Tone in chatbots and voice assistants
Voice and text — different channels, different risks
In articles about tone of voice for bots, text parameters are almost always described: formality, emotionalness, brand persona. The voice channel operates by different rules. Text can be edited before sending, but voice is generated and played immediately — there is no room for error. A user reading a message sets the intonation in their head. The listener receives a ready-made intonation from the system, and if it is off, the meaning of the phrase is distorted physically, not stylistically.
| Parameter | Text chatbot | Voice assistant |
|---|---|---|
| Medium of tone | Lexicon, punctuation, emojis | Tone height (F0), tempo, pauses, timbre |
| Opportunity to reframe | User reads again | Perceived once, in real time |
| Source of error | Poor wording | Incorrect stress, broken question intonation |
| What creates the persona | Letter style, phrase templates | Actor's voice + neural network synthesis model |
How the 'persona' sounds: F0, pauses, intonational stresses
The technical tone of the voice bot is built around the F0 contour — the fundamental frequency, that same pitch of voice that rises at the end of a question and falls at the end of an affirmative sentence. If the system incorrectly places stresses, especially in questions, the phrase sounds fake even with perfect pronunciation of the words. There is an unexpected detail here: in the task of intonational stress, simple rule-based algorithms often work more accurately than neural networks — because the rules of the language regarding question formation are formalizable, while the neural network sometimes 'floats' on non-standard constructions.
Pauses — the second most significant tool. An overly short pause before an important word creates a feeling of rushing and carelessness, while an overly long one reads as hesitation or uncertainty. With Alice, Aphi (MTS), and Joy (Sber), the base voice is recorded by professional actors: this provides a recognizable timbre and individuality, and then the neural network reproduces and scales this voice for any phrases, preserving the character set by the live performer.
Who and how synthesizes emotions: approaches of Yandex, MTS AI, and Sber
Different companies solve the task of emotional tone in different ways, and the differences are telling.
- MTS AI implemented the synthesis of five basic emotions — calmness, joy, anger, sadness, surprise — for voice bots of call centers. The result is measurable: the share of users requiring transfer to an operator has decreased.
- Yandex focused not on discrete emotions, but on smooth interpolation: stylistic parameters of the voice change within a single response without abrupt jumps, so the bot does not 'switch' between moods like between buttons.
- Sber builds personalities (Joy and others) through the linkage of a live actor's voice and a neural network model that preserves the individuality of the timbre when generating new phrases.
The connection with branding here is direct: F0-contour, pauses, and emotional styles are not a technical setting 'for engineers,' but material from which the voice identity of a brand is physically constructed. Managing the tone of a voice bot without understanding these parameters is the same as discussing packaging design without knowing what color is.
When to change the tone: adaptation to user emotions
Static tone — an error regardless of the channel, but in voice it is especially noticeable. Working practice — dynamic routing through sentiment analysis: if the client is irritated, the dialogue switches to the 'empathetic manager' scenario, if they ask a technical question about a tariff — to the 'clear salesperson' scenario. The tone adjusts not to the topic of the request, but to the emotional state of the person at that moment.
What the data shows
Studies confirm the effect of empathetic tone: it increases the perception of the bot as 'human-like' and raises the willingness to recommend the company. The reverse example is no less important: a playful tone in a delivery problem situation worsens all satisfaction metrics — humor at the moment when the client is angry about the order delay is perceived as disrespect, not as friendliness.
Uncanny valley: where the voice stops being 'one's own'
The problem of the uncanny valley effect for synthesized speech exists, but it works differently than for robot faces. The negative reaction does not arise from low synthesis quality — in that case, the user simply recognizes 'it's a bot' — but in the zone where the voice sounds almost human but with a barely noticeable glitch: an unnatural pause, slightly incorrect stress, or overly smooth without a single roughness in intonation. The difference between the text and voice channel is fundamental here: the brain automatically corrects a typo in text without tension, while a voice inaccuracy is perceived as something physically 'not right' — a more bodily and wary reaction. Hence the practice: voices like Alice or Joy deliberately do not strive for photorealistic humanity, keeping a recognizable synthetic character — this reduces the risk of falling into the rejection zone.
Common mistakes when setting the tone of voice for voice bots
- They transfer the ready-made text tone of voice into the voice script without adaptation — without accounting for pauses and intonation, the phrase sounds unnatural, even if the words are chosen correctly.
- They use one emotional register throughout the entire dialogue, ignoring the change in the user's state during the conversation.
- They add humor to scripts with high emotional tension—during complaints and problems, this almost always reduces trust.
- They strive for maximum human-likeness in voice, not considering the risk of falling into the uncanny valley.
- They do not take into account cultural differences in perceiving synthesized voices and emotions: what sounds natural and empathetic in one linguistic context may be perceived as overly theatrical in another.
