The short answer
Keep the approved written copy, then make a separate spoken version that resolves ambiguous dates, numbers, names and abbreviations. Test the hardest passage with the voice and model you plan to use before generating the full narration. ElevenLabs offers normalization and pronunciation controls, but their availability depends on the workflow and model. [1][2] A clear script gives you specific things to check; it does not guarantee a correct recording.
Keep display copy and spoken copy separate
A product screen may need a compact label such as '3.5 GB', while the narration needs an agreed reading. Save the original copy as your reference, then create a spoken version with a matching version number. Avoid changing a price, specification or instruction merely to make a sentence sound smoother. If the intended meaning is uncertain, ask the person responsible for the copy before generating audio.
Add a short delivery brief: audience, language, intended accent, tone and where the audio will sit in the video. Name the person who can approve unusual words. For a software tutorial, list the interface labels the listener must recognize. For a product demonstration, separate the exact claim from optional descriptive language so a later rewrite does not silently change the message.
Resolve numbers and abbreviations before generation
ElevenLabs recommends writing ambiguous numbers, dates, symbols and acronyms in the way they should be spoken. Its web Text to Speech interface applies normalization, while other workflows expose their own controls. [1] An automatic conversion cannot know whether a particular number is a price, a version, a telephone extension or a model name unless the surrounding text makes that clear.
Illustrative written copy: 'Open v2.4, select 3.5 GB, then press Ctrl + S before 04/05.' A possible spoken draft is: 'Open version two point four. Select three point five gigabytes, then press Control S before the fifth of April.' That date is valid only if the author intended month/day order. If the intended date is the fourth of May, resolve it explicitly; do not let the generator choose. This example is invented editorial text, not an ElevenLabs output or test.
- Mark whether an abbreviation should be expanded, read as a word or spelled letter by letter.
- Write the currency name when the symbol could refer to more than one currency.
- Keep a list of model names and version numbers that must remain exact.
Make a small pronunciation brief
Use three fields for each difficult term: its approved spelling, the intended spoken form and who confirmed it. Start with names, place names and technical terms repeated throughout the recording. A short voice reference from an authorized reviewer can be more useful to your human approval process than an unexplained phonetic spelling. Keep it with the brief, and do not upload somebody's recording merely because it was supplied for reference.
ElevenLabs pronunciation dictionaries can use substitutions or phoneme rules, but support differs by model. Its API dictionary guide lists the models that support phoneme rules and describes aliases for other models. Check the current list for your selected model rather than assuming all models support the same rule type. [2] That API capability is not a promise that every Studio control has the same support. Check the documentation for the interface you actually use and test the difficult word in a complete sentence.
Match delivery directions to the selected model
Keep production notes outside the spoken text unless the selected tool explicitly supports them. A note such as 'pause while the menu opens' may belong in the editor's timeline rather than in the text sent for narration. Describe the intended delivery in your brief first, then choose the supported control that gets closest to it.
ElevenLabs' current prompting guide says Eleven v3 does not support SSML break tags; it uses audio tags, punctuation and text structure for pacing. The guide also explains that the selected voice influences how directions work. [3] Do not paste an older pause-tag recipe into a different model and assume it has the same effect. Use one direction at a time so you can identify what changed.
Review one difficult passage, then finish the script
Choose a short passage that includes a difficult name, a number and a transition. Record the script version, voice, model and relevant settings. Listen while following the written copy, then listen with the video. Check omitted words, incorrect readings and whether the speaker finishes before the next action. Save the accepted file with its version instead of replacing every attempt under the same filename.
If the passage fails, change one clear issue: rewrite an ambiguous number, adjust one pronunciation rule or simplify an overloaded sentence. If plain spoken text works, you may not need an elaborate dictionary. If you already have acceptable narration, a new generation is not automatically an improvement. Pick With Context has not generated or measured the examples here; this is a preparation and review workflow. Commercial-use permissions are covered separately in the related guide.
Before you decide
- Keep matching versions of the approved copy and spoken script.
- Resolve dates, units, currencies and abbreviations explicitly.
- Confirm repeated names and difficult technical terms.
- Check pronunciation and pause controls for the actual model and interface.
- Approve a difficult passage before producing the full recording.
Sources & limits
This is a documentation-based guide, not a hands-on product test. Provider terms and exact configurations can change. Confirm the details that apply to your purchase.