A small museum wants a friendly synthetic voice to welcome visitors to a temporary nature exhibit. The script includes an inviting introduction and practical instructions about the accessible route and emergency exits. ElevenLabs supports bracketed audio tags in Eleven v4 and v3 to guide performance, such as a curious tone or a whisper. But expressive direction can easily make a critical instruction harder to hear. This hypothetical tutorial uses an original fictional narrator, not a cloned person’s voice, and shows how an editor can apply delivery tags selectively while protecting the precise meaning and intelligibility of safety information.
Lock factual wording before styling the performance
Create an approved script with clearly separated welcome text, exhibit description, directions, and safety notice. For illustration, the venue’s approved note might say, ‘Use the east corridor for step-free access,’ followed by ‘In an emergency, follow the illuminated exit signs.’ Those are fictional place descriptions and must not be copied into a real museum without verification. The writer should preserve exactly which paths are accessible and which warnings are mandatory. If the script is revised, obtain confirmation from the responsible visitor-services owner before regenerating any audio. A convincing voice cannot make an unapproved route safe or correct.
Select an authorized voice and a suitable model
ElevenLabs documents that Eleven v4 and v3 support audio tags and currently recommends Eleven v4 for most expressive speech work. Choose a voice with permission for the intended project; a licensed synthetic or otherwise authorized fictional voice is preferable to an imitation of a recognizable employee or public figure. Confirm any plan and usage limits in your own account rather than assuming a particular export allowance. The sample here describes a neutral museum guide with a relaxed speaking rhythm. Do not imply that a named historical figure personally narrated the exhibit or that a real visitor supplied a voice sample.
Use audio tags only where delivery matters
Write the factual text in normal prose, then add short bracketed directions at the precise points where the speaker’s tone should change. ElevenLabs documents examples such as [curious], [whispers], and [laughs]. A short [curious] cue before an invitation to look closely at a specimen can suit the narrative. Putting [whispers] immediately before an emergency instruction is a poor choice because some listeners will miss it. Avoid loading every line with competing direction. The illustration shows an intentional mistake and a corrected approach rather than an actual spectrogram or measured speech result from the ElevenLabs service.
Plan pauses without assuming exact SSML timing
ElevenLabs states that Eleven v4 and v3 do not support SSML break tags in the same way some earlier models do. For these models, use suitable punctuation, text segmentation, or audio tags to influence pacing, then listen to what is actually generated. A scripted pause after a welcome is different from silence inside ‘Do not enter’. Do not promise a fixed number of milliseconds from punctuation alone. If the signage text must be understood on the first play, prefer simple sentences and explicit location words. A downstream editor can place deliberate silence after approved sentences when the tool’s rhythm is unsuitable.
Compare normal and over-directed readings
Prepare three candidates: a plain reading, a lightly directed version, and a deliberately over-styled draft for internal review. Listen to the same safety sentence at normal volume and through the kind of speaker visitors will use. Ask whether any word disappears behind an expressive reaction or breath. Check whether the tone feels welcoming without becoming theatrical during instructions. Do not report a measured score or a pass rate without a real evaluation. The illustrated ‘warning whispered’ panel represents a quality risk that the editor is meant to detect, not an assertion that the model produced a particular faulty file.
Audit pronunciation of names and directional words
Review exhibit titles, species names, gallery numbers, accessibility terms, and compass directions in the final output. A wrong ‘east’ can matter more than an imperfect emotional nuance. Maintain a script-to-audio comparison: listen while following the approved transcript and mark every changed word or ambiguous syllable. Where the model consistently mispronounces an uncommon term, adjust the written form or speech workflow and document what changed. Do not silently replace a specialized name with a generic phrase that changes the exhibit content. Ask a knowledgeable museum editor to approve contested pronunciations.
Keep the delivery accessible for varied listeners
Not every visitor hears at the same level or processes rapid speech in the same way. Listen using ordinary speakers, low volume, and a simple caption or transcript reference. Safety notices should be at an understandable speaking level with clear pauses between complete instructions. Avoid laughter, whispering, or dramatic vocal effects in evacuation guidance. Provide a written alternative for visitors who cannot hear the audio, and verify that it matches the approved script. If the installation includes multilingual content, review each language independently rather than assuming the same emotional cue makes equal sense in every translation.
Document the final voice approval and its boundaries
Save the exact text, model choice, voice authorization details, tag placement, and output filename in a versioned editorial record. The final reviewer should know whether the audio is AI-generated and who confirmed its accuracy. Do not imply that an unnamed employee recorded or endorsed it. Confirm rights before redistributing files, and avoid feeding private recordings into a cloning workflow without documented permission. Audio tags guide a performance but do not establish identity rights or factual correctness. For this example, no audio was generated, auditioned, or distributed, and no actual museum signage was consulted.
Establish a recheck when routes or models change
A revised accessible route makes earlier audio obsolete even if the voice performance is excellent. Before a seasonal exhibition opens, compare the script with current site instructions and regenerate or rerecord changed segments as needed. ElevenLabs model recommendations, audio-tag behavior, and available output settings can evolve; consult official documentation for the exact model selected. Review all exported files, not just the preview. The release criteria are straightforward: approved facts, intelligible directions, licensed voice, matching transcript, and a named editor who signs off on the final artifact.
How We Research
We reviewed these named official developer resources: ElevenLabs Docs — How do audio tags work with Eleven v3 and v4?; ElevenLabs Docs — Text to Speech best practices; ElevenLabs Docs — How can I add pauses?. We checked the documented feature entry points, eligible platforms, administrator or licensing controls, and known sharing or source limitations against the tutorial steps. Editorial examples were designed as hypothetical failure-mode checks, not reports of hands-on testing. We did not have a live account, source files from a museum or nonprofit, user interviews, access to private records, or measured model outputs. Every illustration is original explanatory artwork rather than a vendor screenshot or measured before/after result. Features and entitlements can change. A responsible editor should confirm current product documentation, applicable organizational policy, and the original evidence before putting the workflow into consequential use.
Frequently Asked Questions
Yes. ElevenLabs documents audio tags in both models and recommends v4 for most new expressive text-to-speech use cases.
ElevenLabs says v4 and v3 do not support SSML break tags. Use appropriate text, punctuation, and audio direction, then audition the result.
No. Keep essential directions clear, prominent and intelligible. Emotional effects should never obscure approved safety language.
No. Voice permission and factual approval are separate reviews. Confirm the route, signage and message with the appropriate museum owner.
No. They influence delivery but are not deterministic. Listen to the final export and compare it word for word with the approved transcript.