REAL VOICE
Record the person
Best when character, emotion, authority, or personal presence matters most.
- Public introductions and personal messages
- Sensitive training or change communication
- Stories where the speaker is part of the meaning
BrainIT Consulting · Free Field Guide No. 6
Keep the Human Voice
A practical AI audio workflow for marketing, editing, training, and small-business operations.
Real voice firstLocal synthetic optionHuman approval before release
What you will leave withA Voice Workflow Card that records the job, authorized speaker, approved words, method, quality checks, disclosure decision, final approver, and human fallback.
Voice is part of the business
A voice does more than carry words. It can make an instruction easier to follow, make a lesson feel less distant, or make a product explanation sound like it belongs to the people behind the business.
That is why the first question is not, “Can an AI copy this voice?” The first question is:
What job should this voice do, for whom, and who remains responsible for the result?
A studio may need a clean temporary read while an edit is taking shape. A marketing team may need alternate versions of an already approved script. A manager may want the same standard operating procedure available as text and audio. These are different jobs, with different consequences.
IDENTITYWhose voice is heard?A recognizable voice can carry reputation, authority, culture, and expectation.
MEANINGWhat do the words do?A pause, emphasis, name, number, or warning can change what a listener understands.
RESPONSIBILITYWho approves the use?A generated file cannot accept accountability. A named person must decide whether it may be used.
Useful distinctionVoice generation is a production step. Representation is a business decision.
Start with the voice job
Choose the destination before the tool. “Make some narration” is too loose to test. “Prepare a review copy of this approved two-minute training script for the internal onboarding page” is a job.
01 AUDIENCEWho will listen?Customers, employees, editors, learners, or the public.
02 DESTINATIONWhere will it appear?A draft timeline, private course, public video, app, or phone system.
03 WORDSWhat is authoritative?One approved script, pronunciation list, and version owner.
04 FINISHWhat is ready?A labeled draft, a review copy, or an approved publication master.
05 APPROVERWho can release it?A named person who can compare the audio with the script and purpose.
Name what is outside the job as clearly as what is inside it. A training narrator does not invent policy. A marketing draft does not publish itself. A synthetic voice does not gain permission for a second campaign because it was approved for the first.
A good job sentenceWhen this approved script arrives, prepare an internal review recording in the authorized voice, record how it was made, and stop for the named approver before the file can move to its destination.
Choose real, synthetic, or hybrid
No single method is best for every job. Use the least complicated method that produces a trustworthy result.
LOCAL SYNTHETIC
Generate a review draft
Useful when repeatability or editing speed matters and the use is authorized and inspected.
- Scratch narration for an editing timeline
- Approved text revised many times
- Accessibility or internal drafts awaiting a human decision
HYBRID
Use each where it helps
Keep the human at the moments that carry identity; use synthesis for controlled supporting work.
- Human opening and closing, synthetic edit drafts
- Human master with temporary replacement lines
- Generated alternatives followed by a final human recording
A practical decision tree
Does the speaker’s presence carry the meaning?Yes → record the person.
Is this only an internal production draft?Yes → a labeled synthetic version may save time.
Will the audience reasonably believe the person said it?Yes → require specific authorization, close review, and a disclosure decision.
Is the script, permission, or destination uncertain?Stop → clarify the job before creating audio.
Record a trustworthy human baseline
A short, clean human recording gives the project an honest reference. It also shows what matters beyond raw similarity: the speaker’s pacing, emphasis, pronunciation, breath, energy, and ease with the words.
Approve the passage.Use the final words for the comparison. Mark names, numbers, pauses, and words that need emphasis.
Prepare the room and microphone.Reduce fans and notifications. Keep the microphone position steady and run a short level check before the real take.
Record to a recoverable local format.OBS is suitable for the BrainIT workflow. Keep the original recording even if audio is extracted later.
Listen before processing.Check that the take is complete, understandable, not clipped, and free from an interruption that changes the meaning.
