| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| - hi |
| size_categories: |
| - n<1K |
| task_categories: |
| - text-to-speech |
| - automatic-speech-recognition |
| tags: |
| - tts |
| - indian-english |
| - hindi |
| - emotion |
| - speech |
| --- |
| |
| # Indian English + Hindi TTS Mini Dataset |
|
|
| A small (~60 min) speech corpus assembled for TTS training, sourced from |
| public YouTube content and transcribed with the Sarvam `saaras:v3` model. |
|
|
| ## Summary |
|
|
| | Field | Value | |
| |---|---| |
| | Total duration | ~60 min | |
| | Indian English (en-IN) | ~30 min | |
| | Hindi (hi-IN) | ~30 min | |
| | Audio | mono, 16 kHz, WAV | |
| | Per-clip duration | ≤ 30 s | |
| | Emotion tags | Angry, Calm, Excited, Formal, Happy, Instruction, Motivation, Narration, Neutral, News, Sad, Surprised, Whisper | |
|
|
| ## Schema |
|
|
| Each row contains: |
|
|
| - `clip_id` — deterministic id, `{video_id}_{start_sec}_{end_sec}` |
| - `audio` — 16 kHz mono WAV (HuggingFace `Audio` feature) |
| - `transcription` — Sarvam `saaras:v3` transcript, lightly normalized |
| - `language` — `en-IN` or `hi-IN` |
| - `emotion` — one of the categories above |
| - `speaker_id` — set to source `video_id` (one speaker per source video) |
| - `duration_sec`, `start_sec`, `end_sec` — integers, seconds |
| - `video_id`, `source_url` — YouTube provenance |
|
|
| ## Notes & curation |
|
|
| - Clips longer than 30 s in the source spec were split into back-to-back |
| ≤30 s pieces so they fit Sarvam's sync ASR limit. |
| - Transcripts come from Sarvam `saaras:v3`; every clip was then reviewed |
| manually, with multiple iterative passes to correct ASR errors, |
| normalize spelling, and tighten emotion tags. |
| - A separate diarization model was **not** used. Each source URL/timestamp |
| was hand-picked to contain a single speaker, and the manual review |
| pass listened end-to-end to every clip and re-cut or replaced any |
| that drifted into multi-speaker territory. |
| - `speaker_id` is set to the YouTube `video_id` — each source video has |
| exactly one speaker by construction. |
| - Emotion tags come from the curator's labelling of each source clip, |
| refined during the review passes. |
|
|
| ## License |
|
|
| Audio is sourced from YouTube; rights remain with their original creators. |
| Transcriptions and metadata are released under CC-BY-4.0. Please consult |
| the original videos for redistribution constraints. |
|
|
