Voice brand: synthesized voices of assistants and IVR
Synthesis versus live voice: where the boundary lies
Synthesized speech has, in recent years, approached the human level almost closely. Sber, using the Tacotron 2 + LPCNet architecture, achieved a MOS metric of 4.59 — higher than the benchmark live speech metric by 4.526. The numbers are impressive, but the ear still sometimes catches on to details: synthesized speech lacks natural micro-pauses on inhalation, stress shifts in rare words, intonation stays too stable where a human voice would involuntarily "stutter". A careful listener notices this, even if they can't explain in words what exactly is wrong.
| Criterion | Synthesized voice | Live voice |
|---|---|---|
| Scalability | Instant voice-over of any text, no studio required | Requires a new recording for each script |
| Initial cost | High (model training) | Lower, but grows with text volume |
| Emotional accuracy | Limited to a set of intonation presets | Full range of live emotions |
| Application | IVR, chatbots, dynamic content, assistants | Flagship videos, image videos, one-time campaigns |
A practical rule is simple: if the content is dynamic and voluminous — prices, order statuses, personalized messages — the synthesis is economically justified almost always. If the issue is about a key emotional moment of the brand — launch, apology to the client, flagship advertising — a live voice still wins in terms of depth of impact.
Sound branding and voice brand are not the same
Sound branding is broader: it includes musical jingles, notification sounds, sound logos, and background music in sales points. Voice brand is a narrow but critically important part of this system, specifically responsible for the 'personality' of the company in dialogue with a person.
- Timbre — the basic spectral imprint that determines whether a voice is perceived as "warm," "cold," "authoritative," or "friendly"
- Speech rate — the speed of information delivery, directly influencing the perception of competence or, conversely, restlessness
- Prosody — intonational contours and pause placement, what makes speech alive rather than monotonous reading
- Lexicon and formulation style — the vocabulary an assistant uses, which shapes character just as effectively as timbre
The difference in approaches to work is obvious: a sound logo is written once and almost never changes for decades, whereas a voice brand lives in a constant dialogue and must adapt to thousands of scenarios while remaining recognizable
How to design a voice that will be remembered
Working voice branding is not inspiration, but a sequential process of nine stages, and skipping any of them reduces the chance of recognition
- Brand research — which personality and values should sound in the voice
- Voice character formation — age, character, and energy of the character
- Guidelines — fixing the rules of sound, a brand book for voice
- Casting — finding an actor or narrator who best fits the character
- Scriptwriting — taking into account that text for voice is written differently than for reading
- Prototyping — creating test recordings and draft models
- Recording direction — controlling intonation, tempo, and pauses during the studio stage
- Implementation — integrating the voice into all channels and contact points
- Measuring results — collecting metrics of recognition and satisfaction
A telling example of the scale of investments is the case of Alfa-Bank: approximately 50 hours of recording actor Vsevolod Kuznetsov went into training the Yandex SpeechKit Brand Voice model. This is not a one-time session of a couple of hours, but a full-fledged project at the level of advertising video shooting—only the result is used for years and in thousands of dialogues.
Voice cloning: a legal time-bomb
Technically, voice cloning looks deceivingly simple: just 30–60 seconds of quality audio is needed for the model to extract the timbre (spectral fingerprint), prosody (intonational contours and pauses), and speech rhythm. Then come the questions for which there is currently no unified legal answer across all jurisdictions.
Key risks that need to be covered by a contract before the project starts:
- The right to use the actor's voice after contract termination — does the cloned model belong to the company forever or only for the duration of the agreement
- Consent for further training and modification of the model without the involvement of the original voice carrier
- The risk of using the clone for deepfakes by third parties in case of model leakage
- The obligation to disclose to the user that they are speaking with a synthesized, not a live voice
The last point is especially sensitive: customers who later discover that the 'live' consultant was actually a model usually react negatively — trust in the brand suffers more than if the synthesis had been initially indicated.
Voice synchronization across platforms: where identity is most often lost
Different platforms impose different technical limitations, and it is here that the voice brand usually 'spreads out' — on Alexa one character sounds, in a mobile app another, and in a phone IVR a third, mechanical and faceless.
| Platform | The main limitation | Risk for the Brand |
|---|---|---|
| Alexa / Google Assistant | Ready-made voice engines of the platform, limited customization | The brand character dissolves into the platform's standard voice |
| Phone IVR | Narrow frequency range of the line, outdated equipment | "Robot effect" due to audio compression and loss of prosody |
| Mobile app | More flexible audio codec, room for emotions | A discrepancy in sound compared to the company's phone line |
Avoiding the robot effect is less about technology choice and more about discipline: a unified voice guideline, the same base model or actor reference for all channels, and mandatory testing on real equipment for each platform — not just in studio conditions where everything sounds perfect.
Miller's number and the practice of IVR menus
A separate overlooked topic is the cognitive load on the listener. Miller's study (1956) established that a person retains approximately 7±2 elements in memory. Cohen's modern refinement (2001) showed that without information grouping, the actual limit is closer to 4±1 elements. For voice interfaces without chunking, four elements constitute a safe limit. An IVR menu with seven to eight options is a classic mistake: the user forgets the initial options while the announcer reads the final ones, and in the end, they press "0" to connect with an operator, thereby devaluing the very idea of automation. The rule is simple: a maximum of four options per menu level, with nested structure when necessary, rather than a long list.
