Text to Speech for Indian Languages on Calls
Jaseem Abbas
- Published

Text to speech for Indian languages is software that reads written Hindi, Tamil, Telugu and other Indian-language text aloud in a human-sounding voice. A phone call changes how that voice sounds, and it changes which lines go wrong. This guide is for teams that put a voice agent on a phone line. It explains what the line does to a voice, what breaks in Indian languages, and a 20-line test you can run before a voice takes real calls.
Here is a call to picture. A two-wheeler dealership in Madurai has its voice agent phone a customer in Tamil to remind her about a scooter service. The line in the agent's script says "service on 18/11, estimate Rs 12,499". On a laptop, in a quiet room, the demo voice sounded fine. On the call, the voice reads the date digit by digit, stumbles on the amount, and the customer asks what was said. The dealership, the customer and the figures are made up for this example.
What is text to speech for Indian languages?
Text to speech for Indian languages turns written text in scripts such as Devanagari, Tamil and Telugu into spoken audio, and it must also read Indian numbers and mixed Hindi-English sentences correctly. The basic idea is the same as in any language, and our glossary entry on text to speech covers how it works in two steps.
On a voice agent, TTS is the last step of a loop. The agent hears the caller with speech recognition, a large language model writes the reply as text, and text to speech says it. Speech recognition listens, TTS speaks.
Searches for "Hindi TTS" or "Tamil text to speech" are mostly people who want an audio file for a video or a course. This guide is for a different reader: someone whose voice will talk to a customer on a phone call, where the listener cannot pause, replay or read along.
Why does a voice sound different on a phone call?
A phone line carries much less sound than a speaker or headphones, so a voice that sounds rich in a demo can sound thin on a call. The classic phone codec, ITU-T G.711, recommends a sampling rate of 8,000 samples per second (ITU-T G.711, read 15 November 2026). A signal sampled that way cannot hold any sound above 4 kHz.
That is the cut. Voices made for headphones are usually listened to with the high end intact, and the line removes part of it. Our guide to Hindi speech recognition on calls shows the same band from the listening side: it is the reason a computer also finds phone speech harder to hear.
What does that mean in practice? As an illustration, not a measured finding: lines that depend on small sounds, such as the end of a long number or the sounds that tell two similar words apart, have less to go on once the high end is cut. So judge a voice on a phone line, not in a browser tab.
Speed matters as well. A voice that is slow to start also sounds broken on a call, because the caller hears silence and thinks the line is dead. Our guide to voice agent latency covers that side.
What breaks most often in Indian-language text to speech?
Numbers and dates, names, mixed-language sentences and the script the text is written in cause most of the odd-sounding lines. This list comes from how the problem shows up on calls, and it is our own checklist, not a standard.
What to listen for: Numbers, amounts and dates
- What goes wrong
- Digits read one by one, or a date read the wrong way round
- What to check
- Does Rs 12,499 sound like an Indian amount? Is 18/11 said as 18 November?
What to listen for: Names and places
- What goes wrong
- A long place name or a surname is split in the wrong places
- What to check
- Thiruvananthapuram, Visakhapatnam, a customer's surname
What to listen for: Mixed-language sentences
- What goes wrong
- An English word is spelled out or said in the wrong accent
- What to check
- Words such as service, estimate or EMI inside a Hindi or Tamil sentence
What to listen for: Script
- What goes wrong
- The same sentence sounds different in native script and in Roman letters
- What to check
- Send the same line both ways and compare
Numbers, amounts and dates
This is the first place to look, because a wrong digit is a wrong fact. We gave a language model the dealership's line and asked for it in Hindi. In that one run the Hindi came back as "18 नवंबर" and "12499 रुपये", with the digits left as digits. A voice then has to guess whether 12499 is "twelve thousand four hundred ninety-nine" or twelve-four-nine-nine. The same request, told to write everything as words, gave "अठारह नवंबर" (athaarah navambar, eighteenth of November) and "बारह हज़ार चार सौ निन्यानवे रुपये" (baarah hazaar chaar sau ninyaanve rupaye, Rs 12,499). That was one run, not a rule, but it shows where the risk sits.
