IISc Bengaluru & Google Project Vaani Phase 2: 100,000 Hours of Diverse Indic Speech Data Open-Sourced Across 80 Districts
India’s linguistic diversity received its largest open-source data foundation today as IISc Bengaluru and Google released Phase 2 of the landmark Project Vaani speech corpus.
Capturing the Spoken Soul of India
India is a land of profound linguistic complexity: beyond its 22 Eighth Schedule languages, the nation speaks over 780 distinct languages and thousands of regional dialects. Historically, commercial speech AI models were trained almost exclusively on broadcast news, formal parliamentary speeches, or audiobooks recorded in artificial studio conditions by professional voice artists. Consequently, automated speech recognition (ASR) systems failed catastrophically when confronted with colloquial village conversations, regional slang, code-switching (Hinglish, Tanglish, Kanglish), and noisy rural acoustic environments.
Today, the Indian Institute of Science (IISc) Bengaluru, in collaborative research partnership with Google and the ARTPARK (AI & Robotics Technology Park), announced the completion and open-source release of Project Vaani Phase 2. Comprising an astounding 100,000 hours of validated natural spoken audio collected across 80 rural districts, it represents the world’s largest open-access linguistic dataset for non-Western tongues.
1. Grassroots Field Methodology & Dialect Representation
Unlike crowd-sourced web campaigns that disproportionately attract urban, English-educated contributors, Project Vaani deployed dedicated field recording teams directly into rural village squares, local weekly haats (markets), agricultural fields, and panchayat halls across 16 states.
The data collection architecture prioritized true spontaneous speech:
- Natural Conversational Prompts: Rather than asking villagers to read pre-written sentences from a screen, contributors were asked to describe local festivals, share traditional recipes, explain agricultural practices, or discuss village folklore in their mother tongue.
- Acoustic Diversity: Recordings capture authentic real-world background noise—monsoon rainfall, village well activities, diesel pump engines, and cattle bells—ensuring models trained on the corpus develop robust noise-cancellation and acoustic feature invariance.
- Extensive Code-Switching: The dataset faithfully captures the natural, fluid code-mixing characteristic of daily Indian life, where speakers transition effortlessly between regional mother tongues and English loanwords within a single sentence.
2. Open Research Licensing & Model Democratization
In accordance with the IndiaAI Mission’s open-access principles, the entire 100,000-hour corpus—alongside paired phonetic transcriptions, speaker demographic metadata, and acoustic confidence masks—is released freely on OpenSLR and Hugging Face under Creative Commons research and commercial licenses.
Project Vaani Phase 2 Dataset Breakdown
| Linguistic Region / State | Districts Covered | Total Audio Hours | Distinct Dialects Represented | Primary Languages |
|---|---|---|---|---|
| Northern Belt (UP, Bihar, MP) | 24 Districts | 32,000 Hours | 18 Dialects | Bhojpuri, Maithili, Awadhi, Bundeli |
| Southern Belt (TN, KA, AP, TS, KL) | 22 Districts | 28,500 Hours | 14 Dialects | Kannada, Telugu, Tamil, Malayalam |
| Western Belt (Maharashtra, Gujarat) | 16 Districts | 20,500 Hours | 12 Dialects | Marathi, Gujarati, Konkani, Ahirani |
| Eastern & Northeast (WB, Odisha, Assam) | 18 Districts | 19,000 Hours | 16 Dialects | Bengali, Odia, Assamese, Santhali |
Fueling the Sovereign Voice AI Revolution
The availability of 100,000 hours of open-access conversational audio provides immediate rocket fuel for domestic AI labs (including Sarvam AI, Krutrim, Bhashini, and academic IIT labs). Developers can now pretrain native speech foundation models that truly understand the acoustic and cultural cadence of 1.4 billion citizens, unlocking frictionless voice-based public services and commercial mobile commerce.