Blog ·

Indic Parler-TTS online: text to speech in Tamil, Hindi, Telugu and 15 more languages

AI4Bharat's Indic Parler-TTS speaks 18 languages with 69 named voices. Which voices to pick, why you should write in the native script, how to try it in your browser or phone, and how to run it locally.

Indic Parler-TTS is AI4Bharat's open text-to-speech model for Indian languages: Hindi, Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam and more, each with its own named voices. It is released under Apache-2.0, so commercial use is allowed. Running it yourself means a GPU, a Hugging Face account approved for the gated repo and a pinned Python setup; on LiveWeights you can try it straight from the browser, on your phone too.

Indic Parler-TTSAI4Bharat · Apache-2.0AI4Bharat's open text-to-speech for Indian languages: Hindi, Tamil, Telugu, Bengali, Marathi and 13 more, with named voices for each language.Try it online →

What Indic Parler-TTS is

  • About 0.9B parameters, built on Hugging Face's Parler-TTS architecture and fine-tuned by AI4Bharat (IIT Madras) on Indian speech datasets such as IndicTTS, LIMMITS and Rasa.
  • Two inputs: the text to speak, and a short description of the voice in English (who speaks, pitch, pace, expressiveness, recording quality).
  • 69 named voices across 18 languages, with a few recommended voices per language for the most natural results.
  • The language comes from the text: you don't switch languages with a setting, you write in that language.

Languages and voices you can try on LiveWeights

LanguageVoices (recommended first)
HindiRohit, Divya, Aman, Rani
TamilJaya, Kavitha
TeluguPrakash, Lalitha, Kiran
BengaliArjun, Aditi, Tapan, Rashmi, Arnav, Riya
MarathiSanjay, Sunita, Nikhil, Radha, Varun, Isha
KannadaSuresh, Anu, Chetan, Vidya
MalayalamAnjali, Harish, Anju
GujaratiYash, Neha
PunjabiDivjot, Gurpreet
OdiaManas, Debjani
AssameseAmit, Sita, Poonam, Rakesh
English (Indian voices)Thoma, Mary
Nepali, Sanskrit, DogriAmrita, Aryan, Karan
Bodo, Manipuri, ChhattisgarhiBikram, Maya, Kalpana · Laishram, Ranjit · Bhanu, Champa

Write in the language's own script: Devanagari for Hindi and Marathi, Tamil script for Tamil, Telugu script for Telugu, and so on. Text typed in English letters (Hinglish or Tanglish) is read badly, because the model learned each language from its native script. Your phone keyboard can switch to the script, or use voice typing in that language.

How to try it online

  1. Open the Indic Parler-TTS page and sign in, then pick a small credit pack (no subscription).
  2. Write up to 600 characters in your language, or start from the Tamil, Hindi or Telugu example.
  3. Pick the language and a named voice (or keep the recommended one) and how it should sound: natural, expressive, slow and calm, or fast. You can also describe a voice in your own words, in English.
  4. Press Generate: you can listen, download it as MP3 (or WAV), share it on WhatsApp, or change the voice and try again. Long texts are spoken a few sentences at a time.

A page for your language

Each language also has its own page with the language already picked and one-tap templates (Reel voiceover, WhatsApp greeting, festival wish, notes to audio), plus Download MP3 and Share on WhatsApp under every result: Hindi, Tamil, Telugu, Bengali, Marathi, Kannada, Malayalam, Gujarati and Punjabi.

Indic Parler-TTS vs other open TTS models

For Hindi on a budget, Kokoro 82M has two Hindi voices and runs instantly on a CPU, but it does not speak Tamil, Telugu or the other languages here. Chatterbox Multilingual also speaks Hindi and can clone a voice from a short clip. For the other Indian languages, Indic Parler-TTS is the open model to try. We compared the main options in our text-to-speech roundup.

Then run it locally

The weights are on Hugging Face (ai4bharat/indic-parler-tts): the repo is gated, but access is approved automatically once you accept the terms with your account. Install the Parler-TTS code from GitHub (pip install git+https://github.com/huggingface/parler-tts.git), then load ParlerTTSForConditionalGeneration with two tokenizers, one for the text and one for the voice description. A GPU with a few GB of memory is enough; on a CPU it works but is slow.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.