Custom brand voice: where standard synthesis breaks down
Standard speech synthesis for a brand stumbles where the text stops being neutral.
Homographs like "castle" or "worth" are read by the model according to a frequency rule and stress is placed at random, while a product name or founder's surname is spelled out letter by letter.
In a short example this goes unnoticed, but in a stream of videos it is heard immediately and damages perception.
The second place is prosody. Ready-made voices are trained on flat reads, so long sentences fall apart: a pause lands in the middle of a semantic block, a list is read as a single stream, and a question loses its rising contour.
The intonation comes out choppy, and the listener reads it as a "robot," even when phonetics are clean.
The third problem is timbre divergence between materials. Each request to the model produces its own variability, so the brand's voiceover in an advertising video, in an app, and in an auto-attendant sounds like three different people.
A unified voice identity for the company does not come together, and recognition rests only on the logo and music.
Add a live voice talent to this — and editing one word turns into a new studio session: a recording queue, re-reading a paragraph, mixing. Updating a promotion or a section name in a voice menu can take weeks, even though only a line of text changes.
It is precisely at these points that a custom voice is needed: it is built on the company's voice identity — agreed-upon stress, tempo, pauses, and timbre — while phrases are updated as text, without a studio. Next, we break down how such a voice is assembled and how synthesis quality is checked.
What implementation delivers: release speed and sound quality
The voice identity is assembled once — several hours of clean voice talent speech, precise markup, and a pronunciation dictionary.
After that, any line is produced from text, so the speed of audio release is determined by the synthesis queue, not studio availability and the voice talent's schedule.
The reduction in voiceover time is visible on the very first batch: a video, podcast, or set of prompts for a voice menu is assembled in a single pass — from script to finished track.
A unified timbre holds across all channels: video, social media, inbound line, training materials. The effect of "the voice was different in the last episode" does not arise due to fatigue, a different studio, or a microphone change.
Pronunciation stability is ensured by the dictionary: names, abbreviations, numbers, and industry terms are fixed in advance; stress does not drift from take to take.
Edits without re-recording change the cycle itself. Replacing a number, legal wording, or phrase order is a text edit and re-synthesis, not a new studio shift.
The share of lines that must be fully re-voiced falls, and iteration takes hours instead of days — a version is approved faster and without losing timbre.
Output sound quality is determined by three things: reference cleanliness, markup, and post-processing. Mastering levels loudness and dynamics to the requirements of a specific platform, so the track sounds the same in headphones, on a speaker, and on air.
For commercial scenarios, we work with platforms like ElevenLabs, and for closed environments, with on-premises deployment where the voice identity does not leave the client's infrastructure.
Ways to obtain a voice identity: comparison of approaches
A custom voice is assembled in one of four ways, and the choice is dictated not by preference but by what source materials are available as input.
That is why before assessment we always request the available audio material: both the timeline and how stably synthesis will hold timbre on long lines depend on it.
Options and their limitations
| Option | When it fits | Limitations |
|---|---|---|
| Recording a voice talent for the project | The voice is not yet fixed; control over manner, tempo, and emotions is needed | Requires a voice talent and studio sessions; source requirements are strict — a single signal chain, clean unprocessed signal |
| Training on archival voice talent recordings | The brand already has reads: advertising, video voiceovers, training materials | The archive is often heterogeneous in signal chain and manner, and training on recordings with music and overlays produces a "drifting" timbre |
| Adapting a ready-made model | A fast prototype or internal tool is needed without being tied to a recognizable voice | The timbre remains someone else's — brand recognition does not emerge, and the license limits commercial use |
| Hybrid approach | There is a short high-quality reference and a tight timeline | Uses a ready-made model as a base, fine-tuned on the reference; source requirements are an even tone and absence of extraneous noise throughout the fragment |
The question then is no longer which table row to choose, but whether the material can withstand training. We test this on a pilot fragment before launching the full cycle.
