Rasa: Expressive Multilingual TTS with 20 Hours per Speaker for Indian Languages

Rasa provides at least 20 hours of audio per speaker for expressive Text-to-Speech (TTS) across multiple Indian languages. Created by ai4bharat and funded by Bhashini (Ministry of Electronics and Information Technology, India), the collection targets male and female voices for high-quality synthesis. The dataset includes recordings for over 13 languages including Bengali, Malayalam, and Sanskrit.

Use Cases

• Training expressive TTS models using audio and text pairs for languages like Malayalam (ml)
• Developing prosody-aware speech synthesis systems for Bengali (bn)
• Fine-tuning multilingual speech models for low-resource languages such as Sanskrit (sa)

Strengths

• Minimum of 20 hours of audio per speaker
• Covers 13+ Indian languages including low-resource ones like Kashmiri (ks) and Sanskrit (sa)
• Includes both male and female voice representations for gender diversity

Limitations

• Specific expressive labels or emotion categories are not explicitly defined in the provided metadata
• Total number of unique speakers per language is not specified beyond the male/female target

Provenance

Source

ai4bharat and Bhashini (Ministry of Electronics and Information Technology, Government of India)

Freshness

Last updated December 2025.

Geography

India