Text to speech

Speech TechnologyTTSalso: speech synthesisalso: voice synthesis

In one sentence

Text to speech (TTS), also called speech synthesis, turns written text into spoken audio, and it is how a voice agent gets its voice.

4.14 MOS Independent mean opinion score testing, 2026Last reviewed 30 July 2026

Not to be confused with Voice cloning, or Speech-to-speech models.

Definition

Text to speech converts written words into audio a person can hear. It is where a voice agent gets its voice from.

Every visitor hears it, which makes the choice a brand decision as much as an engineering one.

Contemporary synthesis is neural and generative. Two older approaches are responsible for the mechanical voices most people still picture when the subject comes up: concatenative synthesis, which stitched recorded fragments together, and parametric synthesis, which generated audio from acoustic parameters. Neural TTS usually predicts an intermediate acoustic representation from the text, a mel spectrogram being the common one, and hands it to a vocoder that turns it into a waveform. Newer end-to-end and autoregressive audio models drop the intermediate step altogether.

The dimensions quality gets judged on

  • Naturalness. Does it sound like a person rather than a machine.
  • Intelligibility. Can it be understood, which is a different property and sometimes an opposed one. Highly natural speech with casual articulation can be harder to make out than slightly stiff, clear speech.
  • Prosody. The rhythm, stress and intonation carried across a whole sentence. This is the hardest part to get right, and where whatever still sounds synthetic tends to be hiding.
  • Expressiveness. Emotional color that suits the content.
  • Consistency. The same voice sounding like the same voice across sessions.

Streaming, and what synthesis adds to the wait

  • With streaming, audio starts flowing once the first few tokens exist rather than after the whole text is ready. First byte usually arrives somewhere in the 100 to 200 millisecond band.
  • Turn streaming off and the entire synthesis run lands on top of the wait before any sound reaches the visitor, which on a long answer destroys any sense of conversation.
  • Which is the reason to start speaking the opening clause while the language model is still writing the rest.

How delivery gets controlled

  • SSML, the speech synthesis markup language. It is a W3C standard that lets you mark up text with pauses, stress, pronunciation and pace. How much of it a provider implements varies, and partial coverage is normal.
  • Inline prompt-based control in newer models, where the delivery you want is described in natural language instead of marked up.
  • Pronunciation dictionaries and phoneme overrides, which are essential for brand names, place names and industry terms.

Common misconception

That whichever voice sounds most human is the one to pick. In a business agent, being understood, sounding the same in every session and starting quickly nearly always count for more than emotional range. A voice a shade less lifelike that never fumbles the product name and is speaking inside 120 milliseconds beats a beautiful one that hesitates and invents its own pronunciations.

Why it matters commercially

The voice is part of the brand. Everyone who arrives hears it, and it conveys attitude in a way a paragraph on a page never will. It is also the piece most often chosen because a demo charmed somebody and regretted once it is live, since a demo runs on short scripted lines while production runs on long answers stuffed with names and figures.

In voice specifically

Text has no equivalent decision to make. A written answer renders the same way for everyone, and a postcode on a screen cannot be mispronounced. In speech the same sentence still has to be delivered, and delivery is where the brand name, the price, the abbreviation and the question mark each get their chance to go wrong.

Visual

Where synthesis quality actually breaks

Where synthesis quality actually breaksMilliseconds into one spoken answerStreaming synthesisaudio plays while the model is still writing the restWithout streamingwhole answer synthesized first, visitor hears silenceaudio finally startsFive places itbreaksbrand name: pronunciation, needs a dictionary entryprice: number formatting, pounds or the digitabbreviation: spoken or spelled outpostcode: character by character deliveryquestion mark: prosody, does the pitch risestreaming starts here, not at the endwhere a non-streaming voice would start

Demos are short scripted sentences. Production is names, numbers and postcodes.

The sentence is "The Aeron Pro is 1,240 pounds 50 including VAT, in stock at SW1A 2AA. Shall I hold one?" Only the 100 to 200 millisecond time-to-first-byte band is grounded; every other position is illustrative. Quality is judged on four axes at once: naturalness, intelligibility, prosody and consistency. Naturalness and intelligibility pull against each other, because casual articulation sounds more human and is harder to make out.

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.

On that same scale real human speech usually lands between 4.5 and 4.7, so the leading systems now sit at or beside the human reference for naturalness.

4.5 to 4.7industry range

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

Streaming text to speech 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, and it should be presented that way. 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 scale of that valuation relative to the reported revenue is itself a fact about how the market prices voice, and it is worth presenting neutrally rather than as evidence of quality.

Nobody publishes a standard independent benchmark that separates intelligibility from naturalness, and none exists for how accurately a system pronounces domain vocabulary either.

-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.

Examples

In practice

A financial services agent reads account balances out loud. The default voice handles 1,240.50 inconsistently and sometimes produces something that sounds like a date. The fix sits upstream: turn numeric values into explicit spoken form before they ever reach the synthesizer instead of leaving the model to interpret them. Normalizing numbers and currency is dull, routine work, and nearly every production deployment finds that out only after it has shipped.

The everyday version

Text to speech is the voice your customers hear. Choosing one is like choosing how a receptionist should sound: warm or brisk, formal or friendly. The test that counts is not a greeting read aloud. It is whether the thing pronounces your company name properly and gets through a price without tripping.

Usage

Who says it

  • Engineers say TTS, and want to know about streaming, latency, and how much of SSML is genuinely implemented.
  • Marketing and brand teams care which voice gets chosen and whether it stays the same, and that call belongs with them.
  • Procurement meets it through licensing, and asks two questions: is this voice cleared for commercial use, and was it modeled on a real person.

Where it turns up

  • On a spec sheet it sits beside the voice catalog, the language and accent list, whether streaming is supported, time to first byte, how much SSML is implemented, whether pronunciation dictionaries are available, and what the commercial usage rights allow.
  • Licensing terms on a voice carry more weight than buyers anticipate, above all when the voice was built from a real performer.

Common misuse

  • Picking a voice off a vendor showreel instead of off your own most difficult material. Audition it on the single ugliest sentence the agent will ever be asked to read.
  • Reading SSML support as a yes or a no. Partial is the normal answer, and the tags that are missing tend to be the ones you were counting on.
  • Treating text to speech and voice cloning as one thing. Cloning is a particular capability inside synthesis, and it drags its own legal and ethical obligations along with it.

Questions people ask

What is the difference between text to speech and speech synthesis?

They name the same thing. Speech synthesis is the older technical term for generating speech from text and text to speech, or TTS, is the common product name for it. Voice cloning is not a synonym: it is one capability inside synthesis, and it carries its own legal and ethical requirements.

How quickly does text to speech start speaking?

Streaming synthesis typically produces its first byte of audio in around 100 to 200 milliseconds, which is an engineering convention from production practice rather than a published research finding. Without streaming, the whole answer has to be synthesized before anything plays, so the wait grows with the length of the answer.

How should I choose a text to speech voice?

Not on how it sounds reading a greeting. For a business agent, intelligibility, consistency and latency usually matter more than expressiveness, so test candidates on your worst content: product names, prices, postcodes and abbreviations. A voice that never mispronounces your brand name and starts speaking quickly beats a more lifelike one that hesitates.

What is SSML and do I need it?

SSML is the speech synthesis markup language, a W3C standard for annotating text with pauses, emphasis, pronunciation and rate. You need it as soon as the default delivery gets something wrong, most often a brand name or a string of digits. Support is usually partial rather than complete, so check the specific tags you need instead of accepting a yes.

Share this definition

Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.