Voice

Voice AI for Every Indian Language

Starting with Marathi and Hindi — fine-tuned on real speech data so machines can speak India's languages naturally. More languages to follow.

A research project by en2.ai — building voice AI for India's languages, one at a time.

Try It Live

Type any text, pick a language, and hear it instantly. Clone any voice with a reference audio clip.

Language
0/2000

Benchmark

Our best vs leading commercial Indic TTS, same evaluation harness.

Tested against 1000 custom sentences
Word Error Rate (lower is better)
Marathien2 tts-1-dvOurs
10.9% WER
CER: 1.95%
MarathiSarvam Bulbul v3Commercial
6.9% WER
CER: 1.30%
Hindien2 tts-1-dvOurs
4.8% WER
CER: 1.13%
Scored with Sarvam ASR (saaras:v3, language code <lang>-IN). WER and CER are corpus-level. Sarvam Bulbul v3 numbers are from their published benchmark.

Listen to the Best

Switch between languages to hear our best model for each.

No experiments found.

How It Works

We fine-tune an open-source text-to-speech model on real speech data for each target language. The model learns the language's phonemes and prosody from natural recordings, then generates speech via voice cloning — preserving a target speaker's voice while speaking fluently in that language.