Work stages: from recording audit to handover
Speech synthesis implementation begins not with model training but with an audit of what has already been recorded.
We analyze the entire corpus: how many clean hours it contains without noise and overlays, how even the timbre is from session to session, and which sound combinations and intonations are absent entirely.
The output is a voiceover work plan: what we re-record and in what volume so that new material does not stand out from the old.
- Recording audit. We analyze the corpus by session: clean time, timbre evenness, share of retakes, coverage of rare sounds. The result is a list of gaps and the volume of additional recording.
- Voice scheme. We fix the required registers: neutral narration, conversational, announcement, emotional peak. Separately, the pronunciation dictionary: names, abbreviations, numbers.
- Assembly. Training on a verified corpus; we keep some recordings in reserve and do not show them to the system — using them, we check whether the voice has become tailored to specific phrases.
- Listening test. Blind test: we mix synthesis and live recordings, and the client and our engineers mark where they hear a seam. We look not at the overall impression but at the share of correct guesses by phrase type.
- Sound acceptance. Control set: numbers, dates, names, abbreviations, long sentences, questions with rising intonation. We trace each error to its cause — dictionary, stress, or lack of data.
- Solution handover. We hand over access, the pronunciation dictionary, the update procedure, and instructions on how to add a phrase without us. After that — support and fine-tuning on new recordings.
Deliverables: model, pronunciation dictionary, procedures
A custom voice is not a set of audio files but a reproducible process: without a dictionary and procedure, every new phrase is voiced "by eye," and the sound drifts apart.
We structure the deliverables so that your in-house employee can produce voiceovers without contacting us for every paragraph.
- Configured voice model and reference voice talent recordings — we preserve the source files so the voice can be fine-tuned or expanded to new intonations without re-recording.
- Pronunciation dictionary: stress, abbreviations, product and trademark names, numerals, units of measurement, Latin script reading rules — everything synthesis usually stumbles on.
- Voiceover procedure: how to prepare text — pause markup, tempo, emotional markers, line length limits, folder structure for bulk generation.
- Usage rights: written channels (advertising, telephony, video, training), territory, and term — disputes when changing contractors are eliminated in advance.
- Access policies: roles and keys, generation limits, request log, storage rules — the model cannot be exported outside, and incidents are investigated via the log.
- Voiceover documentation: launching synthesis, adding words to the dictionary, updating the model, and rolling back to the previous sound reference.
- Control phrase set with automated checking — after each update, you run the list and see whether pronunciation has shifted.
We consider the deliverables handed over when your team runs the control text itself and gets the expected sound: acceptance by result, not by the fact that files were sent.
Case study: moving brand voiceover to its own voice identity
Before
A media brand voiced broadcast and advertising materials using a pool of three voice talents — with different schedules and different delivery styles.
A single video went through four to five iterations, every release waited for a studio recording slot, and about a third of delivered materials came back for revision.
What we did
We assembled a custom voice from the brand's studio archive — several hours of clean recordings with varying intonation and tempo.
Then the model was trained on marked-up fragments: noise was removed, pauses and stress were marked up, and an exceptions dictionary for names and terms was created separately.
Generation runs through ElevenLabs, while the dictionary and pronunciation rules remain on our side — they can be edited without retraining the model.
What changed
Voiceover turnaround shrank from three days to two to three hours: text goes to generation immediately after proofreading, and there is no longer any need to wait for a studio slot.
The number of takes for the same text decreased manyfold — previously the best of eight to ten reads was chosen, now the first or second generation is enough.
Revisions dropped from a third of materials to 5%, and the share of pronunciation defects in delivery is about 2%. The volume of manual editing decreased from half the text to 5–10%: we correct stress in new terms, while the model holds the rest stably.
Moving voiceover to the model did not break recognition — in a blind test, listeners could not distinguish new material from archival recordings.
How does a custom voice differ from standard synthesis?
Standard synthesis is a library of ready-made voices with a fixed set of styles: you choose the voice talent, speed, and emotional shade, but not the timbre itself.
