DEV Community

Cover image for TTS Full Form: How Text-to-Speech Works on Phone Calls
Piyush Sahoo
Piyush Sahoo

Posted on

TTS Full Form: How Text-to-Speech Works on Phone Calls

The TTS full form is text-to-speech: software that reads written text aloud in a generated voice. On a phone call, that audio must then cross a network that often carries only 300 Hz to 3.4 kHz at 8,000 samples per second. The engine creates the voice, but the telephone path decides what callers hear.

This guide follows one sentence from your server to a caller's handset. It covers the engine stages, the telephone limits, the problems local languages and formats cause, how to choose an approach, and the rules for automated voice.

Note: Key takeaways

  • TTS is a component, not a product category. An IVR or voice agent uses it as its output stage.
  • A voice has to be judged at 8 kHz. Studio-rate samples tell you little about a narrowband call.
  • Local text formats break generic front ends first. Amounts, dates, and mixed-language sentences, such as dot-grouped thousands or Spanish-English scripts, need testing before voice quality does.
  • Time to first audio matters more than total render time. Streaming hides render time from the caller, and barge-in needs queued audio to be discarded.
  • Artificial voice counts as an automated call. The US TCPA covers AI-generated voices, and India's TRAI adds an advance declaration duty.

What does TTS stand for?

The letters expand to text-to-speech, and engineers also call it speech synthesis. Outside technology, the same letters can mean "true to size" on clothing listings.

Four related terms are often mixed up, so here is how they differ:

Term Direction or role Example on a call
TTS Text in, audio out Reads "your order ships today" to the caller
STT or ASR Audio in, text out Turns the caller's reply into words
IVR A call flow that may use TTS or recordings "Press 1 for sales"
Voice agent STT, a language model, and TTS in a loop Holds an unscripted conversation

How does text become audio on a call?

The work splits into two parts, and most explanations cover only the first. The engine produces audio, and the telephone path then reshapes it.

Nine-stage pipeline. Inside the TTS engine: normalise, letters to sounds, plan the delivery, acoustic model, vocoder. On the telephone path: resample to 8,000 samples a second, encode with G.711 at 64 kbps, packetise into 20 ms RTP packets, then network and handset at about 300 to 3,400 Hz.

Stages inside the engine

The front end works on language. It expands symbols into words, maps words to phonemes, and decides on stress and pauses. W3C's SSML 1.1 specification describes these same stages, and it offers markup to override them when the engine guesses wrong.

The back end works on sound. Tacotron 2 set a pattern that many systems still resemble: one network predicts a mel spectrogram from text, and a vocoder renders that into a waveform. Later research targeted speed, with FastSpeech generating spectrograms through a feed-forward network and HiFi-GAN offering a more efficient vocoder.

Older methods still help explain the vocabulary you meet in vendor documents:

Method How it makes sound What it means for a call
Concatenative Joins stored fragments of recorded speech Bound by what was recorded, and joins can be audible
Statistical parametric Generates acoustic parameters from a model Compact and flexible, with a buzzy character
Neural Learns to generate audio from training data Natural prosody, but compute and latency need planning

The Tacotron 2 authors reported a mean opinion score of 4.53, against 4.58 for professional recordings. That is their own lab result, so don't read it as a phone-line figure.

Stages on the telephone path

Telephony has fixed arithmetic. RFC 3551 defines PCMA and PCMU, the RTP names for ITU-T G.711 A-law and μ-law, at 8,000 samples per second and 8 bits per sample. That works out at 64 kbps. A 20 ms packet therefore holds 160 samples, which is 160 bytes of audio.

A-law is used in most countries outside North America and Japan, which mainly use μ-law. Your TTS output may pass through a conversion if the engine and the network use different laws, so it pays to generate as close to the final format as you can.

Why does a phone line change how a voice sounds?

A sampled signal can only represent frequencies up to half its sample rate. An 8 kHz call therefore stops at 4 kHz in theory and about 3.4 kHz in the usual narrowband definition. The engine's output, often generated at 22 or 24 kHz, loses everything above that line.

Bar chart on a 0 to 12 kilohertz scale. Narrowband call at 8 kilohertz with G.711 covers 300 hertz to 3.4 kilohertz. Wideband call with G.722 covers 50 hertz to 7 kilohertz. A 16 kilohertz stream reaches 8 kilohertz and 24 kilohertz TTS reaches 12 kilohertz. The zone above 4 kilohertz holds most of the energy of English s.

The loss falls hard on some consonants. Phonetician Geoff Lindsey explains that English /s/ carries its strong energy mainly above about 4 kHz, while /f/ is weak and spread out. Listeners fill those gaps from context. A synthetic prompt that reads an unfamiliar name or code gives them little context to use.

