---
license: cc-by-4.0
language:
  - en
  - hi
tags:
  - tts
  - text-to-speech
  - indian-english
  - hindi
  - speech
  - single-speaker
  - sarvam
  - audio
task_categories:
  - text-to-speech
pretty_name: Sarvam Indian TTS Dataset (en-IN + hi-IN)
size_categories:
  - 1K<n<10K
---

# Sarvam Indian TTS Dataset

A curated Text-to-Speech dataset containing **Indian English** and **Hindi** 
single-speaker audio segments, built as part of the Sarvam AI hiring assignment.

## Dataset Summary

| Property | Value |
|---|---|
| Total duration | 78.8 minutes |
| English (en-IN) | 57.5 minutes |
| Hindi (hi-IN) | 21.3 minutes |
| Total segments | 175 |
| Human reviewed | 99.4% |
| Mean quality score | 92.1/100 |
| Mean SNR | 49.0 dB |
| Mean ASR confidence | 0.900 |

## Dataset Card

### Data Collection

Audio sourced from YouTube educational and motivational content featuring 
single speakers with minimal background noise. All audio was processed through:

1. **Speaker Diarization** (Sarvam API) — isolate dominant speaker
2. **Segmentation** — 15–60 second clips, no overlap
3. **Quality Filtering** — SNR ≥ 15 dB, speech ratio ≥ 65%
4. **ASR Transcription** (Sarvam saarika:v2) — with word timestamps
5. **Emotion Tagging** (Sarvam LLM) — style/prosody labels
6. **Human Review** — transcript correction and quality verification

### Audio Format

- Format: WAV (PCM 16-bit)
- Sample rate: 16,000 Hz
- Channels: Mono

### Schema

```
id              — unique segment identifier
audio           — WAV audio file (HF Audio feature)
transcript      — verbatim speech transcript
language        — en (Indian English) | hi (Hindi)
emotion         — style label (see taxonomy below)
duration        — segment length in seconds
confidence      — ASR confidence score [0, 1]
source_url      — original YouTube URL
human_reviewed  — whether a human verified this segment
```

### Emotion Taxonomy

  - informative: 158
  - emphatic: 16
  - energetic: 1

### Source Videos

| Language | Description | Accent |
|---|---|---|
| English | Stanford GSB lecture (Matt Abrahams) | American English |
| English | Indian tech talk | Indian English |
| Hindi | Hindi motivational speech | Standard Hindi |
| Hindi | Hindi motivational speech | Standard Hindi |

> **Note on English accent**: The primary English source (Stanford lecture) 
> features an American English speaker. This is documented transparently in 
> the `accent` metadata field. The Sarvam en-IN ASR model was used for 
> transcription regardless.

### Pipeline

Full reproducible pipeline available at: https://github.com/[username]/sarvam-tts-dataset

```
scripts/
  00_setup_env.py         # environment validation
  01_download_audio.py    # yt-dlp download
  02_convert_audio.py     # ffmpeg → 16kHz mono WAV
  03_diarization.py       # Sarvam speaker diarization
  04_segment_audio.py     # 15–60s segmentation
  05_asr_transcription.py # Sarvam saarika:v2 ASR
  06_emotion_tagging.py   # Sarvam LLM emotion labels
  07_quality_scoring.py   # composite quality score
  08_human_review.py      # interactive CLI review
  09_build_metadata.py    # final CSV/JSONL assembly
  10_dataset_statistics.py # stats + visualisations
  11_upload_huggingface.py # HF Hub upload
```

### Licensing

- Dataset: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
- Source videos used under YouTube's fair use provisions for research/education
- Transcripts are derived works; see individual source URLs for original content

### Citation

```bibtex
@dataset{sarvam_tts_dataset_2025,
  title={Sarvam Indian TTS Dataset},
  year={2025},
  url={https://huggingface.co/Avishii0309/sarvam-tts-dataset},
  note={Indian English and Hindi single-speaker TTS dataset}
}
```