Built to answer a specific question: which short phrases exist in open Indic speech corpora, and how many are in the register a voice assistant actually speaks? TTS models trained on sentence-level read speech degrade on very short inputs because such corpora contain almost no short utterances. This is the text-side map of what is available.

For the audio, see the per-language companion sets (Hindi: hindi-short-utterance-1to3w).

Per language

Language	Code	Rows	In-register	share
Malayalam	ml	5,344	1,457	27.3%
Kannada	kn	4,965	771	15.5%
Marathi	mr	4,425	421	9.5%
Telugu	te	4,328	764	17.7%
Punjabi	pa	4,237	263	6.2%
Assamese	as	3,728	404	10.8%
Hindi	hi	3,669	371	10.1%
Tamil	ta	3,240	1,010	31.2%
Gujarati	gu	2,842	339	11.9%
Bengali	bn	2,362	568	24.0%
Total		39,140	6,368	16.3%

Columns

Column	Meaning
`language`, `lang_code`	e.g. `malayalam` / `ml`
`transcript`	the utterance, native script
`word_count`	1, 2, or 3
`duration`	seconds of the source clip (audio not included)
`style`	Rasa's own register label, carried through unchanged
`register`	`in-register` if `style` ∈ {CONV, ALEXA}, else `out-of-register`
`tag`	secondary seed-word-list heuristic — see caveat
`gender`	speaker gender of the source clip

The `register` column is the point

Raw short-utterance counts are misleading. Most short rows in these corpora are proper nouns — monuments, cities, festivals, company names — not the conversational phrases an assistant needs. style separates them:

CONV — conversational speech: ठीक है, सर। · जी नहीं। · बिलकुल!
ALEXA — voice-assistant commands: अलार्म जोड़ें · थोड़ी आवाज़ बढ़ाओ · लाइट बंद करो
PROPER NOUN — ताज महल · कुतुब मीनार · गोमती एक्सप्रेस

Malayalam and Tamil are 3–4× richer in in-register short speech than Hindi, despite similar raw counts. If you are choosing a language to start with, that matters more than total rows.

Caveat on tag. Derived from a hand-written Hindi seed word list, kept only for cross-checking. It undercounts the in-register set ~2.5× (151 vs 371 on Hindi) and cannot recognise the ALEXA register at all. It is meaningless for non-Hindi rows. Use register.

Caveat on style for non-Hindi. style values were spot-checked against actual transcripts for Hindi only. The other nine languages assume CONV/ALEXA mean the same thing there. Verify on a sample before relying on the per-language figures.

Convenience text files

text/<language>_all.txt and text/<language>_in_register.txt — deduplicated, one phrase per line, frequency-ordered. Handy for picking a recording script or building a test set without touching Parquet.

Selection criteria

Word count 1–3 after NFC normalization, stripping Unicode P*/S* categories while preserving combining marks — stripping via [^\w\s] shatters Indic conjuncts at the virama and inflates word counts roughly 3×, which silently reduced measured yield 4× during development. Plus: ≥90% of letters in the language's own script; not empty, digits-only, or single-character; source duration 0.2–4.0 s; ≤20 rows per unique phrase.

License and attribution

CC-BY-4.0, inherited from the source. Derived from AI4Bharat's Rasa:

Rasa: Building Expressive Speech Synthesis Systems for Indian Languages in Low-resource Settings. AI4Bharat. https://huggingface.co/datasets/ai4bharat/Rasa

The source is access-gated on the Hub; please accept its terms there if you use the original.

Downloads last month: -

Total file size:

2.13 MB

You need to agree to share your contact information to access this dataset

Indic Short-Utterance Transcripts (1–3 words), 10 languages