Record the receipt.Keep the passage version, filename, date, speaker authorization, and a file hash with the project.
Do not train the listener to prefer loudnessWhen comparing versions, match their listening level. Louder often sounds more impressive even when it is not the better reading.
Listen to the BrainIT experiment
BrainIT tested a simple question: can a local synthetic option help with production while the real recording remains the default? The earlier pilot compared identical words without showing the listener which file was real. The real OBS voice was preferred. One locally generated V3 profile was retained as an optional draft or fallback.
This guide uses a fresh passage so the public example is not selected after hearing which text happens to suit a model.
Approved comparison passage
“A useful business voice workflow begins with an approved script and a clear destination. Record the human version first. If a synthetic voice could save editing time, generate the same words, compare them at the same listening level, and let a person decide which version represents the business honestly.”
Blind listening module
Listen before revealing the sources
Both samples use the same approved passage and have been prepared at the same listening level. Listen to Sample A and Sample B before opening the reveal. This guide does not collect a vote or any personal information.
SAMPLE A
Approved comparison passage
SAMPLE B
Approved comparison passage
SAMPLE AEmile’s original OBS recordingRecorded specifically for this guide and preserved unchanged in the private production folder.
SAMPLE BLocal Chatterbox Multilingual V3 generationGenerated from the same approved words with the accepted BrainIT reference profile.
Preparation receipt: Both listening files are mono MP3 copies measured at −18.85 LUFS. The originals, hashes, model settings, and local transcript evidence remain in the project record.
What to listen for
Check correctness, natural phrasing, distracting artifacts, appropriate tone, and whether the file honestly represents the intended use. Similarity alone does not decide which sample belongs in a real workflow.
This is not a personality contestListen for correctness, natural phrasing, distracting artifacts, appropriate tone, and whether the file honestly represents the intended use.
Create a local synthetic version
Local generation can keep the working files on the production computer and avoid per-generation API charges. It does not remove the need for authorization, secure storage, license checks, or human review.
01Authorized referenceUse a clean recording from the speaker and a record of the allowed purpose.
02Approved textGenerate the same version that a person reviewed. Do not let the model improvise policy or claims.
03Recorded settingsSave model, code version, reference hash, text hash, parameters, device, and output hash.
04Labeled draftKeep the result inside the review workflow. Do not make the generation step the publication step.
The BrainIT local toolchain
OBSRecords the human reference and default narration.
Chatterbox V3Creates the optional local synthetic comparison or draft.
Whisper.cppHelps compare the spoken result with the approved script.
FFmpegExtracts, measures, converts, and prepares audio for a fair listening test.
Remotion or an editorPlaces the approved audio into the final production.
These tools are examples from the tested BrainIT workflow, not a claim that every business needs the same stack. Check the exact installed model, code, dependencies, and licenses for the intended use.
Inspect before accepting
A successful generation is only a file. Acceptance is a separate decision.
WORDSDoes it match the approved script?Check omissions, additions, substitutions, numbers, names, and negation.
MEANINGDoes emphasis preserve the intent?A warning, exception, or approval condition must not sound optional.
VOICEDoes it sound appropriate—not merely similar?Listen for pacing, strain, emotion, artifacts, and abrupt changes in room or timbre.
SOUNDIs it technically usable?Check clipping, noise, loudness, silence, edits, and the destination’s format.
HONESTYWould the audience be misled?Decide whether labeling or disclosure is needed for the destination and context.
OWNERWho signs off?Record the named reviewer, decision, date, and fallback if the file is rejected.
- Job
- Approved passage for blind listening module
- Speaker
- Authorized voice owner recorded in project ledger
- Method
- OBS baseline + local V3 alternative
- Checks
- Script · pronunciation · artifacts · listening level · destination
- Decision
- Accept, revise, use real recording, or do not use
Apply the workflow to business work
The same discipline supports several useful jobs. The value comes from fitting voice into an owned process, not from producing more audio.