A custom voice is trained on recordings of a specific person, and the model reproduces their articulation, timbre, and speech manner while remaining controllable at the text level.
The difference between approaches is most noticeable in intonation control.
In a custom voice, markup at the phrase and word level is available: stress, pauses, tempo, rising or falling tone in a question, and separate coloring for an advertising address and a technical instruction.
Standard synthesis responds to such edits with smoothing — the style remains averaged, and brand recognition becomes blurred.
Complex constructions — numbers, abbreviations, product names, inflections — are read according to general rules in standard synthesis, and it is almost impossible to influence them.
In a custom voice, we configure the pronunciation dictionary and text normalization rules: the model pronounces terms the way they are accepted inside the company, without re-voicing the entire material.
Editing without a new voice talent is the key difference in operation. With the standard approach, changing the wording means a new studio recording with mixing, and the timbre may differ between sessions.
A custom voice allows you to rewrite a paragraph as text and get the same timbre: the edit cycle shrinks from days to hours, and voiceover versions remain compatible with each other.
We check naturalness of sound through blind listening — we compare custom and standard synthesis on the same phrases and calculate the preference share. Separately, we record the number of iterations to approval and the time per edit cycle.
The quality benchmark is ElevenLabs, but the result of a custom voice is determined by the volume and cleanliness of the voice talent's recordings.
Common concerns: rights, data, and complex texts
Concerns at this stage are almost always the same: whether the source recordings are sufficient, who owns the voice identity, and what will happen to terms and abbreviations in live text.
We address them before signing the contract — this way the project boundaries are clear to both parties in advance.
How many voice talent recordings are needed?
For a stable model, several hours of clean studio speech are enough — without music, echo, or extraneous voices. If there is less material, we collect it in short sessions according to an agreed script, and we align intonation and tempo during preparation.
Who owns the rights to the voice identity?
The rights to the trained model and its use are assigned to the client in a separate contract clause. The voice talent signs consent with a specified scope: where the voice sounds, in which channels, and for what term.
What happens to the voice talent's data?
Recordings and markup are stored in an isolated environment with role-based access, are not transferred outside, and do not enter shared datasets. Upon project completion, they can be deleted, leaving only the model itself.
How does the voice read terms and abbreviations?
For each project, we build a pronunciation dictionary: product names, Latin script, abbreviations, numbers, and units of measurement. Before handover, we run a control list of complex strings — discrepancies are fixed with settings, not by re-recording the entire model.
What remains after handover, and what about support?
Configurations, the pronunciation dictionary, synthesis scripts, and launch instructions. Fine-tuning on new recordings and editing individual words are carried out as part of support — without rebuilding the model from scratch.
Let's discuss the task and come back with a plan and estimate
To make the estimate substantive and the plan feasible, for the first conversation it is enough to describe the task in words and attach what you already have at hand. A full set of materials is not needed: we will collect some ourselves and clarify some along the way.
Below is what is worth preparing so that we come back not with generalities but with a work scheme and timelines.
- Base voice recordings: 2–3 fragments of clean speech, 10–15 minutes each, without music and noise, with varying intonation — question, statement, emotional delivery.
- Text corpus: scripts of future lines, terms, brand names and abbreviations, with notes on exactly how they should sound.
- Application scenarios: where the voice will sound — videos, auto-attendant, training videos, interface voiceover; line length and required emotion matter.
- Voice rights: voice talent consent or confirmation of rights to archival recordings — without this, we do not launch the project.
- Languages and accents: current ones and those planned later; this affects the scope of work and the order of stages.
- Platforms and output format: where the sound is integrated and in what form your contractor or department will receive it.
In response, we provide a work scheme with stages and checkpoints, the composition of source materials, the verification procedure — how we compare synthesis with live speech and by what criteria we accept the result — and timelines with a buffer for edits. Let's discuss the task: send what you have, and we will collect what is missing together.