Four practical rules follow:

Higher rates still help in some cases. G.722 reaches 7 kHz where both ends support it. But a high-rate payload doesn't guarantee that rate at the handset (audio formats docs), because a later leg may use a lower-rate codec. Here is the size of each format:

Format Highest frequency possible Payload per 20 ms Audio bit rate
μ-law, 8 kHz 4 kHz 160 bytes 64 kbps
L16, 8 kHz 4 kHz 320 bytes 128 kbps
L16, 16 kHz 8 kHz 640 bytes 256 kbps
L16, 24 kHz 12 kHz 960 bytes 384 kbps

We calculated the sizes from each format's sample rate and bit depth.

Why do local languages and number formats trip up TTS?

Real call scripts break an engine's front end before voice quality does, usually on the parts that carry the money and the dates. Every market has its own traps, such as month-first dates in the US or a dot as the thousands separator in much of Europe. The example below is a Spanish payment sentence with an English word in it.

Illustrative Spanish sentence about a payment. €1.500 should be read as one thousand five hundred euros, because the dot groups thousands, not as one point five. SMS should be read as letters. 12/11/2026 should be read day first as 12 November, not December eleventh. The English word online should keep its English sound. The spoken result is written out below.

  • Number grouping. Separators differ by market. Much of Europe writes 1.500 for fifteen hundred, and Indian notation groups digits around lakh and crore. A generic rule may misread either.
  • Irregular number words. Some languages need a real lookup rather than a tens-and-units rule. In Hindi, for example, nearly every number from 1 to 99 is irregular.
  • Dates. Check the day and month order on every template. SSML's say-as date format exists to remove that ambiguity, where your engine supports it.
  • Silent vowels. Some scripts don't mark which vowels are silent. In Devanagari this is a known problem for speech synthesis. For example, नमक is said namak, not namaka.
  • Language mixing. Sentences that switch languages, such as Spanish-English or Hindi-English, are common in call scripts, and an engine may change accent or misread the borrowed word. Multilingual markets have many such pairs; India alone has 22 languages in the Eighth Schedule of its Constitution.
  • Names and codes. Local names, branch names, and policy numbers are the terms a caller can't guess from context.

A simple defence works with any engine. Write amounts and dates out as words in your own template layer, keep a pronunciation list for names and product terms, and read reference codes in groups of digits. Your application then controls the reading instead of each engine's front end.

What engines publish for non-English voices

This table uses Indian languages as the example. It reports vendor documentation checked on November 3, 2026, plus our own Deepgram guide. It compares coverage only, not quality.

Engine Indian-language coverage Telephony output
Sarvam Bulbul 10 Indian languages plus English on v3, the default model. The newer v4 Flash lists 8 Indian languages so far 8,000 Hz available, but the default is 22,050 Hz if you omit it
Amazon Polly Hindi and English (Indian) voices μ-law and A-law at 8,000 Hz
Google Cloud TTS Hindi and other Indian-language voices IVR audio profile
ElevenLabs Hindi and Tamil among listed languages μ-law and A-law at 8 kHz
Deepgram Flux English voices with an Indian accent See the integration guide

Should you use TTS or recorded prompts?

Use recordings when the wording is fixed and matters, and use TTS when the wording changes per call. The real choice has four options, because you can also render TTS once and treat the result as a file.

Decision map. If wording does not change per call, use a recorded file or TTS rendered once, played as an audio file. If it changes and the caller converses freely, use a streaming engine through a bidirectional audio stream or a SIP trunk to an AI platform. If it changes in a template and your language and voice are on the built-in text-to-speech list, use built-in text-to-speech. Otherwise render templates with your chosen engine, cache them, and play them as audio files.

Recorded file TTS at call time TTS rendered and cached Streaming engine
Best for Brand intros, legal notices Balances, dates, order status Templates with a limited set of values Voice agents
Updating wording New recording session Edit the text Edit and re-render Prompt or model change
Voice choice Any human voice The platform's list Any engine Any engine with streaming
Failure to plan for Stale content A misread number A cache miss on a new value Delay and interruption

Mixing a studio voice with TTS in one sentence usually sounds odd, because the caller hears the voice change at the variable. Keep one voice across a whole prompt where you can. If you pre-render, cache by the exact text, voice, and format, so a voice change never serves stale audio.

Why does time to first audio matter?

Callers react to silence, so the wait before the first sound matters more than total render time. Total render speed is measured as real-time factor, where a value below 1 means audio is generated faster than it plays.

