Chaashini (चाशनी)

Chaashini — Hindi/Urdu for sugar syrup — is a continuously growing corpus of clean, single-speaker, studio-grade Indian-language speech built for training speech models (text-to-speech, speech recognition, speech language models). Every clip in the corpus has passed a strict multi-stage quality gate; the aim is purity over volume.

• Total: 42,593 clips · 85.35 hours · 19 languages
• Format: mono 24 kHz FLAC (audio column) with a verbatim transcript and rich per-clip metadata
• Clip length: 0.5 s – 30 s, cut at natural pauses
• Last refreshed: 2026-09-05 03:34 IST

The corpus grows automatically: new shards are appended every ~2 hours of newly accepted audio.

Languages

code	language	clips	hours	avg clip	avg quality (OVRL)
hi	Hindi	21,124	40.11	6.8 s	3.24
en	English	5,807	12.82	7.9 s	3.23
te	Telugu	2,939	6.47	7.9 s	3.24
ml	Malayalam	2,031	4.36	7.7 s	3.22
mr	Marathi	2,360	3.94	6.0 s	3.18
bn	Bengali	1,776	3.41	6.9 s	3.25
ne	Nepali	828	2.65	11.5 s	3.31
ta	Tamil	974	2.42	8.9 s	3.22
pa	Punjabi	1,489	2.29	5.5 s	3.21
or	Odia	624	1.95	11.3 s	3.27
as	Assamese	664	1.46	7.9 s	3.23
sat	Santali	833	1.02	4.4 s	3.15
kn	Kannada	342	0.82	8.7 s	3.22
mni	Manipuri	298	0.76	9.1 s	3.23
kok	Konkani	305	0.47	5.5 s	3.27
gu	Gujarati	126	0.28	8.0 s	3.22
brx	Bodo	44	0.07	5.8 s	3.25
mai	Maithili	28	0.06	7.8 s	3.35
lus	Mizo	1	0.00	4.6 s	3.16

What makes a clip "pristine"

Audio is sourced from publicly available spoken-word recordings (talks, interviews, narration, lectures, podcasts and similar long-form speech). Each recording then passes through:

1. Source-level screening – recordings dominated by music, singing, or non-speech content are discarded whole.
2. Speech activity detection – frame-accurate speech/non-speech decisions; clips are cut only at genuine pauses.
3. Speaker purity – a neural diarizer labels speakers; every clip contains exactly one speaker and any region where voices overlap (plus a safety margin) is removed. Speaker labels are consistent within a source.
4. Perceptual quality scoring – each clip receives non-intrusive MOS-style scores for signal quality, background intrusiveness and overall quality, plus tagger probabilities for music / singing / noise, an SNR estimate, loudness, clipping and effective bandwidth. Clips outside strict thresholds are rejected; borderline clips are passed through a speech enhancer and re-scored (never accepted blindly) — such clips are flagged enhanced=true.
5. Transcription – a multilingual Indic ASR system transcribes each accepted clip; clips with implausible character rates (empty, hallucinated or clipped transcripts) or low recogniser confidence are rejected.
6. Language identification – script-aware identification on the transcript gives the language, a confidence, and the full language composition of the clip (code-mixing is common in Indian speech and is preserved, not filtered; language is the dominant language and language_mix holds the shares).

Schema

column	type	description
id	string	unique clip id
audio	Audio	mono 24 kHz FLAC
text	string	transcript (native script; borrowed English words may appear in Latin script)
language	string	dominant language (ISO 639 code)
language_name	string	human-readable language name
language_confidence	float	confidence of the language decision, 0–1
language_mix	string (JSON)	share of each language/script in the clip, e.g. {"hi": 0.82, "en": 0.18}
script	string	dominant writing system of the transcript
code_mixed	bool	true if a secondary language exceeds 15 % of the tokens
duration_s	float	clip duration in seconds
sample_rate	int	24000
speaker_id	string	speaker label, consistent within a source recording
source_id	string	opaque, stable id of the source recording (for grouping / de-duplication)
segment_index	int	order of the clip within its source
enhanced	bool	clip was passed through speech enhancement before re-scoring
dnsmos_sig / dnsmos_bak / dnsmos_ovrl / dnsmos_p808	float	perceptual quality scores (1–5): signal, background, overall, P.808 MOS
music_prob / speech_prob / noise_prob	float	tagger probabilities (0–1)
snr_db	float	estimated signal-to-noise ratio
rms_dbfs / peak_dbfs	float	loudness and peak level
clipping_ratio	float	fraction of clipped samples
bandwidth_hz	float	effective audio bandwidth
vad_speech_ratio	float	fraction of the clip that is active speech
speaker_dominance	float	fraction of speaker-labelled frames belonging to the clip's speaker
chars_per_sec	float	transcript characters per second
asr_confidence	float	mean posterior of the recogniser's emitted tokens (0–1); low values flag hard audio
genre	string	coarse content genre of the source (talk, narration, interview, …)
created_at	string	ISO-8601 timestamp when the clip was accepted

Usage

from datasets import load_dataset

ds = load_dataset("kapturecx/Chaashini", "hi", split="train", streaming=True)   # one language
row = next(iter(ds))
print(row["text"], row["audio"]["sampling_rate"], row["language_mix"])

all_langs = load_dataset("kapturecx/Chaashini", "default", split="train", streaming=True)

Filtering tips: dnsmos_ovrl >= 3.2 and enhanced == False gives the most conservative subset; language_confidence >= 0.8 and code_mixed == False gives monolingual clips.

Licensing and intended use

The corpus is released under Apache-2.0 for research and commercial speech-technology development. Clips are short excerpts of publicly available spoken-word material processed for machine learning; no personal identifiers are stored beyond an opaque source id. If you believe content should be removed, open a discussion on this repository.

Citation

@misc{chaashini2026,
  title  = {Chaashini: a quality-gated multilingual Indian speech corpus},
  author = {Kapture CX},
  year   = {2026},
  url    = {https://huggingface.co/datasets/kapturecx/Chaashini}
}