Robotic vs natural voices: how modern TTS voices are made

Updated October 31, 2026

Some computer voices sound like a robot from an old film. Others are hard to tell apart from a person. The difference comes down to how the sound is made.

Formant synthesis: the classic robot voice

Every vowel you speak has a few strong resonances, called formants, created by the shape of your mouth and throat. Formant synthesis builds speech from scratch by generating a buzzing or hissing sound and filtering it to create those resonances, following rules written by experts.

• Strengths: very small and fast, perfectly clear even at high speeds, and easy to extend to new languages. That is why screen reader users often like it, and why eSpeak NG supports so many languages.
• Weaknesses: it sounds artificial, because real voices have countless small details that rules don't capture.

On our site, the standard voices use this method.

Concatenative synthesis: recorded pieces

A voice actor records many hours of carefully chosen sentences. The recordings are cut into tiny units, such as half-syllables. To say something new, the system finds the best units and joins them.

• Strengths: each piece is a real human voice, so it can sound very natural.
• Weaknesses: needs a large recording database for every voice; the joins can be audible; and the voice can't easily change its tone.

Neural synthesis: learning to speak

Neural text to speech uses machine learning. A model is trained on many hours of recorded speech together with the matching text. It learns how text maps to the sounds, rhythm and melody of speech, then generates new audio for any sentence.

Most modern systems work in two parts. An acoustic model predicts what the speech should look like (often as a spectrogram, a picture of sound frequencies over time), and a vocoder turns that into an audio wave. Newer models like VITS, which Piper is based on, do both in one network.

• Strengths: natural rhythm, intonation and voice quality; pleasant to listen to for a long time.
• Weaknesses: needs more computing power, and each language needs good training recordings. Neural models can occasionally mispronounce rare words or names.

The voices on our site

• Premium voices (Kokoro) are among the most natural open-source voices available. Kokoro is a compact neural model with about 82 million parameters. It takes a little longer to generate, so we mark these voices as premium.
• Natural voices (Piper) are fast neural voices trained for dozens of languages. They are a great balance of speed and quality.
• Standard voices (eSpeak NG) use formant synthesis. They are less natural but cover languages that neural voices don't yet support, and they are very clear.

All three run on our own servers. Your text is not sent to a third-party AI service.

How to choose a voice

• For listening to long articles or narration, pick a premium or natural voice.
• For very fast listening, or a language without neural voices, a standard voice works well.
• Try a few: the same text can sound quite different between voices, and some suit certain content better.

Compare the voices yourself by converting the same sentence with each.