Neural TTS

Speech Technologyalso: neural text to speechalso: neural speech synthesis

In one sentence

Neural text to speech (neural TTS) uses neural networks to generate speech directly, replacing the older concatenative and parametric methods, and it is the reason synthetic voices stopped sounding robotic.

4.14 MOS Independent mean opinion score testing, 2026Last reviewed 31 July 2026

Definition

Neural text to speech uses neural networks to generate speech. It replaced two older methods: one that stitched together recorded fragments, and one that generated audio from acoustic parameters.

It is the reason synthetic voices stopped sounding robotic.

The generational history is worth knowing, because it explains why most people's expectations of synthetic speech are years out of date.

The three generations

  • Concatenative. Recorded speech fragments from a voice actor, selected and joined together. It sounded natural inside a fragment and produced audible seams and abrupt prosody at the joins, and it needed an enormous recorded corpus for each voice.
  • Parametric. Audio generated from acoustic parameters by a statistical model. Smooth and flexible, and characteristically buzzy or muffled. This is the voice most people mean when they say robotic.
  • Neural. Networks generate the speech directly, usually by predicting an intermediate acoustic representation and converting it to a waveform, or by generating audio tokens end to end.

What neural approaches changed

  • Naturalness reached the human reference band on short samples, and on some measures matched it.
  • Making a voice became far cheaper. Adapting to a new voice takes minutes of audio rather than many hours.
  • Expressiveness became controllable instead of fixed.
  • Prosody improved a lot, though it is still the weakest area, particularly over long spans.

What did not change

  • Pronunciation of unusual words still needs a lexicon. The model guesses, and the guess is not deterministic across contexts.
  • Numbers, dates and codes still need explicit handling.
  • Prosodic drift over long passages persists.

The architectural direction

  • Two-stage designs, predicting an acoustic representation then vocoding it, have given way in part to end-to-end and autoregressive audio generation.
  • Streaming became a first-class requirement, because conversation needs output to begin before generation finishes.
  • Time to first byte of roughly 100 to 200 milliseconds is the practical expectation for streaming synthesis.

Common misconception

That neural TTS solved speech synthesis. It solved timbre and it largely solved naturalness on short samples. Prosody over long spans, pronunciation of domain vocabulary and the correct rendering of numbers are still live problems, and every production deployment runs into them.

Why it matters commercially

The jump in quality is big enough that a buyer whose mental model was formed on parametric voices will underestimate what is now possible. Here a demo carries more than a description, because the gap between what someone expects and what they hear is the thing that persuades.

In voice specifically

Text has nothing to synthesize. A written answer renders the same for everyone, and a price on a screen cannot be mispronounced. Neural TTS is the step that gives a voice agent its voice, and it is where the brand name, the price and the abbreviation each get a chance to come out wrong.

Where AsqVox fits

The Orb speaks with a neural voice. Voice quality is a brand decision, so the useful test is candidate voices on the worst content the agent will handle, meaning prices, names and reference codes drawn from your uploaded documents, rather than a scripted greeting.

Visual

Three generations, and what each still gets wrong

Three generations, and what each still gets wrongHow it generates soundWhat it sounds likeWhat it still gets wrongMaking a new voiceConcatenativeRecorded fragments selected andjoined togetherNatural inside a fragment,seams between themAudible joins and abruptprosodyMany hours of recorded audioper voiceParametricAudio built from acousticparameters by a statisticalmodelBuzzy and muffled, the soundpeople mean by roboticSounds synthetic throughoutNo recorded corpus,expressiveness fixedNeuralA network generates thewaveform directlyAt or near the human referenceon short samplesLong-span prosody, domainpronunciation, numbersMinutes of audio to adapt avoice

It solved how the voice sounds, not what it says the numbers are.

The quality figures are in the statistics below: ElevenLabs was measured at around 4.14 mean opinion score in independent testing, against a human reference band of 4.5 to 4.7. The three-generation split is the standard account of synthesis history; the characteristic failures are what every production deployment still meets.

Statistics

Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.

ElevenLabs was measured at around 4.14 mean opinion score in independent testing.

4.14 MOSindependent

Independent mean opinion score testing, 2026 - MOS is an averaged subjective judgment, not a measurement, and figures from different studies are not on a common scale. Read it against the human reference band below rather than against another vendor number.

The strongest open-source result cited is Sesame CSM at around 4.7 mean opinion score.