MARKETINGApproved campaign variationsCreate internal timing drafts or authorized versions while keeping claims, destination, and final release under review.
EDITING STUDIOScratch reads and pickup planningGive editors a useful placeholder, find timing problems, then choose a human pickup or an approved generated line.
HR TRAININGConsistent internal lessonsNarrate approved content with a named policy owner. Regenerate only after the text change is approved.
STANDARD PROCEDURESListen while doing the workPair the authoritative written procedure with audio. The audio must point back to the controlled version.
PRODUCT DEMOSRepeatable draft narrationTest pacing against screen actions before recording or approving the final voice.
ACCESSIBILITYAnother way to receive informationOffer clear audio alongside usable text and captions; do not treat a generated voice as complete accessibility.
For policy, safety, and HR materialThe approved text remains authoritative. Give it an owner and version. If the wording changes, return to approval before generating or publishing new audio.
Integrate without publishing automatically
The safest first integration ends in a review queue, not a public channel. Automation may prepare evidence; a person decides whether the result may move.
01 REQUESTApproved job entersScript, audience, destination, due date, and voice authorization are present.
02 PREPAREAudio draft createdReal, synthetic, or hybrid according to the card.
03 INSPECTChecks and receipt attachedTranscript, audio QC, provenance, disclosure flag, and exceptions.
04 APPROVENamed person decidesAccept, revise, replace with real voice, or reject.
05 RELEASESeparate deliberate actionMove the approved master to its recorded destination.
Automation may
- prepare a labeled draft
- transcribe and compare words
- measure technical properties
- assemble the review receipt
- notify the named reviewer
Automation must not
- invent or approve claims
- reuse a voice for a new purpose
- hide the generation method
- publish because generation succeeded
- place outbound AI calls from this guide
Consent, licenses, disclosure, and security
Voice files are easy to copy and hard to call back. Treat the reference, model output, and publication master as controlled production assets.
Permission is specific
Record who authorized the voice, for what project, purpose, audience, destination, period, and revision process. An old recording is not blanket permission.
Licenses belong to the exact stack
Check the model, code, dependencies, music, fonts, samples, and destination terms. “Free to download” does not answer every use question.
Disclosure follows context and destination
Decide what a reasonable listener could believe, what the content depicts, and what the platform currently requires. Recheck platform rules at release time.
Storage should match the risk
Keep reference audio, generated files, logs, and hashes in an access-controlled project location. Set retention, deletion, and backup expectations before scale.
Watermarks and detectors are layers
They may add evidence, but they do not replace permission, secure handling, review, and clear communication with the audience.
Outbound calling is a different project
Automated voice calls bring additional consent, identity, consumer-protection, telecommunications, and operational requirements. This guide does not teach or authorize them.
Plain-language ruleIf you would be uncomfortable explaining the voice workflow to the speaker and the listener, stop and redesign it.
Complete the Voice Workflow Card
Use the card before generating audio. It turns a vague request into a reviewable production decision.
Voice Workflow Card
One job, one authorized voice, one destination, one approver.
1. Job and destination
2. Voice and permission
3. Method and approved words
4. Inspection and release
Use it alone—or ask for a second pair of ears
You can use this guide and card entirely on your own. If another perspective would help, BrainIT can review the job, compare a small local prototype, or help place an approved voice inside a controlled business workflow.
Sources and limits
This guide draws on current official material from Resemble AI’s Chatterbox model card, OBS Studio, Whisper.cpp, FFmpeg, YouTube’s synthetic-content guidance, and FTC voice-cloning risk research. Product, license, and platform rules must be checked again for the exact stack and destination at publication time.
The Voice Workflow Card and blind comparison are BrainIT teaching devices. This is general educational guidance, not legal advice, a security certification, consent verification, or a guarantee that generated audio will be correct or appropriate.
Emile du Toit · BrainIT Consulting