The glossary has the plain case for dates: 12/11 can mean 12 November or 11 December, and the voice should not have to guess. How a voice sounds on a number comes down to its prosody: pitch, rhythm and pauses.
Names and places
Names are hard in every language, and Indian names are long and varied. The voice has no dictionary entry for a customer's surname. Listen for a name that is split in the wrong place, or stressed on the wrong syllable. A caller notices a mangled name faster than a mangled sentence, because it is their own.
Mixed-language sentences
Indian callers and agents mix languages in one sentence. The Hindi line from the same run reads "सर, आपके स्कूटर की service का time अठारह नवंबर सुबह दस तीस बजे का रखा है", with "service" and "time" in English letters inside a Devanagari sentence. A good voice switches smoothly in the middle of the sentence. A poor one slows down, changes accent or spells the English word out. Our guides on Hinglish voice agents and on code-switching explain why this is normal speech, not a mistake.
Script
The same sentence can be sent in native script or in Roman letters, and the two can sound different. Test both for each language. For Hindi, see our page on Hindi voice agents. For Tamil and Telugu, here are the two lines the same run gave, written for the ear and kept in script:
- Tamil: "உங்க ஸ்கூட்டர் சர்வீஸ் நவம்பர் பதினெட்டாம் தேதி காலை பத்தரை மணிக்கு இருக்கு" (your scooter service is on the eighteenth of November at half past ten in the morning). Note the date and time are words, not digits.
- Telugu: "మీ స్కూటర్ సర్వీస్ నవంబర్ పద్దెనిమిదవ తేదీ ఉదయం పదిన్నరకి ఉంది" (the same sentence in Telugu).
Lines like these were drafted with a language model. Have a native speaker check any line before your agent says it to a customer. See our pages on Tamil and Telugu voice agents.
How are text to speech voices scored?
Voices are usually rated with a mean opinion score (MOS), a listening test in which people score speech on a scale of 1 to 5. The method is standardised in ITU-T P.800, "Methods for subjective determination of transmission quality" (ITU-T P.800, read 15 November 2026).
Researchers use MOS for Indian languages too. Kumar et al. (ICASSP 2023) note that advances in neural TTS "have not been thoroughly investigated for Indian language speech synthesis". They compare acoustic models and vocoders for Dravidian and Indo-Aryan languages, train models for 13 Indian languages, and rate them by mean opinion score (arXiv 2211.09536, abstract read 15 November 2026). That is research, and it says nothing about any product.
Two cautions when you read a score. First, MOS is usually measured on clean audio, so a high score does not say how the voice sounds after the phone line. Second, a score for one language does not carry over to another. Treat a score as a starting point, then run your own test.
How do you fix a mispronounced line?
Change what the voice is given: write numbers, dates and names the way they should be said, and test again on a call. Most odd lines are fixed in the text, not in the voice.
- Write amounts and dates as spoken. "Rs 12,499" becomes "बारह हज़ार चार सौ निन्यानवे रुपये" in the line the voice reads. Show the digits on screen or in a message, and say the words on the call.
- Keep sentences short. One idea per sentence. A long sentence with a name, a date and an amount gives the voice three chances to slip.
- Say the key fact twice. State the date and time once in the middle and once at the end of the call, so a caller who missed it on the first pass gets another.
- Fix names by spelling them as said. If a name is wrong, write it the way it sounds in that language's script and retest.
- Use markup when your engine offers it. Many TTS engines accept SSML, the W3C markup for speech. Its
phonemeelement gives a phonetic pronunciation, and itssay-aselement marks text as a date or a number (W3C SSML 1.1, read 15 November 2026). Without aphonemeelement, the processor decides how to pronounce the word on its own.
How do you test a voice before it takes real calls?
Play 20 real lines over a real phone call, and have two native speakers mark each one clear, odd or wrong. This is our own method, not a standard. To compare voices side by side, read how to compare TTS voices.
- Write 20 lines your agent will really say. Include 5 with numbers or dates, 5 with names, and 5 that mix two languages. The other 5 are plain sentences.
