Voice assistants and company voice branding
What makes a voice recognizable: five parameters and the logic of their selection
The Mercedes MBUX study on name-brand voice assistants (NBVA) identified a specific set of characteristics that need to be adjusted consciously, not 'by ear': sound quality, frequency range (pitch), prosody — melody and stress distribution in speech — and voice gender. All of them must correspond to the brand's stated traits: if the brand is built on warmth and care, a lower pitch and smooth prosody are needed; if on competence and accuracy — a more neutral tone and clear articulation without unnecessary emotionalism.
Timbre and pitch
Timbre is the physical coloring of the voice, pitch is the emotional coloring of the message. The same timbre can sound friendly or cold depending on how the prosody of a specific phrase is structured. Therefore, a voice brand guide should fix not 'the voice as sound,' but the voice in specific scenarios: greeting, refusal, error, sale.
Tempo, intonation, address
The speech rate sets the feeling of confidence or restlessness — an overly fast assistant is perceived as intrusive. Intonation determines whether the phrase sounds like an expert's statement or a question from an uncertain trainee. The model of address (using 'you' or 'you', the user's name or the depersonalized 'you' — you are executing a command) forms the distance between the brand and the person — a parameter that is often not documented at all, despite being more critical than the tone.
Synthetic voice versus live announcer: when to choose which
Voice cloning technologies in the 2024–2025 years eliminated the technical limitation: ElevenLabs clones a voice from a one-minute audio sample, CosyVoice 3 — from a three-second recording. Neuro.net demonstrated the limits of capabilities by recreating the voice of Vladimir Vysotsky. A technically alive announcer is no longer required for quality sound. The choice question is now not about quality, but about strategy.
| Criterion | Synthetic voice (TTS/cloning) | Live announcer |
|---|---|---|
| Scalability | Instant generation of any text volume, multilingualism | Requires new recordings for each phrase |
| Cost per volume | Low with large volumes | High, grows with content volume |
| Authenticity and emotional connection | Limited if the voice is not tied to a real personality | High, especially if it's the voice of a founder or expert (Barnes, 2018) |
| Legal risks | Rights to cloned voice, consent issues | Contract with a specific person, risk of ambassador leaving |
| Flexibility of adjustment via SSML | Full control over intonation, pauses, tempo through markup | Depends on the actor, it's harder to standardize |
A live voice actor is justified when a brand bets on a specific personality — a founder, expert, or ambassador: this creates a clear emotional association, but requires a systemic approach, not just one lucky clip. Synthesis wins where scale is needed — call centers, IVR, multiple languages, and constantly updated content.
Can an AI voice be human — and why it risks sounding 'like everyone else'
Here lies the main practical divide in the topic. Research by WPP Media, amp, Choreograph, and MediaScience (2025) showed: less than half of listeners could distinguish an AI voice from a human one — 42% on individual sentences and 47% on full audio clips. Yet synthetic voices were no less engaging to the brand, drawing attention and intent to buy compared to human voices. Technically, an AI voice has long been capable of convincingly conveying emotions.
The problem is not with the quality of synthesis, but with unification. If ten brands use the same voice engine with 'default' settings, their voices sound almost identical — this is the very effect that studies by ElevenLabs and WPP Media capture: quality increases, but distinctiveness decreases. Differentiation is no longer about whether the voice sounds natural, but whether it has a unique sonic signature — specific prosody, recognizable pauses, brand-specific verbal turns. Ipsos confirms the value of the question in its study The Power of You (2020): ads with a distinctive sonic identity are 8,5 times more likely to achieve high ad recognition rates than those without it.
How to measure the effectiveness of a voice brand
The model applied to NBVA (name-brand voice assistants) using Mercedes MBUX as an example structures the outcome in three levels, and this chain should form the basis of metrics:
- Drivers — voice uniqueness and quality of dialog scenarios;
- Intermediate outcomes — perception of brand identity through the voice and strength of the emotional connection with the user;
- Final outcomes — multi-dimensional loyalty: willingness to recommend, repeat usage, resilience to switching to a competitor.
