Build an AI voiceover workflow with review

Text-to-speech can voice a product video or a training module in minutes. The hard part is catching what the voice gets wrong. A synthetic read rarely sounds broken. It sounds confident while it mispronounces a client's name, reads a year as a quantity or rushes the sentence that matters.

This tutorial sets out a workflow that puts human review where those errors are caught. It complements our explainer on how AI video generators work, since narration often holds a generated video together. AI Chronicle has not built or run this workflow in any product, and no audio or results are claimed. Where we mention a tool, we report what its official documentation says, as read on 16 October 2026. The worked example is a labelled hypothetical.

Scope box

This workflow covers: short and mid-length narration (explainers, training, product videos, podcast segments) read by a stock synthetic voice or a voice cloned with documented consent.

AI may: convert approved script text to speech; regenerate segments a reviewer flags; suggest spellings or phonetic respellings for the dictionary, which a person checks.

People must: obtain and file voice consent; approve the script and dictionary; do the full listen-through; sign off loudness and disclosure; decide what is released.

Out of scope: live or conversational voice agents, impersonating real people, dubbing someone else's performance, and anything where the listener could reasonably believe a specific real person spoke the words.

The workflow at a glance

The figure lays the workflow out like a multitrack session: four tracks running left to right, crossed by four sign-off gates. Nothing moves past a gate until a named person has checked it.

AI Chronicle

Step 1: Write the script for the ear

A script written for reading often fails when it is spoken. Before any audio exists, rewrite it for listening.

• Spell out what should be said. Write numbers, dates, currencies and units as they should sound: "twenty twenty-six", "four point five percent". Don't leave a model to guess whether "1/4" is a date or a fraction.
• Shorten sentences that carry facts. A listener cannot re-read. One figure per sentence is a sound default.
• Mark segment breaks. Split the script into paragraph-sized chunks, each with its own ID.
• Lock it. Give the approved script a version number. Later edits create a new version, never a silent change.

The approver should be someone who knows how the product, place and people names are actually said.

Step 2: Build a pronunciation dictionary

Names, brands and acronyms are where synthetic voices most often stumble, and listeners notice those errors first. Keep a pronunciation dictionary for the project and reuse it.

The W3C's Pronunciation Lexicon Specification (PLS) is a long-standing open format for this. It defines "the syntax for specifying pronunciation lexicons" for speech recognition and synthesis engines.[1] ElevenLabs' documentation describes accepting PLS files with two kinds of rule: a phoneme rule that spells the sound in IPA or CMU notation, and an alias rule that swaps in a respelling.[2] Its guide adds two practical cautions. The files are case-sensitive, so its example lists "tomato" and "Tomato" separately. And only some of its models honour phoneme rules: "Other models skip dictionary phoneme tags and use the default pronunciation."[2] Our ElevenLabs review covers the platform's other documented voice and editing controls.

Whatever the tool, assume a rule can be silently ignored, and test each entry with a short line containing the word. An alias ("Ngozi" → "n-GO-zee") is cruder than a phoneme but more portable across engines.

Version the dictionary too. The ElevenLabs example passes both a dictionary ID and a version ID at generation time,[2] a habit worth copying: record which dictionary version produced which audio.

Step 3: Choose the voice and settle the rights

Voice choice is two decisions: what the voice sounds like, and whether you have the right to use it. Gate A is about the second.

A stock voice from a vendor's library is the simplest route. OpenAI's text-to-speech guide, for example, describes its built-in voices as "currently optimized for English".[3] Choose for intelligibility, not demo appeal, and keep the same voice across a series.

A cloned voice raises consent questions that the platforms themselves now build into the process. OpenAI's custom-voices guide requires a consent recording in which the voice actor reads a set phrase, and says "The voice must match the consent recording." The feature is limited to eligible customers.[4] ElevenLabs' documentation goes further for its Professional Voice Clones, which require a verification step:

“Even with their consent, you cannot clone someone else’s voice.”

