevenAgents for conversational voice and chat agents.
At work it suits narration for training videos and demos, placeholder voiceovers and translated versions of existing videos. It is not a replacement for a real person where trust depends on hearing that person.
How do I get started with ElevenLabs?
Create an account, open Text to Speech, choose a voice from the library, paste your script and generate. You can download the audio once you are happy. ElevenLabs' documentation describes voice settings including stability, similarity and style, and a speed control from 0.7 to 1.2.
Check the plan before you publish anything. ElevenLabs' help centre says the free plan does not include a commercial licence and that content published from it must credit ElevenLabs in the title, while paid plans carry commercial rights where you hold the necessary rights in your content.
A 10-minute ElevenLabs starter tutorial
The example is fictional. Mei is the learning lead at Kestrel Payroll, a fictional 60-person software company. She needs a 90-second narration for a new-starter video explaining how to submit expenses.
Minutes 0 to 3: write for the ear
She rewrites her written guide as speech: short sentences, one idea each, numbers spelled the way they should be said, and the product name spelled out phonetically in a note.
Minutes 3 to 4: choose a library voice
She picks a calm voice from the library rather than cloning a colleague, which avoids consent questions entirely.
Minutes 4 to 6: control pacing
On the newer models ElevenLabs describes, she writes [pause] between steps instead of SSML break tags, and keeps them sparing because the docs warn that too many can cause odd speed changes.
Minutes 6 to 8: fix pronunciation
The company name comes out wrong, so she uses the pronunciation option ElevenLabs documents for her model, IPA on the newest model or a pronunciation dictionary for terms that repeat across scripts.
Minutes 8 to 10: generate twice and listen
The docs say output is non-deterministic, so she generates twice, picks the better take, and listens to the full clip against the script for skipped words and wrong numbers.
How do I get better results from ElevenLabs?
Most quality problems come from the script and settings rather than the voice.
Start from the suggested settings
ElevenLabs' best practices describe stability around 50, similarity around 75 and style at 0 as a common starting point. Lower stability adds range but can become unstable; higher can sound flat.
Use delivery tags carefully
On newer models, tags such as [pause], [whispers] or [excited] shape delivery. One or two per paragraph is plenty for work narration.
Build a pronunciation dictionary
For product names, client names and acronyms that recur, a dictionary keeps every video consistent.
Match speed to the content
Slow slightly for step-by-step instructions and leave normal speed for summaries, within the 0.7 to 1.2 range.
What should I not put into ElevenLabs, and what do its consent and rights terms say?
Never clone or imitate a colleague's, customer's or public figure's voice without explicit consent, and do not paste confidential scripts or personal data your employer has not cleared. Follow your company's policy on synthetic media and disclosure; ElevenLabs' rules are the floor.
ElevenLabs' Prohibited Use Policy bans cloning or imitating a living person's voice without explicit, informed consent, using recordings of anyone under 18 to create a voice, impersonating organisations, implying false endorsement and removing watermarks or provenance signals. It also requires you to disclose AI interaction where the law requires it or where not doing so would mislead. Its documentation says a Professional Voice Clone can only be made of your own voice and needs verification, even if someone else consents; Instant Voice Cloning asks you to confirm you have the right and consent.
On data, ElevenLabs' privacy policy says it uses personal information to train its models and that you can opt out under Data use in Terms and Privacy settings. Its help centre says it does not train on Enterprise customers' data by default, and its Zero Retention Mode for selected Enterprise customers covers API requests only, not the web interface.
Where does ElevenLabs fall short?
Synthetic voices still misread numbers, names and abbreviations, and the same script can sound different on each generation. Sensitive messages, such as news about redundancies, should be delivered by a person.
Model features differ: pause tags, SSML support and pronunciation methods vary between ElevenLabs models, so a tip from an older tutorial may not apply to the model you are using. Dubbed translations need a fluent reviewer.
How does Tosheo.ai teach ElevenLabs inside a role path?
Tosheo.ai includes AI voice in learning and development, marketing and customer education paths, usually after lessons on scripting and presentations.
Each 10 to 15 minute lesson has a short explanation, a worked example from the learner's role and a hands-on task marked against a visible rubric with specific feedback. A final practical assessment at the end of a path leads to a recognised certificate. On the Team plan, managers assign lessons and see the team report, never raw chat logs. Every plan starts with a 7-day free trial and no credit card is needed.
Sources and scope
ElevenLabs, official homepage (products and features), checked 19 October 2026 ↗
ElevenLabs, Prohibited Use Policy, checked 19 October 2026 ↗
ElevenLabs Docs, Voice Cloning overview, checked 19 October 2026 ↗
ElevenLabs Docs, Text to Speech best practices, checked 19 October 2026 ↗
ElevenLabs, Privacy Policy, checked 19 October 2026 ↗
ElevenLabs Docs, Zero Retention Mode (Enterprise), checked 19 October 2026 ↗