- Call from a real phone. Use a mobile on a normal network, and try one from a noisy place. A browser tab uses wider audio than a phone line, as the section above explains.
- Mark each line. Two native speakers listen without the script and mark each line: clear, odd or wrong. Odd means they understood it but it sounded off.
- Count the result by group. "4 of 5 amount lines clear" tells you more than one overall score.
- Fix the text, not only the voice. Rewrite the "odd" and "wrong" lines as spoken, then run the same 20 again.
Our guide on how to test a voice agent shows how to test numbers, dates and amounts across a whole call.
Should every language use the same voice?
No. Pick a voice for each language, because a voice that is good in Hindi may be weaker in Odia or Malayalam. Test every language you will use on calls, with its own 20 lines.
Callers expect to be spoken to in their own language. Our pages for each language show the numbers and the call scenarios: Hindi, Tamil and Telugu are three to start with. The full list is on Languages.
How does Bananaflow fit?
Bananaflow is an AI voice agent platform. These are the parts that touch the voice:
- English plus 10 Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, each heard and spoken in its own script. See Languages.
- Following the caller's language, and switching in the middle of a call when the caller does.
- Hinglish written down as spoken, so a mixed sentence is written the way people say it.
- A voice for each agent, which you can change without rebuilding the agent.
- Pre-recorded audio mixed with generated speech, and template variables for names and amounts.
- Testing before real calls: call any agent by voice in your browser, then place the 20-line test on a real phone.
- Interruptions (barge-in) set per step, so the caller can cut in.
Run the 20-line test with the lines your callers will hear, in every language you switch on.
What should you remember about text to speech on calls?
- Text to speech for Indian languages reads each script, Indian numbers and mixed sentences aloud. On a phone call the voice reaches the caller with less sound than in a demo.
- The phone codec samples 8,000 times a second, so nothing above 4 kHz survives.
- Listen for four things: numbers and dates, names, mixed-language sentences and script.
- Fix pronunciation in the text first: amounts and dates written as spoken, short sentences, the key fact said twice.
- A voice score from a studio does not predict a call. Test 20 real lines on a real phone, per language, with two native speakers.
FAQ
Frequently asked questions
Which is the best text to speech for Hindi?
There is no single best one, so test on a call. Play 20 real lines over a real phone call, including 5 with numbers, 5 with names and 5 that mix Hindi and English, and have two native speakers mark each line clear, odd or wrong. The voice that scores best on your lines, on a phone line, is the best one for you.
Does text to speech work for Tamil and Telugu?
Yes. Research teams have trained speech models for Dravidian languages such as Tamil and Telugu, and tools exist for both. What varies is quality on your own lines, so test each language with its own 20 lines and a native speaker, and write dates and amounts the way they should be said.
Why does a text to speech voice sound worse on a phone call?
Because the phone line carries less sound. The classic phone codec, ITU-T G.711, samples audio 8,000 times a second, so nothing above 4 kHz survives. A voice that sounds full in headphones loses part of its high end on a call, and a slow start can make it sound broken as well.
Can text to speech read Hinglish?
It can, and it is a common case on Indian calls. A sentence such as a Hindi line with the English words service and time inside it needs a voice that switches in the middle of the sentence without slowing down or spelling the English word out. Test mixed lines on their own, because they are the ones that expose a weak voice.
How do I make a text to speech voice say numbers in the Indian way?
Give the voice the words, not the digits. Write 12,499 as twelve thousand four hundred ninety-nine in the language of the call, and write dates as the day and the month name. Many engines also accept SSML, the W3C speech markup, whose say-as element marks text as a number or a date. Then retest on a call.
What is a good MOS score for text to speech?
MOS stands for mean opinion score, a listening test where people rate speech from 1 to 5, standardised in ITU-T P.800. We give no target number, because scores depend on the test, the language and the audio. A score from clean audio also does not tell you how a voice sounds on a phone call.
Which Indian languages can a Bananaflow voice agent speak?
English plus 10 Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, each heard and spoken in its own script. The agent follows the caller's language and can switch in the middle of a call.




