auraCodes/indian-english-hindi-tts-60min

Indian English + Hindi TTS Dataset

A small, heavily-curated Text-to-Speech corpus: 73.2 minutes (37.7 min Indian English, 35.5 min Hindi) of single-speaker, studio-grade clips. Every clip's audio was listened to and its transcript corrected against automated Sarvam ASR output; resulting WER against the corrected text is 0.05% (en-IN) and 0.0% (hi-IN), showing both very clean source audio and very accurate ASR. Built for the Sarvam AI ML & Speech Data Pipeline assignment using a quality-first methodology.

• —Pipeline and report: https://github.com/auraCodesKM/sarvam-tts-dataset
• —Licence: CC-BY-4.0. Sources carry YouTube's Creative Commons Attribution flag, verified per clip, and are single-speaker:
• —Indian English - NPTEL Psychrometry lectures (nptelhrd), one instructor.
• —Hindi - a single-narrator reading of Premchand's essay "साहित्य का उद्देश्य" (Bhoopendra Pandey Hindi Channel); the text is public domain, the recording is CC-BY. Per-clip attribution and the recorded licence flag are in the manifest.

Schema

How it was built

Clips pass an ordered quality funnel: voice-activity segmentation, a pre-ASR acoustic gate (SNR, clipping, loudness, duration), Sarvam speech-to-text, language/transcript verification, a human listen-and-correct review, 24 kHz normalization with edge-trimming, and conservative emotion tagging. Automated gates run before any API call; every decision is logged in review/review_log.csv. ASR quality is quantified by word/character error rate against the human-corrected transcripts (reports/wer.json). Candidate sources were auditioned for TTS suitability before selection (src/audition.py).

Intended use and limitations

• —Intended use: research and educational TTS / ASR fine-tuning for Indian English and Hindi; small-scale voice modeling; a worked reference for curation methodology.
• —Limitations: lecture and narration content skews neutral/formal (439/555 clips tagged neutral), so affective emotion diversity is limited (documented rather than inflated). Emotion tags are LLM-derived from text, not independently confirmed by listening to each clip. The corpus is small (~73 minutes). The English source contains domain-specific terminology (thermodynamics); the Hindi source is literary prose.
• —Ethics: only YouTube CC-BY-flagged sources are used, with attribution preserved per clip. Not intended for impersonation or voice cloning of identifiable individuals without consent.

Citation

@misc{indian_en_hi_tts_2026,
  title  = {Indian English + Hindi TTS Dataset (curation-first)},
  author = {Thakur, Kavin},
  year   = {2026},
  note   = {Sources: NPTEL + Bhoopendra Pandey (CC-BY).
            Pipeline: github.com/auraCodesKM/sarvam-tts-dataset}
}