Chaashini

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: 1,032,690 clips · 2208.36 hours · 33 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: 2026-09-18 19:10 IST

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

Languages

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

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}
}