DviTTS: 1-click text-to-speech for English + 6 Indian languages ๐ฎ๐ณ๐
Text-to-speech for English, Hindi, Marathi, Bengali, Punjabi, Gujarati and Odia, with two models in one web UI:
๐๏ธ DviTTS (NVIDIA GPU, 6 GB+)
- Clone a voice from a 3-10 s clip (upload or record it, then type its exact transcript)
- Or use one of 18 preset voices
- Splits long text at sentence ends
โก DviTTS-Nano (runs on any CPU)
- 110M-parameter ONNX model, so no GPU is needed
- 18 preset voices
- On x86 Windows/Linux, stage 1 runs on llama.cpp, about 1.3x faster than ONNX Runtime alone
Both models
- 24 kHz output with an inaudible AudioSeal watermark
- Numbers, dates and money are spelled out before speaking
- Detects the language from the script (choose Marathi by hand for Devanagari Marathi text, since Devanagari is read
as Hindi)
- The same seed gives the same audio
- WAVs are saved to app/outputs
- Gradio API (/generate_dvitts, /generate_nano) with curl, Python and JS examples in the README
Voices: English (emma, grace, james, oliver), Hindi (ananya, kavya, arjun, rohan), Marathi (sneha, aditya), Bengali
(ishita, arnab), Punjabi (jasleen, gurpreet), Gujarati (hetal, kunal), Odia (subhashree, debasis)
Install: Click Install, then Start. It sets up Python 3.12, PyTorch 2.8 (the CUDA 12.8 build on NVIDIA, so RTX
50-series works) and the model weights (561 MB for Nano, about 4.8 GB more on NVIDIA). Reset keeps the downloaded
weights.
๐ฆ Launcher: https://github.com/PierrunoYT/DviTTS-Pinokio
๐ค Models: dvitts (https://huggingface.co/prakharkhatri/dvitts) ยท dvitts-nano
(https://huggingface.co/prakharkhatri/dvitts-nano). Apache 2.0, launcher MIT.
โ ๏ธ Only clone voices you have permission to use. Don't use it for impersonation, and label published output as
AI-generated where your local rules require it.
DviTTS: 1-click text-to-speech for English + 6 Indian languages ๐ฎ๐ณ๐
Text-to-speech for English, Hindi, Marathi, Bengali, Punjabi, Gujarati and Odia, with two models in one web UI:
๐๏ธ DviTTS (NVIDIA GPU, 6 GB+)
โก DviTTS-Nano (runs on any CPU)
Both models
as Hindi)
Voices: English (emma, grace, james, oliver), Hindi (ananya, kavya, arjun, rohan), Marathi (sneha, aditya), Bengali
(ishita, arnab), Punjabi (jasleen, gurpreet), Gujarati (hetal, kunal), Odia (subhashree, debasis)
Install: Click Install, then Start. It sets up Python 3.12, PyTorch 2.8 (the CUDA 12.8 build on NVIDIA, so RTX
50-series works) and the model weights (561 MB for Nano, about 4.8 GB more on NVIDIA). Reset keeps the downloaded
weights.
๐ฆ Launcher: https://github.com/PierrunoYT/DviTTS-Pinokio
๐ค Models: dvitts (https://huggingface.co/prakharkhatri/dvitts) ยท dvitts-nano
(https://huggingface.co/prakharkhatri/dvitts-nano). Apache 2.0, launcher MIT.
โ ๏ธ Only clone voices you have permission to use. Don't use it for impersonation, and label published output as
AI-generated where your local rules require it.