In practice, specific metrics are needed for evaluation: voice recognition without visual association to the brand (blind recall test), attractiveness (pleasantness/irritating scale), uniqueness relative to competitors' voices, and correlation with NPS or repeat purchases.
How to conduct A/B testing of voices
- To record the same script and deliver it with two different voice settings (pitch, tempo, prosody) to two comparable audience groups.
- To measure the percentage of correct brand identification without visual cues — only by sound.
- To collect subjective evaluations on scales of 'warmth', 'competence', 'trust' — these should align with the target brand archetype.
- To test behavioral response: listening time until abandonment, conversion to target action after voice message.
- To repeat the test after 3–6 months — voice recognition improves with contact frequency; a single test yields artificially low results.
Voice Guide: how to document the brand voice
Tone of voice guide for text — a familiar practice, but there is almost no systematic approach to creating a similar document for sound. This creates a risk: different contractors configure TTS differently, and the company's voice gradually spreads across incompatible versions. A Voice Guide should close this gap and include at least the following sections:
- Reference audio recordings of the voice in 5–7 basic scenarios (greeting, refusal, apology, sale, error, farewell);
- Specific SSML parameter values — speech rate, volume range, pause duration, points of intonational rise;
- Prohibited voice states — what must never sound (for example, 'too cheerful' when refusing a refund);
- User address rule — fixed form of address using 'you'/'you', use of name;
- Voice connection with the rest of the brand's sound system (sonic identity system) — jingles, sound notifications, musical logo, so that the voice does not exist in isolation from the rest of the brand's sound landscape, as emphasized in the works of Barnes and subsequent sonic branding research;
Without such a document, even the most carefully selected synthetic voice will gradually start to sound more neutral and impersonal — not because the technology degrades, but because nobody monitors compliance with the original parameters at scale;
What makes a voice recognizable: five parameters and the logic of their selection
The Mercedes MBUX study on name-brand voice assistants (NBVA) identified a specific set of characteristics that need to be adjusted consciously, not 'by ear': sound quality, frequency range (pitch), prosody — melody and stress distribution in speech — and voice gender. All of them must correspond to the brand's stated traits: if the brand is built on warmth and care, a lower pitch and smooth prosody are needed; if on competence and accuracy — a more neutral tone and clear articulation without unnecessary emotionalism.
Timbre and pitch
Timbre is the physical coloring of the voice, pitch is the emotional coloring of the message. The same timbre can sound friendly or cold depending on how the prosody of a specific phrase is structured. Therefore, a voice brand guide should fix not 'the voice as sound,' but the voice in specific scenarios: greeting, refusal, error, sale.
Tempo, intonation, address
The speech rate sets the feeling of confidence or restlessness — an overly fast assistant is perceived as intrusive. Intonation determines whether the phrase sounds like an expert's statement or a question from an uncertain trainee. The model of address (using 'you' or 'you', the user's name or the depersonalized 'you' — you are executing a command) forms the distance between the brand and the person — a parameter that is often not documented at all, despite being more critical than the tone.
Synthetic voice versus live announcer: when to choose which
Voice cloning technologies in the 2024–2025 years eliminated the technical limitation: ElevenLabs clones a voice from a one-minute audio sample, CosyVoice 3 — from a three-second recording. Neuro.net demonstrated the limits of capabilities by recreating the voice of Vladimir Vysotsky. A technically alive announcer is no longer required for quality sound. The choice question is now not about quality, but about strategy.