— ElevenLabs documentation, Professional Voice Cloning FAQ.[5]

Under that model, a voice owner who wants to share their voice creates and verifies the clone in their own account, then shares it.[5]

The usage policies point the same way. ElevenLabs prohibits cloning a living person's voice "without their explicit and informed consent or other valid authorization".[6] OpenAI's usage policies bar use of someone's likeness, including their voice, "without their consent in ways that could confuse authenticity".[7] ElevenLabs' ElevenCreative terms also require users who submit an identifiable person's voice to keep records demonstrating consent for at least five years, or longer where the law requires.[8]

File a consent record for every cloned voice before Gate A opens. We suggest it names the speaker, the scope of use, the duration, how consent can be withdrawn and payment terms. That is editorial guidance, not legal advice; publicity-rights law varies by country. Our profile of ElevenLabs and the AI voice business looks at how one company builds these safeguards into its model.

Step 4: Generate in segments

Generate one script chunk at a time. A reviewer's fix then touches one short file, not a ten-minute render, and a regenerated segment can be compared directly with the one it replaces.

For each segment, log the script version, segment ID, voice ID, model, dictionary version, settings and date. Regenerating the same text does not guarantee the same audio, and platforms change their models over time. ElevenLabs makes a related point about clones in its FAQ: "Each clone will sound slightly different, even when you use the same audio."[5]

Ask for a lossless format for the working files where your tool offers one. OpenAI's guide lists MP3 as the default and WAV, FLAC and raw PCM among the alternatives. It notes that its PCM output is "raw samples in 24kHz (16-bit signed, low-endian)".[3] That matters at delivery: Apple Podcasts Connect's minimum for WAV is a 44.1 kHz sampling rate,[9] so plan for sample-rate conversion in mastering rather than discovering the mismatch at upload.

Keeping one narrator stable across segments is the audio twin of keeping characters consistent in AI video.

Step 5: The listen-through review

Gate C is a full, real-time listen by a person following the approved script, on headphones and then once on a phone speaker. Spot checks miss joins, repetition and drifting pace, which only show up end to end. Log each issue with segment ID and timecode, fix it, then re-listen to the stretch around it.

Review against the locked script, not memory: a reviewer who knows the script hears what should be there. And budget the time. A ten-minute piece needs at least ten minutes of listening, plus fixes. Our analysis of measuring AI productivity without hiding review work explains why that time belongs in any estimate of savings.

Step 6: Master to a published loudness spec

Loudness targets differ by destination. All three specifications below measure loudness under ITU-R BS.1770.

LUFS and LKFS are the same unit under different names; the EBU document says so explicitly.[10] For long-form content, ATSC's July 2026 revision of A/85 stresses that "It is vital that the average Dialogue Level be measured and reported as the Loudness for Long-form Content".[11] For narration-led pieces that is usually the voice. Apple's guidance adds an order of operations: normalise first, encode second. "The preconditioning steps need to occur before the encoding process," because compression can clip a signal that breaks the true-peak limit.[9]

A client or platform may have its own delivery spec that overrides these defaults, so ask first.

Step 7: Disclose the synthetic voice

Disclosure can be a contractual requirement, not just good manners. OpenAI's text-to-speech guide says its usage policies require developers to:

“…provide a clear disclosure to end users that the TTS voice they are hearing is AI-generated and not a human voice.”

— OpenAI, Text to speech guide.[3]

ElevenLabs requires disclosure where the law does or where omitting it "would materially mislead" people, and says you "must not remove, alter, or conceal" its watermarks, metadata or other provenance signals to disguise AI content.[6] Decide where the disclosure goes (spoken intro, description, credits) and confirm it at Gate D.

Regulators are watching voice cloning used for deception. The US Federal Trade Commission's rule on impersonation of government and businesses took effect on 1 April 2024.[12] When the rule was finalised, the FTC also proposed extending it to impersonation of individuals. Then-Chair Lina Khan said:

“With voice cloning and other AI-driven scams on the rise, protecting Americans from impersonator fraud is more critical than ever.”

