IndicVoices (2024)

In one paragraph

IndicVoices is a speech dataset of natural, spontaneous speech from 16 237 speakers in 145 districts of India, covering all 22 languages in the Eighth Schedule of the Constitution. Most of the audio is extempore speech in response to prompts, with smaller conversational and read portions. Of 7 348 hours collected, 1 639 hours were transcribed, a median of 73 hours per language. The paper is as much about process as data: it publishes the protocols, prompt libraries, conversation scenarios, quality-control steps, transcription guidelines and tools used, so that other groups can collect comparable data. Using the dataset, the authors train IndicASR, which they describe as the first ASR system covering all 22 scheduled languages. The data, tools and models are released openly.

Key idea

Inclusive, spontaneous data at scale, collected on purpose. Existing Indian-language corpora were mostly read speech from a few cities, which is not what a recogniser meets in production. IndicVoices samples by district, gender, age and education; uses prompts and scenarios designed per language; records on ordinary phones; and standardises transcription conventions (script, numerals, code-switched English words, disfluencies) across all 22 languages. The transcription guidelines are themselves a contribution, because for Hindi and its neighbours the biggest source of WER disagreement is convention, not acoustics. The ASR models are Conformer-family encoders trained in NeMo on this data; the released checkpoints on the Hugging Face hub are branded IndicConformer and come in per-language and multilingual variants, the latter carrying both CTC and RNN-T heads.

Why it matters for vaak

The indic-conformer:<lang> backend loads these checkpoints. They are the strongest open-weight option for Hindi spontaneous speech in vaak’s benchmark, especially measured by CER, and they decode with CTC so their RTF is low. Because the training transcripts follow the IndicVoices conventions (Devanagari for code-switched English, spelled-out or digit numerals per the guidelines), indic-conformer output differs systematically from Whisper’s, and part of the WER gap between them on any dataset is convention. That is why vaak’s benchmark normalises both sides identically and reports CER. The tokenization page discusses these conventions; the tiny-model page uses IndicVoices’ transcribed hours as the yardstick for what training a real Hindi model requires. The dataset is also a candidate evaluation set for vaak bench beyond the FLEURS, Common Voice and Kathbath sets shipped today.

Numbers to remember

• 7 348 hours of audio; 1 639 hours transcribed; median 73 transcribed hours per language.
• 16 237 speakers, 145 districts, 22 languages.
• Composition: 74 % extempore, 17 % conversational, 9 % read speech.

Citation

Tahir Javed, Janki Atul Nawale, Eldho Ittan George, Sakshi Joshi, Kaushal Santosh Bhogale, Deovrat Mehendale, Ishvinder Virender Sethi, Aparna Ananthanarayanan, Hafsah Faquih, Pratiti Palit, Sneha Ravishankar, Saranya Sukumaran, Tripura Panchagnula, Sunjay Murali, Kunal Sharad Gandhi, Ambujavalli R, Manickam K M, C Venkata Vaijayanthi, Krishnan Srinivasa Raghavan Karunganni, Pratyush Kumar, Mitesh M Khapra. “IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languages.” AI4Bharat, 2024. arXiv:2403.01926. https://arxiv.org/abs/2403.01926