4.7 MOSindependent

Reported mean opinion score results, cited as the leading open-source figure, 2026 - Cited from scattered study results rather than a single maintained leaderboard, so treat the ranking as directional. Above roughly 4.5 the scale stops separating systems usefully.

Human speech is generally placed at 4.5 to 4.7 on the same scale, so the top of the field now sits at or near the human reference for naturalness on short samples.

4.5 to 4.7industry range

Standard mean opinion score reference band, 2026 - A genuinely recent development. It also means naturalness on short samples has stopped being the axis worth choosing on, because the leaders are already there.

Streaming synthesis time to first byte typically lands at 100 to 200 milliseconds.

100 to 200 msindustry range

Common production budgeting convention, 2026 - An engineering convention drawn from production practice rather than a published research finding. It applies to streaming synthesis only.

ElevenLabs raised USD 500m at an USD 11bn valuation in a Series D led by Sequoia on 4 February 2026, roughly tripling its USD 3.3bn valuation a year earlier, on total funding of USD 781m and reported ARR above USD 330m at the end of 2025.

USD 11bnindependent

ElevenLabs Series D announcement, 4 February 2026, 2026 - The size of that valuation relative to the reported revenue is a fact about how the market prices voice, and it is worth presenting neutrally rather than as evidence of quality.

No standard independent benchmark separates intelligibility from naturalness, and none measures pronunciation accuracy on domain-specific vocabulary.

-no reliable figure

Both matter more than mean opinion score for a business deployment, and neither can be looked up. The substitute is testing candidate voices on your own product names, prices and postcodes.

No benchmark measures prosodic quality over long-form content, which is where the remaining weakness concentrates.

-no reliable figure

It is the reason a voice can be rated natural on one sentence and mechanical over five, and there is no published way to measure the drift.

Examples

In practice

A team picks a voice from demo samples and ships it against real content full of prices, postcodes and product names. Complaints arrive at once, and none of them are about the voice as a voice. The demos were short scripted lines; production is thick with numbers and names. The team reruns voice selection on the agent's own worst-case content, and the voice it keeps is a different one.

The everyday version

Neural text to speech is why computer voices stopped sounding like robots. If your picture of a synthetic voice is more than a few years old, it is out of date. What has not changed: it still guesses at unusual words and prices unless you tell it how to say them.

Usage

Who says it

  • Text-to-speech engineers and vendors, often as a differentiator, though it is now the baseline rather than a distinguishing feature.
  • It turns up in product documentation and voice quality discussions.

Where it turns up

  • On a spec sheet it sits next to the available voices, the language and accent list, streaming support, time to first byte, SSML support and the commercial usage rights.

Common misuse

  • Presenting neural synthesis as a differentiator when it is now standard.
  • Evaluating voices on scripted demo content instead of real production content.
  • Assuming the gains in naturalness removed the need for pronunciation configuration.

Questions people ask

What is neural TTS?

Neural text to speech uses neural networks to generate speech directly, in place of two older methods: concatenative synthesis, which stitched together recorded fragments, and parametric synthesis, which generated audio from acoustic parameters. Neural approaches are the reason synthetic voices stopped sounding robotic, and they are now the baseline rather than a distinguishing feature.

What is the difference between neural, concatenative and parametric text to speech?

Concatenative synthesis selects and joins recorded speech fragments, which sounds natural inside a fragment but produces audible seams. Parametric synthesis generates audio from acoustic parameters and sounds buzzy or muffled, which is the voice people mean by robotic. Neural synthesis generates the waveform directly with a neural network and reaches the human reference band for naturalness on short samples.

Did neural TTS solve text to speech?

It solved timbre and largely solved naturalness on short samples, but not the whole problem. Prosody over long spans, pronunciation of domain vocabulary such as brand and product names, and the correct rendering of numbers, dates and codes are all still live problems that production deployments run into. Neural quality alone does not fix them; a pronunciation lexicon and explicit number handling do.

How natural do neural TTS voices sound?

The leading systems are at or near the human reference band on short samples. ElevenLabs was measured at around 4.14 mean opinion score in independent testing and the strongest open-source result cited, Sesame CSM, at around 4.7, against a human reference band of 4.5 to 4.7. Above roughly 4.5 the scale stops separating systems usefully, so naturalness on short samples is no longer the axis worth choosing on.

Share this definition

Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.