AAII - Bodo Neural TTS Model

This model has been trained on the female voice subset of the AI4Bharat Rasa Bodo TTS dataset and is built using an ESPnet-based FastSpeech2 + Multi-band MelGAN TTS pipeline.

Model Overview

The TTS system is based on:

Acoustic model: FastSpeech2
Vocoder: Multi-band MelGAN
Toolkit: ESPnet
Language: Bodo (brx)
Training data: AI4Bharat Rasa Bodo TTS Dataset (Female Voice)
Input: Bodo Devanagari text
Output: Bodo speech waveform

The goal of this model is to provide a lightweight and usable Bodo neural TTS system for browser-based and community-focused language technology applications.

Applications Built Using This Model

Using this Bodo model, we have developed two free public-facing Bodo reading tools.

1. Bodo Text-to-Speech Web App

A browser-based Bodo TTS web application that can read Bodo Devanagari text directly from webpages, pasted text, or uploaded photos/PDFs (via on-device OCR).

Key features:

Reads Bodo Devanagari text
Supports webpage URL input
Supports direct text input
Supports Photo and PDF text extraction (OCR)
Runs speech synthesis inside the browser using WebAssembly / WebGPU
No server-side API call required for synthesis
Model is cached in the browser after first load
Can work offline after the model is loaded
Useful for accessibility, reading support, education, and Bodo digital content consumption

Acknowledgement

We express our sincere gratitude to AI4Bharat for creating and open-sourcing the Rasa Dataset. The training of this model relies on the female Bodo voice subset provided in the Rasa corpus.

Disclaimer

This is an early version of the Bodo TTS model. The model is designed to generate clear and natural Bodo speech, but pronunciation, prosody, and expressiveness may still improve with further data tuning and future model updates.

Downloads last month: 13