Skip to content

Bananaflow MCP is live.

Read the docs

TTS full form: text to speech and how it works on calls

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published

How does text to speech work on a phone call?

Text to speech works in two steps: the system first turns the written reply into the words to be spoken, then generates audio of a voice saying them. The W3C's SSML 1.1 specification (read 15 November 2026) names text normalization as one of the steps: it converts the written form of a text into the spoken form, so "$200" may be spoken as "two hundred dollars".

The same step explains most odd-sounding calls. Picture an AI voice agent at a Pune diagnostics lab that reads out "Your appointment is on 12/11 at 4:30 PM." In India, 12/11 means 12 November, but a synthesizer has to decide that; the SSML specification itself notes that "1/2" can be spoken as "half", "January second" or "February first". An agent that writes "12 November" gives the voice nothing to guess. The lab and the sentence are made up for this example.

A phone call adds two limits that a screen reader does not have. The audio must start quickly, because a pause makes the caller think the line is dead, and the voice must suit the language the caller speaks. For the speed side, see voice agent latency.

Before a voice speaks, written shortcuts like 12/11 are turned into spoken words.

What is the difference between TTS and speech synthesis?

Speech synthesis is the wider term: the W3C's SSML specification defines it as the automatic generation of speech output from data input, which may be plain text, marked up text or binary objects. Text to speech is the case where the input is text.

In everyday use the two names mean the same thing on a call: the agent has a sentence, and a voice says it. The reverse step, which turns the caller's voice into text, is automatic speech recognition.

How does AI change text to speech?

Older systems stitched speech together from stored pieces of recorded speech, while newer AI voices generate the audio from the text. A generated voice has no fixed set of pieces, so it can say any sentence a language model writes, at the pace and tone the sentence calls for. Many AI voices end with a vocoder, the step that turns the model's description of the sound into audio. A few systems skip the separate voice step and use a speech-to-speech model, which takes in speech and answers in speech.

For a voice agent this matters in one way: the reply is new on every call. A recorded IVR menu can only play the prompts someone recorded, while an agent writes a fresh reply to each caller and its voice says it. See IVR, interactive voice response for the menu side.

On Bananaflow, you choose the voice for each agent, and the agent speaks English plus 10 Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. Pre-recorded audio can be mixed with generated speech, so a fixed opening line sounds exactly as you recorded it and the rest is spoken live. For callers who mix languages, such as Hinglish, see how voice agents handle Hinglish. Read how a Bananaflow agent speaks English and 10 Indian languages.

Why does the voice matter more on a call than in a chat?

On a call, the voice is the whole interface, so it decides whether the caller trusts and understands the agent. There is no screen to fall back on. A caller who cannot make out an amount or a date has to ask again, and every repeat costs time.

Three things shape that: how clearly the voice says numbers and names, whether it speaks the caller's language, and how soon the first word arrives. The first is mostly about how the reply is written: short sentences, full dates, amounts in words. The second is about the voice covering the language. The third is covered in voice agent latency. To test a voice in each language, read text to speech for Indian languages on calls. How a voice sounds on a number or a question comes down to its prosody: pitch, rhythm and pauses.

FAQ

Frequently asked questions

  • What is the full form of TTS?

    TTS stands for text to speech. It is software that turns written text into spoken audio, so a computer or an AI voice agent can say a reply out loud instead of showing it on a screen.

  • What is the meaning of TTS?

    TTS means the automatic generation of speech from text. You give it a sentence, and it returns audio of a voice saying that sentence. A phone agent uses it to speak each reply to the caller.

  • Is TTS the same as speech synthesis?

    Nearly. Speech synthesis is the wider term: making speech from any input. Text to speech is the common case, where the input is written text. Some synthesis systems start from phonetic symbols instead of text.

  • What is the difference between TTS and STT?

    They run in opposite directions. STT, speech to text, turns what the caller says into text. TTS, text to speech, turns the agent's written reply into audio. A voice agent uses both in every turn of a call.

  • Why does a TTS voice sometimes say a number or date wrongly?

    Written text has shortcuts, like 12/11 or a rupee sign, that must be turned into spoken words first. This step is called text normalization, and the same characters can be read several ways, so a good agent writes dates and amounts out in full.

  • Can TTS speak Hindi and other Indian languages?

    Yes, if the voice was built for the language. A Bananaflow agent speaks English plus 10 Indian languages, Hindi among them, and you choose the voice for each agent.

  • Can an agent mix recorded audio with a TTS voice?

    On Bananaflow, yes. Pre-recorded audio, such as a fixed greeting, can be mixed with generated speech in the same call, so fixed lines sound exactly as you recorded them and the rest is spoken live.

  • What is SSML?

    SSML is the Speech Synthesis Markup Language, a W3C standard for controlling how a synthesizer speaks, such as pronunciation, volume, pitch and rate. It is a markup language written around the text that is to be spoken.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot