nitya2405/sarvam-tts-dataset

Sarvam TTS Dataset

A curated dataset of clean, single-speaker audio clips in Indian English and Hindi, assembled for training and evaluating text-to-speech (TTS) and speech recognition models. Each clip is 45–60 seconds, loudness-normalized to broadcast standard, and tagged with a primary emotion label by a human reviewer.

Dataset Summary

Total clips	72
Total duration	62.5 minutes
Languages	Indian English (EN), Hindi (HI)
Created	2026-06-18
Audio format	24 kHz mono 16-bit PCM WAV, −23 LUFS

Duration and Emotion by Language

English — 37 clips, 32.0 minutes

Emotion	Clips
conversational_casual	6
happy_excited	6
intense_dramatic	6
formal	4
neutral	4
sad	4
calm_reverent	4
angry	3

Hindi — 35 clips, 30.5 minutes

Emotion	Clips
happy_excited	6
sad	5
calm_reverent	5
formal	4
intense_dramatic	4
neutral	3
angry	3

Genre Distribution

Across 20 source videos spanning: speech (21), conversational (14), podcast (13), news (11), dramatic (11), vlog (7), standup (6), cooking (2), poetry (3).

Methodology

1. Source selection

Sources are publicly available YouTube videos featuring natural Indian English and Hindi speech across diverse genres (news anchors, political speeches, stand-up comedy, poetry recitals, cooking shows, vlogs, podcasts). Genre and speaker diversity was prioritised to cover the full emotion taxonomy.

2. Download

Audio extracted with yt-dlp at the best available quality, then converted to 16 kHz WAV for diarization input.

3. Speaker diarization and the single-speaker interpretation

The central design decision in this pipeline is how to handle multi-speaker audio.

YouTube recordings almost always contain more than one speaker — interviews, panel discussions, call-in shows, and ambient audience noise are the norm. The brief asked for "single-speaker audio clips for TTS," but this constraint is genuinely ambiguous when applied to real-world sources: does it mean the source video must have one speaker, or that each output clip must contain only one speaker's voice?

We resolved this as follows: each output clip must contain exactly one speaker's voice, but source videos may contain multiple speakers. This allows a much wider pool of sources and more natural speech, while still producing clean single-speaker clips suitable for TTS.

Implementation:

• Speaker separation was performed with Sarvam AI's saaras:v3 batch API (mode=transcribe, with_diarization=True), which returns a turn-by-turn transcript annotated with speaker_id labels.
• For each source, a human reviewer inspected the speaker_summary field of the diarization output (which reports total speaking time per speaker) and identified the target speaker — the single voice that should represent that source in the dataset. This was recorded in source_log.csv as target_speaker_label.
• Only turns attributed to the target speaker were retained. Turns from other speakers, including interviewers, audience, and background voices, were discarded.

This means every clip contains exactly one speaker's voice throughout. However, the transcription reflects only what the target speaker said, which may include responses to questions or reactions to other speakers present in the original recording.

4. Clip construction

After diarization:

• Adjacent turns from the target speaker separated by ≤ 3 s were merged into a single continuous segment.
• Long segments were split at natural silence boundaries (≥ 200 ms pause, threshold −40 dB) into clips of 45–60 s. The 45 s lower bound ensures enough context for prosody modelling; the 60 s upper bound fits within practical batch job limits and keeps annotation effort per clip tractable.

5. Audio processing

Each clip was processed in order:

1. Resample to 24 000 Hz mono 16-bit PCM WAV (the standard for TTS training).
2. High-pass filter at 80 Hz to remove low-frequency rumble and HVAC hum.
3. Loudness normalise to −23 LUFS integrated loudness (EBU R128), −1 dBTP true-peak ceiling — the same standard used in broadcast and podcast mastering.
4. Silence trim at clip edges with ~175 ms of padding preserved.

6. Quality control flags

Before writing each clip, the pipeline runs two automated checks and records the results in the qc_flags column of the metadata CSV. Flagged clips are not automatically rejected — the human reviewer made the final keep/discard decision in Phase 5.

Flag	Condition	Meaning
clipping	Any sample ≥ 0.99 absolute value	The source recording hit full scale before normalisation; may have transient distortion
low_snr	Integrated loudness of the raw clip < −35 LUFS	Likely a poor-quality microphone or high ambient noise floor

The presence of a QC flag in an approved clip means the reviewer listened and judged the artefact acceptable (e.g., a single clipped transient in an otherwise clean recording).

7. Transcription

Verbatim ASR via Sarvam AI's saaras:v3 batch API (mode=verbatim, diarization off). Clips were grouped by language and batched into jobs of ≤ 20 files (the API upload limit). The Sarvam SDK names output files after the uploaded filename, so transcript–clip matching was done by filename lookup rather than by position to avoid order-dependent misalignment across chunks.

8. Human tagging and approval

Each clip was reviewed in a purpose-built Gradio application. For every clip, the reviewer:

• Listened to the audio and read the auto-generated transcript.
• Corrected any transcription errors (disfluencies, proper nouns, code-switching).
• Assigned a primary emotion tag from the taxonomy below (required to approve).
• Optionally assigned a secondary emotion tag.
• Approved or rejected the clip.

Only approved clips are included in this dataset. Clips with corrected transcripts reflect the human-reviewed version, not the raw ASR output.

Emotion taxonomy: neutral, formal, happy_excited, sad, angry, calm_reverent, conversational_casual, intense_dramatic.

Data Fields

Column	Type	Description
audio	Audio (24 kHz)	Audio clip
clip_filename	string	Original WAV filename
language	string	en or hi
genre	string	Source genre
speaker_name	string	Speaker name from source log
source_index	int	Index into source video list
video_id	string	YouTube video ID
source_channel	string	YouTube channel name
source_url	string	YouTube URL
clip_start_time_s	float	Start time in the original source audio (seconds)
clip_end_time_s	float	End time in the original source audio (seconds)
duration_s	float	Clip duration in seconds
transcript	string	Human-reviewed verbatim transcript
primary_emotion	string	Primary emotion tag
secondary_emotion	string	Secondary emotion tag (may be empty)
approved	bool	True = approved by human reviewer
reviewed	bool	True = clip was seen by reviewer
qc_flags	string	Pipe-delimited QC flags (clipping, low_snr), or empty

Audio Format

Property	Value
Sample rate	24 000 Hz
Channels	Mono
Bit depth	16-bit PCM
Format	WAV (stored as parquet via HuggingFace Audio feature)
Loudness	−23 LUFS (EBU R128)
True peak	−1 dBTP
High-pass	80 Hz

License and Copyright

Audio clips: Sourced from publicly available YouTube videos. The clips are provided here for non-commercial research purposes under fair use. Copyright in the original audio remains with the respective creators and YouTube channels. Downstream users who wish to use the audio commercially should obtain clearance from the original creators.

Metadata (transcripts, emotion tags, QC flags, and all other structured data): This is original annotation work released under CC-BY 4.0. Attribution: Nitya Shukla, 2026.

The CC-BY 4.0 tag on this repository applies to the metadata. We do not claim or grant any rights over the audio content itself beyond what fair use permits.

Pipeline

The full pipeline (download → diarize → preprocess → transcribe → tag → publish) is open-source at nitya2405/sarvam-tts-dataset-pipeline.

Citation

@dataset{sarvam-tts-dataset-2026,
  title  = {Sarvam TTS Dataset},
  author = {Shukla, Nitya},
  year   = 2026,
  url    = {https://huggingface.co/datasets/nitya2405/sarvam-tts-dataset},
  note   = {Indian English and Hindi single-speaker TTS clips, human-reviewed,
            emotion-tagged, 74.4 minutes total}
}