Text to speech for Indian languages: open voices and checks

Text to speech for Indian languages no longer needs a paid API: AI4Bharat's Indic Parler-TTS (Apache 2.0, 20 Indian languages plus English) and IndicF5 (MIT, 11 languages) both speak Tamil and Hindi and allow commercial use on your own server. Meta's MMS voices cover more languages but forbid commercial use. Whichever you choose, test names, numbers and dates with native speakers before a caller hears them.

These notes are from our voice AI lab and follow our guide to self hosted text to speech, which compares licences for English and Arabic voices too. Here we look only at Indian languages. For the basics, see text to speech in our glossary.

Which open models speak Tamil and Hindi?

Indic Parler-TTS has 0.9 billion parameters. You pick a named speaker and describe the delivery in plain words: pitch, pace, expressiveness, background noise. For Tamil the card recommends the speaker Jaya; for Hindi, Rohit or Divya. Adding "very clear audio" to the description improves quality.

IndicF5 copies the voice in a short reference clip, so you need the clip and the exact text spoken in it. That makes it the choice when you want one consistent brand voice, and it is also why its consent terms matter (see below).

Indic-TTS is AI4Bharat's earlier work, published at ICASSP 2023: separate FastPitch and HiFi-GAN models per language. They are smaller and simpler to run, and include Rajasthani, which neither model above lists.

How good are they, by published scores?

Few models publish numbers per Indian language. Indic Parler-TTS does, using a native speaker score, a listening test similar to a mean opinion score:

Two lessons. Tamil scores well below Hindi, so a model that sounds good in Hindi needs its own Tamil test. And fine-tuning helped Tamil far more than Hindi, which is a reason to keep a fine-tuning option open for the languages your customers use.

The Indic-TTS paper reports that its models significantly improve on existing ones in all its languages, measured by mean opinion score, but its summary gives no figures. We have not run our own listening tests yet, so we quote only these.

What should you check before callers hear it?

Most complaints about Indian language voices are not about the voice. They are about what it reads wrongly.

1. Numbers and money. Does ₹1,250 come out as an amount in the right language? Write numbers out in words before synthesis if the model stumbles.
2. Dates and times. "09/12" means 9 December in India. Expand dates to words yourself rather than trusting the model.
3. Phone numbers. Read in groups, digit by digit, not as one large number.
4. Names and places. Patient names, doctors, street names and your brand. Keep a pronunciation list and test every entry.
5. English inside Tamil or Hindi. "Appointment confirmed" in a Tamil sentence is normal speech. Check whether the model reads the English words naturally or spells them out.
6. The phone line. Play the output through a real call. Telephony audio is narrowband, and a voice that sounds rich on headphones can sound thin on a call.

Ask three or four native speakers per language to rate a fixed set of 30 sentences built from your real messages. Repeat the test whenever you change model or speaker.

Where is it used?

• IVR menus. Prompts rarely change, so generate them once, have a native speaker approve each file, and play recordings. Speed does not matter.
• WhatsApp voice notes. A reminder or order update read aloud in Tamil or Hindi. A few seconds to generate is fine.
• Appointment calls. Live, so speed matters: a caller expects a reply within a second or two. Stream the reply sentence by sentence. Our guide to voice AI appointment booking covers the full call.

Consent and voice cloning

IndicF5's terms say to "only clone voices for which you have explicit permission" and that unauthorised cloning is strictly prohibited. Treat that as the minimum for any model:

• Get written consent from the person whose voice you clone, stating where it will be used, and let them withdraw it.
• Do not clone a staff member's voice because their recordings happen to exist.
• Never imitate a doctor, official or public figure.
• Tell callers they are speaking to an automated assistant, and transfer anyone who asks for a person.

Our self hosted text to speech post covers the US rule on AI voices in calls. Check the rules in each country where you call.

Self host or use an API?

Hosted APIs exist too. Sarvam's Bulbul, for example, covers 10 Indian languages plus Indian English, takes up to 2,500 characters per request on its current models, and can return 8 kHz audio for telephony.

Self host when the text contains personal details such as names, appointments or amounts that should stay on your network, or when volume is high. Our voice AI lab will rate open and hosted voices on how they say Indian names and terms, and our voice AI agents combine speech to text, a language model and one of these voices. Book a consultation.