Illustrative timeline for a 6 second reply. Full-file synthesis gives 1.2 seconds of silence and ends at 7.2 seconds. Streaming starts at 0.2 seconds and ends at 6.2 seconds. If the caller interrupts at 3 seconds, 3.2 seconds of queued audio must be cleared.

The numbers above are assumed, but the shape holds. Three design points follow from it:

  • Stream sentence by sentence. Your language model and your engine both produce output gradually, so don't wait for a full reply.
  • Keep the queue short. Audio already sent still has to play or be discarded when the caller interrupts. Send small outbound chunks so barge-in stays responsive.
  • Measure the tail. Ask for first-byte latency at p95 or p99 on your actual audio format, because an average hides the slow calls.

How do you test a TTS voice for telephony?

Vendor scores won't answer this for you. Mean opinion score comes from listener ratings, so two vendors' figures compare only under the same test conditions. Run your own short test instead:

  1. Collect real lines. Take a sample from your own scripts, weighted towards amounts, dates, names, and mixed-language sentences.
  2. Render in the final format. Use the rate and codec your calls will really use, and keep the settings identical across engines.
  3. Listen on real calls. Place them to ordinary handsets on your intended routes. Test calls cost money, so use your own credentials and confirm each number first.
  4. Score token types separately. Count errors for amounts, dates, names, and borrowed words, because an overall average hides the failures that matter.
  5. Try an ASR back-test. Running TTS output through a recogniser gives a cheap, repeatable intelligibility signal. Use it alongside human listeners, not instead of them.
  6. Time it. Record first-audio delay from your own region, at p95, during your busy hours.

Also check where the text goes. A prompt may contain a customer's name and balance, and a cloud engine will receive it. Vobiz's XML best practices say never to speak card numbers or passwords in TTS.

What rules apply to TTS calls?

Regulators treat a synthetic voice as an automated call, so the voice technology can change your obligations. These rules change often, so confirm the current text and take legal advice.

United States. The FCC's February 2024 declaratory ruling confirms that the TCPA's limits on "artificial or prerecorded voice" cover AI technologies that generate human voices. These calls need the called party's prior express consent, unless they are for an emergency or an exemption applies.

India. TRAI's Third Amendment counts calls using "pre-recorded/artificial voice technologies" as A2P calls. Senders must declare them and their caller ID ranges in advance, or the calls count as unsolicited, and terminating providers may charge up to ₹0.05 per minute, with the designated commercial series exempt. Vobiz's India calling regulations page covers KYC and caller ID, and the 140 and 160 series guide covers commercial numbers.

Vobiz docs for this topic

The setup details for this are in the Vobiz docs:

Anything that places calls or sends messages is billed and reaches real people, so run it only with your own credentials after confirming the details.

Frequently asked questions

What is the full form of TTS?

TTS is short for text-to-speech, also called speech synthesis. It is software that reads written text aloud in a generated voice. On shopping sites the same letters can mean true to size, though software and telephony use only the first meaning.

What does TTS mean in a phone call or IVR?

It means the voice the caller hears is generated from text rather than recorded by a person. An IVR or voice agent sends words to an engine, and the audio plays on the call. This suits content that changes per caller, such as amounts, dates, and names.

What is the difference between TTS and STT?

They run in opposite directions. TTS turns text into speech, so it is the system's mouth. STT, also called ASR, turns speech into text, so it is the system's ears. A voice agent needs both, with a language model deciding what to say between them.

Why does TTS sound worse on a phone call than in a demo?

Demos usually play at a high sample rate through good speakers. A narrowband call carries roughly 300 Hz to 3.4 kHz, sampled at 8 kHz, then compressed by G.711. High consonants such as s lose energy. Judge a voice only after hearing it on a real call.

Can TTS speak languages other than English on calls?

Yes. Most engines publish voices in many languages, including Hindi and other Indian languages; check each engine's current list. Test numbers, names, and mixed-language sentences, because those fail first on real scripts.

Are TTS calls regulated?

Yes, in most markets. In the US, the TCPA's limits on artificial or prerecorded voice cover AI-generated voices, so these calls need prior express consent. In India, TRAI's Third Amendment of September 2026 defines A2P calls to include artificial voice, and senders must declare them and their caller ID ranges in advance. Confirm the start date and take legal advice.

Further reading on Vobiz

Sources

Standards and research

Regulators

Engine documentation (checked November 3, 2026)

Industry and reference


If your TTS voice belongs to a voice AI agent on a phone line, Vobiz provides the telephony, with integrations for ElevenLabs, Cartesia and Pipecat.

Top comments (0)