Synthesis versus live voice: where the boundary lies
Synthesized speech has, in recent years, approached the human level almost closely. Sber, using the Tacotron 2 + LPCNet architecture, achieved a MOS metric of 4.59 — higher than the benchmark live speech metric by 4.526. The numbers are impressive, but the ear still sometimes catches on to details: synthesized speech lacks natural micro-pauses on inhalation, stress shifts in rare words, intonation stays too stable where a human voice would involuntarily "stutter". A careful listener notices this, even if they can't explain in words what exactly is wrong.
| Criterion | Synthesized voice | Live voice |
|---|---|---|
| Scalability | Instant voice-over of any text, no studio required | Requires a new recording for each script |
| Initial cost | High (model training) | Lower, but grows with text volume |
| Emotional accuracy | Limited to a set of intonation presets | Full range of live emotions |
| Application | IVR, chatbots, dynamic content, assistants | Flagship videos, image videos, one-time campaigns |
A practical rule is simple: if the content is dynamic and voluminous — prices, order statuses, personalized messages — the synthesis is economically justified almost always. If the issue is about a key emotional moment of the brand — launch, apology to the client, flagship advertising — a live voice still wins in terms of depth of impact.
Sound branding and voice brand are not the same
Sound branding is broader: it includes musical jingles, notification sounds, sound logos, and background music in sales points. Voice brand is a narrow but critically important part of this system, specifically responsible for the 'personality' of the company in dialogue with a person.
- Timbre — the basic spectral imprint that determines whether a voice is perceived as "warm," "cold," "authoritative," or "friendly"
- Speech rate — the speed of information delivery, directly influencing the perception of competence or, conversely, restlessness
- Prosody — intonational contours and pause placement, what makes speech alive rather than monotonous reading
- Lexicon and formulation style — the vocabulary an assistant uses, which shapes character just as effectively as timbre
The difference in approaches to work is obvious: a sound logo is written once and almost never changes for decades, whereas a voice brand lives in a constant dialogue and must adapt to thousands of scenarios while remaining recognizable
How to design a voice that will be remembered
Working voice branding is not inspiration, but a sequential process of nine stages, and skipping any of them reduces the chance of recognition
- Brand research — which personality and values should sound in the voice
- Voice character formation — age, character, and energy of the character
- Guidelines — fixing the rules of sound, a brand book for voice
- Casting — finding an actor or narrator who best fits the character
- Scriptwriting — taking into account that text for voice is written differently than for reading
- Prototyping — creating test recordings and draft models
- Recording direction — controlling intonation, tempo, and pauses during the studio stage
- Implementation — integrating the voice into all channels and contact points
- Measuring results — collecting metrics of recognition and satisfaction
A telling example of the scale of investments is the case of Alfa-Bank: approximately 50 hours of recording actor Vsevolod Kuznetsov went into training the Yandex SpeechKit Brand Voice model. This is not a one-time session of a couple of hours, but a full-fledged project at the level of advertising video shooting—only the result is used for years and in thousands of dialogues.
Voice cloning: a legal time-bomb
Technically, voice cloning looks deceivingly simple: just 30–60 seconds of quality audio is needed for the model to extract the timbre (spectral fingerprint), prosody (intonational contours and pauses), and speech rhythm. Then come the questions for which there is currently no unified legal answer across all jurisdictions.
Key risks that need to be covered by a contract before the project starts:
- The right to use the actor's voice after contract termination — does the cloned model belong to the company forever or only for the duration of the agreement
- Consent for further training and modification of the model without the involvement of the original voice carrier
- The risk of using the clone for deepfakes by third parties in case of model leakage
- The obligation to disclose to the user that they are speaking with a synthesized, not a live voice
The last point is especially sensitive: customers who later discover that the 'live' consultant was actually a model usually react negatively — trust in the brand suffers more than if the synthesis had been initially indicated.
Voice synchronization across platforms: where identity is most often lost
Different platforms impose different technical limitations, and it is here that the voice brand usually 'spreads out' — on Alexa one character sounds, in a mobile app another, and in a phone IVR a third, mechanical and faceless.
| Platform | The main limitation | Risk for the Brand |
|---|---|---|
| Alexa / Google Assistant | Ready-made voice engines of the platform, limited customization | The brand character dissolves into the platform's standard voice |
| Phone IVR | Narrow frequency range of the line, outdated equipment | "Robot effect" due to audio compression and loss of prosody |
| Mobile app | More flexible audio codec, room for emotions | A discrepancy in sound compared to the company's phone line |
Avoiding the robot effect is less about technology choice and more about discipline: a unified voice guideline, the same base model or actor reference for all channels, and mandatory testing on real equipment for each platform — not just in studio conditions where everything sounds perfect.
Miller's number and the practice of IVR menus
A separate overlooked topic is the cognitive load on the listener. Miller's study (1956) established that a person retains approximately 7±2 elements in memory. Cohen's modern refinement (2001) showed that without information grouping, the actual limit is closer to 4±1 elements. For voice interfaces without chunking, four elements constitute a safe limit. An IVR menu with seven to eight options is a classic mistake: the user forgets the initial options while the announcer reads the final ones, and in the end, they press "0" to connect with an operator, thereby devaluing the very idea of automation. The rule is simple: a maximum of four options per menu level, with nested structure when necessary, rather than a long list.
Recent articles
MegaFon, BMW, VTB and more 240+ companies. See how it looks in the showcase reel:
Our branding agency is approached to order:
Branding agency services
