Rasa: Towards Building an Expressive Multilingual Text-To-Speech Dataset for Indian Languages

Funded by: Bhashini, Ministry of Electronics and Information Technology, Government of India
Supported by: EkStep Foundation and Nilekani Philanthropies

Overview

We introduce Rasa, the first high-quality multilingual expressive Text-to-Speech (TTS) dataset for any Indian language. It comprises a minimum of 20 hours per speaker with a target of covering a female and male voice for each of the 22 officially recognized languages of India. In our initial version, we explore a practical recipe for collecting high-quality data for resource-constrained languages, prioritizing easily obtainable neutral data alongside smaller amounts of expressive data. This approach enables us to extend our dataset to encompass a diverse array of speaking styles and contexts. These include neutral readings from Wikipedia and IndicTTS texts, expressive speech capturing the six Ekman emotions (happy, sad, angry, fear, disgust, and surprise), as well as command-based interactions from platforms like Alexa, BigBasket, UMANG, and DigiPay. Additionally, Rasa includes natural conversations on various topics, news-reading, and narration from book readings. Currently, we release the data for 8 speaker-language pairs. Through this release, we aim to provide a valuable resource for developing expressive TTS models in multilingual settings for the officially recognized languages of India.

Key Features

• Multilingual Coverage: Covers diverse Indian languages
• Expressive Speech: Includes Ekman emotions (happy, sad, angry, fear, disgust, and surprise)
• Multiple Speaking Styles:
  • Neutral speech from Wikipedia texts
  • Command-based interactions from Alexa, BigBasket, UMANG, and DigiPay
  • Natural conversations on various topics
  • News reading and narration from book readings
• High-Quality Data: 48 KHz, Mono
• Current Release: 27 speaker-language pairs available now

Through this release, we aim to provide a valuable resource for multilingual expressive TTS models, helping advance text-to-speech synthesis for Indian languages.

---

Dataset Statistics

Language	Speaker	Hours	Utterances
Assamese	Female	29.09	15,085
Assamese	Male	28.44	16,282
Bengali	Female	29.64	15,570
Bengali	Male	25.89	15,055
Bodo	Female	27.32	16,329
Bodo	Male	24.99	13,163
Dogri	Female	25.69	13,178
Dogri	Male	21.13	9,856
Gujarati	Female	23.66	11,324
Kannada	Female	27.02	14,915
Kannada	Male	27.60	16,002
Konkani	Female	26.33	17,585
Maithili	Male	29.34	12,918
Malayalam	Female	26.42	16,974
Malayalam	Male	25.22	16,878
Marathi	Female	28.81	15,473
Marathi	Male	26.55	14,483
Nepali	Female	28.74	16,016
Nepali	Male	24.13	14,426
Odia	Female	24.06	11,756
Punjabi	Female	26.72	13,422
Punjabi	Male	29.12	15,620
Sanskrit	Female	25.27	12,757
Sanskrit	Male	26.20	11,002
Tamil	Female	29.94	19,871
Telugu	Female	27.29	15,406
Telugu	Male	24.98	15,007
Total		739.47	408,019

---

License

CC-BY-4.0

Citation

If you use this dataset, please cite:

@inproceedings{ai4bharat2024rasa,
  author={Praveen Srinivasa Varadhan and Ashwin Sankar and Giri Raju and Mitesh M. Khapra},
  title={{Rasa: Building Expressive Speech Synthesis Systems for Indian Languages in Low-resource Settings}},
  year=2024,
  booktitle={Proc. INTERSPEECH 2024},
}