ai4bharat/indicvoices_r

IndicVoices-R: Multilingual, Multi-Speaker Speech Corpus for Indian TTS

Dataset Summary

IndicVoices-R (IV-R) is the largest multilingual Indian text-to-speech (TTS) dataset derived from an automatic speech recognition (ASR) dataset. It contains 1,704 hours of high-quality speech from 10,496 speakers across 22 Indian languages. This dataset is designed to enhance the development of robust Indian TTS models by providing diverse speaker demographics, natural conversational speech, and high-quality audio samples.

Key Features

• Comprehensive Coverage: Includes all 22 scheduled Indian languages, ranging from 9 to 175 hours per language.
• Speaker Diversity: 10,496 speakers, ensuring rich demographic and linguistic variation.
• High-Quality Samples: Speech quality comparable to gold-standard TTS datasets (LJSpeech, LibriTTS, IndicTTS).
• Natural Speech Recordings: 93.25% extempore speech, leading to more expressive and natural-sounding synthesis.
• Zero-shot, Few-shot, and Many-shot Benchmarking: Includes a benchmark dataset for evaluating speaker generalization capabilities of TTS models.

Data Processing Pipeline

The dataset was created by restoring and enhancing ASR-quality speech using:

1. Demixing: HTDemucs model to separate speech from background noise.
2. Dereverberation: VoiceFixer to reduce reverb and enhance speech clarity.
3. Speech Enhancement: DeepFilterNet3 to remove remaining artifacts.
4. Filtering: Samples filtered based on speech clarity, SNR, pitch variation, and speaking rate.
5. Normalization: Volume adjusted using PyDub for consistency.

Dataset Format

Each entry in the dataset includes:

• Audio file: .wav format, 48 kHz sampling rate.
• Text transcript: Available in both verbatim and normalized formats.
• Metadata JSON: Includes speaker ID, gender, age group, duration, and language.

Benchmarks & Evaluation

IndicVoices-R includes a benchmark suite to evaluate zero-shot, few-shot, and many-shot TTS model performance across different demographics and speech styles.

Key Evaluation Metrics:

• NORESQA-MOS: Measures naturalness of synthesized speech.
• SNR & C50: Assess speech clarity and reverberation levels.
• Speaker Similarity (S-SIM): Evaluates zero-shot speaker generalization.

Usage & Applications

IndicVoices-R can be used for:

• Training multilingual TTS models with high speaker diversity.
• Improving zero-shot speaker generalization for Indian languages.
• Building expressive and natural-sounding synthetic voices.
• Evaluating TTS performance with a standardized benchmark.

License

CC-BY-4.0

Acknowledgements

This project was supported by Digital India Bhashini, the Ministry of Electronics and Information Technology (Government of India), EkStep Foundation, and Nilekani Philanthropies. Special thanks to CDAC Pune for access to the PARAM-Siddhi supercomputer, enabling our speech enhancement pipeline and model training. We also appreciate the unwavering support from the AI4Bharat team.

Citation

If you use IndicVoices-R, please cite the following paper:

@inproceedings{ai4bharat2024indicvoices_r,
  author       = {Ashwin Sankar and
                  Srija Anand and
                  Praveen Srinivasa Varadhan and
                  Sherry Thomas and
                  Mehak Singal and
                  Shridhar Kumar and
                  Deovrat Mehendale and
                  Aditi Krishana and
                  Giri Raju and
                  Mitesh M. Khapra},
  editor       = {Amir Globersons and
                  Lester Mackey and
                  Danielle Belgrave and
                  Angela Fan and
                  Ulrich Paquet and
                  Jakub M. Tomczak and
                  Cheng Zhang},
  title        = {IndicVoices-R: Unlocking a Massive Multilingual Multi-speaker Speech
                  Corpus for Scaling Indian {TTS}},
  booktitle    = {Advances in Neural Information Processing Systems 38: Annual Conference
                  on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver,
                  BC, Canada, December 10 - 15, 2024},
  year         = {2024},
  url          = {http://papers.nips.cc/paper\_files/paper/2024/hash/7dfcaf4512bbf2a807a783b90afb6c09-Abstract-Datasets\_and\_Benchmarks\_Track.html},
}