Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html> <h"... is not valid JSON

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Indian TTS Dataset — Hindi + Indian English

A curated, single-speaker TTS training dataset with 144 segments (~60 minutes total) sourced from YouTube, transcribed using Sarvam AI ASR, and annotated with emotion/style tags.

Dataset Summary

Split Segments Duration
Indian English 71 29.6 min
Hindi 73 30.8 min
Total 144 60.3 min

Audio Specs

  • Sample rate: 16 kHz
  • Channels: Mono
  • Format: WAV (16-bit PCM)
  • Segment length: 20–28 seconds

Columns

Column Description
audio Audio segment (16kHz mono WAV)
transcript ASR transcript (Sarvam saarika:v2.5)
language indian_english or hindi
language_code en-IN or hi-IN
emotion_tag Emotion/style label
source_channel YouTube channel name
style Content style (news_formal, conversational_podcast, etc.)
duration_s Segment duration in seconds
mean_dbfs Mean audio level in dBFS

Emotion Distribution

Emotion Count
questioning 56
formal 27
informational 19
motivational 18
conversational 17
emphatic 2
excited 2
storytelling 1
angry 1
sad 1

Sources

Indian English: The Seen and the Unseen

Hindi: Aaj Tak, Dhruv Rathee, Josh Talks Hindi

Pipeline

  1. Audio downloaded via yt-dlp, converted to 16kHz mono WAV via ffmpeg
  2. Silence-based segmentation into 20–28s chunks
  3. Transcription via Sarvam AI saarika:v2.5 ASR
  4. Automated QC: SNR check, transcript length, ASR artifact detection
  5. Emotion tagging: rule-based classifier on transcript + source style

See the full pipeline code at: https://github.com/itsharshi/indian-tts-dataset

License

Audio sourced from YouTube under fair use for research/educational purposes. Dataset annotations (transcripts, tags) are released under CC BY 4.0.

Downloads last month
13