Datasets:
metadata
license: other
language:
- hi
- en
task_categories:
- text-to-speech
- automatic-speech-recognition
tags:
- audio
- speech
- tts
- indian-languages
- hindi
- indian-english
- emotion
pretty_name: Indian English + Hindi TTS Dataset (emotion-tagged)
size_categories:
- n<1K
Indian English + Hindi TTS Dataset (emotion-tagged)
A curated, single-speaker-per-clip speech dataset for Text-to-Speech research, covering Indian English and Hindi. Every clip is sourced from YouTube, transcribed with Sarvam Saaras v3, and emotion-tagged via acoustic cues + a Sarvam LLM.
- Total: 82 clips, 55.6 minutes
- Hindi: 28.8 min | Indian English: 26.8 min
- Audio: mono, 24 kHz, 16-bit WAV
- Single speaker per clip, clean (no background music / overlapping speakers)
Schema
| field | description |
|---|---|
audio |
the waveform (24 kHz mono) |
transcript |
source-language transcript (Sarvam Saaras v3, mode=transcribe) |
emotion |
style/emotion label (acoustic cues + Sarvam LLM; 8-label taxonomy) |
language |
hi-IN or en-IN |
speaker |
source speaker |
speaker_gender |
speaker gender |
style |
content style |
duration |
seconds |
snr_db |
estimated SNR (QC) |
source_url |
original YouTube URL |
license |
per-clip source license |
Emotion distribution
| emotion | hi | en | total |
|---|---|---|---|
| serious | 20 | 14 | 34 |
| neutral | 4 | 12 | 16 |
| calm | 9 | 4 | 13 |
| happy | 4 | 3 | 7 |
| excited | 1 | 4 | 5 |
| angry | 2 | 1 | 3 |
| sad | 2 | 0 | 2 |
| surprised | 0 | 2 | 2 |
Speakers / sources
- Pronabesh Das
- Raj Shamani podcast (diarized single speaker)
- Sandeep Maheshwari
How it was built
- Source vetting — candidate YouTube channels were probed and rejected when they turned out to be AI two-host "audiobooks", laughter-heavy standup, or skits with background music. Only genuine, clean, single-human narration was kept.
- Download & segment —
yt-dlp+ffmpeg silencedetectcut audio at natural pauses into <=18s sub-clips. - QC — per-clip SNR / clipping / silence metrics auto-flag bad clips.
- Diarization — Sarvam Batch API extracted a single speaker's turns from a 2-speaker podcast.
- Transcription — Saaras v3 (
mode=transcribe, source language), kept clips only, cached (never billed twice). - Merge — consecutive clean sub-clips merged into ~30-50s clips by re-cutting one continuous span from the source; transcripts concatenated (no extra ASR).
- Emotion — auto-tagged via acoustic energy/dynamics + a Sarvam LLM (8-label taxonomy). A custom offline audition tool (in the repo) supports human keep/reject + emotion refinement.
- Packaging — attributed and published here.
See the project repository for the full pipeline, scripts, and an iteration log.
Intended use & limitations
- Intended for research / educational TTS and ASR experimentation.
- Audio is sourced from YouTube; rights remain with the original creators
(per-clip
license+source_urlare included for attribution). - Emotion labels are auto-generated (acoustic cues + LLM) and may be imperfect; treat as weak supervision, refine with the audition tool if needed.