— Lina M. Khan, FTC press release, 15 February 2024.[13]

A legitimate voiceover is not an impersonation scheme, but an unauthorised voice that passes as a real company or person is what these rules target. Disclosure and a consent file are your evidence of good faith. Labels can also travel with the file, as our explainer on content credentials describes.

Step 8: Version everything

Keep one release record per published file: script and dictionary versions, voice ID with its consent record, model and settings per segment, segment files, the signed review log, master loudness readings and the disclosure text. Make versions visible in file names, such as demo_v3_dict2_S04_take2.wav.

That record answers the questions that arrive months later. Which files used a voice whose actor has withdrawn consent? What changed between the approved cut and the live one?

Worked example: a product walkthrough narration

This is a hypothetical example. It describes a plan, not a project AI Chronicle has run.

A small software company is narrating a six-minute walkthrough of a new invoicing feature. The script names the product, two staff members, prices in two currencies and a release date. There will be a video cut and an audio-only podcast cut. A marketing lead owns the project; a support manager is second listener.

The team picks a stock voice, so Gate A records the voice ID and vendor terms rather than a consent file. It builds a small dictionary, splits the script into 18 segments and masters the podcast cut to Apple's target. The reviewer runs these cases. The Result column is empty on purpose: these are cases to run, not results we observed.

Failure modes specific to AI voiceover

• Confident mispronunciation. A name is said wrongly with perfect fluency, and nobody knows the right way.
• Silent rule skipping. The model ignores a dictionary rule type, so everyone assumes the word is fixed.
• Number ambiguity. Dates and decimals are read in a way the writer never intended, especially across regions.
• Segment drift. Regenerated segments differ in tone, pace or level, and the joins show.
• Consent gaps. A cloned voice is used beyond its agreed scope or after withdrawal, because no record links files to voices.
• Spec mismatch. Audio is mastered for one platform and delivered to another, or normalised after encoding.
• Missing disclosure. The label is dropped in the final edit or the audio-only cut.
• Rubber-stamp review. The listen-through shrinks to spot-checks as deadlines tighten.

When not to use this workflow

• You need a specific real person's voice without their documented consent. No review step fixes that.
• The moment calls for a human voice, such as a personal apology or a statement on behalf of a named individual.
• Nobody can do the listen-through. Without it, the review gate doesn't exist.
• The content is high-stakes and specialised, such as medical instructions, and no subject expert can review it.
• The language is poorly served by the engine. OpenAI, for instance, describes its built-in voices as optimised for English.[3] Test with native listeners first.

Questions readers commonly ask

Can we skip the listen-through if we use a pronunciation dictionary?

No. A dictionary fixes known words. The listen-through catches what nobody anticipated: numbers, pacing, artefacts and rules the model ignored.[2]

Is a stock voice free of consent questions?

Largely, because the vendor licenses it. You still need to follow the vendor's terms and disclosure rules,[3][6] and you should avoid using it to suggest a real person is speaking.

Do we have to say the voice is AI?

If you use OpenAI's text-to-speech, its docs say its policies require clear disclosure to end users.[3] ElevenLabs requires it where the law does or where omitting it would materially mislead.[6] Check the terms of the tool you use and the rules where you publish.

What if a voice actor withdraws consent?

Your release records show every file that used the voice. Stop new generations, then follow your contract on existing files.

The next step

Disclosure tells listeners what they are hearing; provenance can make that claim travel with the file. Read our explainer on content credentials and what provenance can show.

About these sources

All pages were accessed on 16 October 2026. Vendor documentation establishes what a product is described as doing or requiring; it does not show how well it works. AI Chronicle has not built, run or tested the workflow. The scope box, figure, review checklist, consent-record list, worked example and test cases are illustrative.

Related reading

• How to keep characters consistent in AI video
• Vapi's approach to voice agents