Making Voice AI Actually Sound Human
Punctuation, number formatting, and pacing tricks that fix 80% of robotic voice output without changing the model.
When a voice agent sounds robotic, the instinct is to blame the voice model and go shopping for a better one. In most cases the text is at fault, and the same model will sound dramatically more natural once the input is prepared properly — at no additional cost.
Numbers are the biggest offender
TTS engines read digits inconsistently. "12/05/2026" may be read as three separate numbers; "€250,000" often comes out as a digit-by-digit recitation. Convert every number to words before synthesis, and choose the form a human would use: "two hundred and fifty thousand euros", not the literal numeral.
Phone numbers are the opposite case: they should be grouped and paused, not read as one enormous number. Insert separators so the listener can actually write it down.
Punctuation is prosody
The engine derives pauses and intonation almost entirely from punctuation. Long clause chains with no commas produce a breathless monotone. Add commas where a human would breathe, end questions with a question mark so the pitch rises, and break any sentence longer than about twenty words into two.
An ellipsis or an em dash creates a longer, more natural hesitation than a comma — useful right before a price or a name you want emphasised.
Homographs and stress
Written text often omits the cues a reader relies on for stress and meaning, so the engine must guess. For the words that matter most in your domain — product names, place names, recurring phrases — insert explicit phonetic spelling in the pronunciation dictionary and keep it in version control.
Brand names and mixed vocabulary
Unusual brand names embedded in a sentence are the fastest way to break the illusion; the engine either mispronounces them or applies the wrong accent mid-sentence. Add them to the pronunciation dictionary based on how your customers actually say them.
Pacing and streaming
For real-time agents, synthesise sentence by sentence and start playback on the first sentence while the rest generates. This cuts perceived latency dramatically. Slightly slowing the overall rate — around 0.95x — makes output noticeably easier to follow without sounding sluggish.
Also add a short pause after the greeting. A voice that starts talking the instant the call connects sounds like a machine; a beat of silence sounds like a person picking up.
Choosing a voice
Test candidate voices with your worst text, not your best. Long sentences, numbers, product names, and a question. Have three native speakers rate them blind. Voices that score well on marketing samples routinely fall apart on real domain content.
The pipeline we ship
Normalise digits to words, with special handling for dates, currency and phone numbers.
Apply the pronunciation dictionary for domain terms and brand names.
Insert punctuation-based breathing points and split long sentences.
Synthesise per sentence and stream the audio.
Cache the audio for fixed phrases such as greetings and hold messages — it saves both money and latency.
None of this requires a different model. It is text preparation, and it is the difference between a voice customers tolerate and one they do not notice is synthetic.