Voice and text — different channels, different risks
In articles about tone of voice for bots, text parameters are almost always described: formality, emotionalness, brand persona. The voice channel operates by different rules. Text can be edited before sending, but voice is generated and played immediately — there is no room for error. A user reading a message sets the intonation in their head. The listener receives a ready-made intonation from the system, and if it is off, the meaning of the phrase is distorted physically, not stylistically.
| Parameter | Text chatbot | Voice assistant |
|---|---|---|
| Medium of tone | Lexicon, punctuation, emojis | Tone height (F0), tempo, pauses, timbre |
| Opportunity to reframe | User reads again | Perceived once, in real time |
| Source of error | Poor wording | Incorrect stress, broken question intonation |
| What creates the persona | Letter style, phrase templates | Actor's voice + neural network synthesis model |
How the 'persona' sounds: F0, pauses, intonational stresses
The technical tone of the voice bot is built around the F0 contour — the fundamental frequency, that same pitch of voice that rises at the end of a question and falls at the end of an affirmative sentence. If the system incorrectly places stresses, especially in questions, the phrase sounds fake even with perfect pronunciation of the words. There is an unexpected detail here: in the task of intonational stress, simple rule-based algorithms often work more accurately than neural networks — because the rules of the language regarding question formation are formalizable, while the neural network sometimes 'floats' on non-standard constructions.
Pauses — the second most significant tool. An overly short pause before an important word creates a feeling of rushing and carelessness, while an overly long one reads as hesitation or uncertainty. With Alice, Aphi (MTS), and Joy (Sber), the base voice is recorded by professional actors: this provides a recognizable timbre and individuality, and then the neural network reproduces and scales this voice for any phrases, preserving the character set by the live performer.
Who and how synthesizes emotions: approaches of Yandex, MTS AI, and Sber
Different companies solve the task of emotional tone in different ways, and the differences are telling.
- MTS AI implemented the synthesis of five basic emotions — calmness, joy, anger, sadness, surprise — for voice bots of call centers. The result is measurable: the share of users requiring transfer to an operator has decreased.
- Yandex focused not on discrete emotions, but on smooth interpolation: stylistic parameters of the voice change within a single response without abrupt jumps, so the bot does not 'switch' between moods like between buttons.
- Sber builds personalities (Joy and others) through the linkage of a live actor's voice and a neural network model that preserves the individuality of the timbre when generating new phrases.
The connection with branding here is direct: F0-contour, pauses, and emotional styles are not a technical setting 'for engineers,' but material from which the voice identity of a brand is physically constructed. Managing the tone of a voice bot without understanding these parameters is the same as discussing packaging design without knowing what color is.
When to change the tone: adaptation to user emotions
Static tone — an error regardless of the channel, but in voice it is especially noticeable. Working practice — dynamic routing through sentiment analysis: if the client is irritated, the dialogue switches to the 'empathetic manager' scenario, if they ask a technical question about a tariff — to the 'clear salesperson' scenario. The tone adjusts not to the topic of the request, but to the emotional state of the person at that moment.
What the data shows
Studies confirm the effect of empathetic tone: it increases the perception of the bot as 'human-like' and raises the willingness to recommend the company. The reverse example is no less important: a playful tone in a delivery problem situation worsens all satisfaction metrics — humor at the moment when the client is angry about the order delay is perceived as disrespect, not as friendliness.
Uncanny valley: where the voice stops being 'one's own'
The problem of the uncanny valley effect for synthesized speech exists, but it works differently than for robot faces. The negative reaction does not arise from low synthesis quality — in that case, the user simply recognizes 'it's a bot' — but in the zone where the voice sounds almost human but with a barely noticeable glitch: an unnatural pause, slightly incorrect stress, or overly smooth without a single roughness in intonation. The difference between the text and voice channel is fundamental here: the brain automatically corrects a typo in text without tension, while a voice inaccuracy is perceived as something physically 'not right' — a more bodily and wary reaction. Hence the practice: voices like Alice or Joy deliberately do not strive for photorealistic humanity, keeping a recognizable synthetic character — this reduces the risk of falling into the rejection zone.
Common mistakes when setting the tone of voice for voice bots
- They transfer the ready-made text tone of voice into the voice script without adaptation — without accounting for pauses and intonation, the phrase sounds unnatural, even if the words are chosen correctly.
- They use one emotional register throughout the entire dialogue, ignoring the change in the user's state during the conversation.
- They add humor to scripts with high emotional tension—during complaints and problems, this almost always reduces trust.
- They strive for maximum human-likeness in voice, not considering the risk of falling into the uncanny valley.
- They do not take into account cultural differences in perceiving synthesized voices and emotions: what sounds natural and empathetic in one linguistic context may be perceived as overly theatrical in another.
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