| Criterion | Synthetic voice (TTS/cloning) | Live announcer |
|---|---|---|
| Scalability | Instant generation of any text volume, multilingualism | Requires new recordings for each phrase |
| Cost per volume | Low with large volumes | High, grows with content volume |
| Authenticity and emotional connection | Limited if the voice is not tied to a real personality | High, especially if it's the voice of a founder or expert (Barnes, 2018) |
| Legal risks | Rights to cloned voice, consent issues | Contract with a specific person, risk of ambassador leaving |
| Flexibility of adjustment via SSML | Full control over intonation, pauses, tempo through markup | Depends on the actor, it's harder to standardize |
A live voice actor is justified when a brand bets on a specific personality — a founder, expert, or ambassador: this creates a clear emotional association, but requires a systemic approach, not just one lucky clip. Synthesis wins where scale is needed — call centers, IVR, multiple languages, and constantly updated content.
Can an AI voice be human — and why it risks sounding 'like everyone else'
Here lies the main practical divide in the topic. Research by WPP Media, amp, Choreograph, and MediaScience (2025) showed: less than half of listeners could distinguish an AI voice from a human one — 42% on individual sentences and 47% on full audio clips. Yet synthetic voices were no less engaging to the brand, drawing attention and intent to buy compared to human voices. Technically, an AI voice has long been capable of convincingly conveying emotions.
The problem is not with the quality of synthesis, but with unification. If ten brands use the same voice engine with 'default' settings, their voices sound almost identical — this is the very effect that studies by ElevenLabs and WPP Media capture: quality increases, but distinctiveness decreases. Differentiation is no longer about whether the voice sounds natural, but whether it has a unique sonic signature — specific prosody, recognizable pauses, brand-specific verbal turns. Ipsos confirms the value of the question in its study The Power of You (2020): ads with a distinctive sonic identity are 8,5 times more likely to achieve high ad recognition rates than those without it.
How to measure the effectiveness of a voice brand
The model applied to NBVA (name-brand voice assistants) using Mercedes MBUX as an example structures the outcome in three levels, and this chain should form the basis of metrics:
- Drivers — voice uniqueness and quality of dialog scenarios;
- Intermediate outcomes — perception of brand identity through the voice and strength of the emotional connection with the user;
- Final outcomes — multi-dimensional loyalty: willingness to recommend, repeat usage, resilience to switching to a competitor.
In practice, specific metrics are needed for evaluation: voice recognition without visual association to the brand (blind recall test), attractiveness (pleasantness/irritating scale), uniqueness relative to competitors' voices, and correlation with NPS or repeat purchases.
How to conduct A/B testing of voices
- To record the same script and deliver it with two different voice settings (pitch, tempo, prosody) to two comparable audience groups.
- To measure the percentage of correct brand identification without visual cues — only by sound.
- To collect subjective evaluations on scales of 'warmth', 'competence', 'trust' — these should align with the target brand archetype.
- To test behavioral response: listening time until abandonment, conversion to target action after voice message.
- To repeat the test after 3–6 months — voice recognition improves with contact frequency; a single test yields artificially low results.
Voice Guide: how to document the brand voice
Tone of voice guide for text — a familiar practice, but there is almost no systematic approach to creating a similar document for sound. This creates a risk: different contractors configure TTS differently, and the company's voice gradually spreads across incompatible versions. A Voice Guide should close this gap and include at least the following sections:
- Reference audio recordings of the voice in 5–7 basic scenarios (greeting, refusal, apology, sale, error, farewell);
- Specific SSML parameter values — speech rate, volume range, pause duration, points of intonational rise;
- Prohibited voice states — what must never sound (for example, 'too cheerful' when refusing a refund);
- User address rule — fixed form of address using 'you'/'you', use of name;
- Voice connection with the rest of the brand's sound system (sonic identity system) — jingles, sound notifications, musical logo, so that the voice does not exist in isolation from the rest of the brand's sound landscape, as emphasized in the works of Barnes and subsequent sonic branding research;
Without such a document, even the most carefully selected synthetic voice will gradually start to sound more neutral and impersonal — not because the technology degrades, but because nobody monitors compliance with the original parameters at scale;
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
