surindersinghssj/gurbani-asr-whisper-aligned

Gurbani ASR Whisper-Aligned Dataset

A forced-aligned corpus of Gurbani kirtan audio segments paired with canonical ground-truth Gurmukhi text from the SikhiToTheMax (STTM) database. Built for training low-resource Gurbani speech recognition models.

Whisper large-v2 provides word-level timestamps; canonical STTM text provides ground-truth labels. Whisper is never used as a transcriber — only as a timestamp oracle.

Dataset Summary

Language	Punjabi (Gurmukhi script)
Domain	Gurbani kirtan (devotional singing)
Audio format	FLAC, 16 kHz, mono
Segment duration	1–30 seconds
Source	SikhNet kirtan tracks with shabadId linking to STTM
Ground truth	STTM database — exact canonical Unicode Gurmukhi
Alignment	Whisper large-v2 word timestamps + matra-normalised F1 matching

How the Data Was Created

1. Source Selection

Kirtan tracks were scraped from SikhNet's artist APIs. Only tracks with a shabadId field (linking to the STTM database) were included, filtered by:

• Duration: 2–90 minutes
• Gurmukhi content ratio: ≥80%
• Exclusion of katha, akhand path, and non-kirtan content

2. Canonical Text Lookup

Each track's shabadId maps to exact canonical Gurbani text from the local STTM database.sqlite. This provides the ground-truth training labels — shabad lines (tuks), ang (page number), raag, and writer attribution.

3. Forced Alignment

Stage 1 — Whisper timestamp extraction:

• Model: Whisper large-v2 with word_timestamps=True, language="pa", beam_size=3
• Output: word-level timestamps {word, start, end, probability}

Stage 2 — Canonical text matching:

1. Matra stripping: Gurmukhi vowel signs (ਾਿੀੁੂੇੈੋੌੰੱ) are removed to compare consonant skeletons
2. Vishram splitting: Lines with primary vishram (;) are expanded into three match targets — full line, first half, second half — to handle kirtanis singing each half separately
3. Forward-scan F1: A sliding window over Whisper words computes F1 overlap with each canonical tuk partition
4. Thresholds: Match score ≥ 0.5, segment duration 1–30s, avg word confidence ≥ 0.3
5. Repetition tracking: Same tuk can match multiple times (captures kirtan repetitions)

4. Segmentation

Audio is extracted at Whisper-determined word boundaries and saved as FLAC (16 kHz, mono). The training label is always the canonical STTM text.

Features

Feature	Type	Description
audio	Audio	FLAC-encoded kirtan segment (16 kHz, mono)
segment_id	string	{recording_id}_{sequence:04d}
recording_id	string	MD5 hash of source URL (first 16 chars)
sentence	string	Ground truth — canonical Gurmukhi from STTM
whisper_text	string	Raw Whisper transcription (reference only)
tuk_index	int32	Index of canonical line in source shabad
start	float32	Segment start time (seconds)
end	float32	Segment end time (seconds)
duration	float32	Segment duration (seconds)
match_score	float32	F1 score of Whisper words vs canonical text (0.5–1.0)
avg_confidence	float32	Mean Whisper word confidence (0.0–1.0)
repetition	int32	How many times this tuk appeared so far in the recording
partition_type	string	"full", "first_half", or "second_half"
pipeline	string	"whisper_v2_gpu"
ang	int32	Page number in Guru Granth Sahib
style_bucket	string	Recording style (hazoori, puratan, akj, taksali, live, mixed)
artist_name	string	Kirtani name
shabad_id	int64	STTM database shabad ID
raag	string	Raag (melodic framework)
writer	string	Composer/saint credited with the verse

Quality Metrics

Match Score (F1): Overlap between Whisper-detected words and canonical STTM text after matra stripping.

• ≥ 0.8: high confidence match
• 0.6–0.8: good match
• 0.5–0.6: acceptable but noisy

Avg Confidence: Mean Whisper word-level probability.

• ≥ 0.7: Whisper confident
• 0.5–0.7: moderate
• < 0.5: Whisper uncertain (noisy audio)

Why Ground-Truth STTM Text?

Using Whisper's own transcription as training labels creates a closed loop where errors propagate. Instead, this dataset uses:

1. STTM database as the single source of truth for Gurbani text
2. Whisper timestamps only to locate where each word occurs in the audio
3. Matra-normalised fuzzy matching to handle Whisper's imperfect transcription

Even if Whisper misidentifies words, the canonical label remains correct.

Intended Use

• Fine-tuning Whisper or similar ASR models on authentic kirtan audio
• Curriculum learning (clean studio recordings → noisy live gurdwara audio)
• Vocabulary-constrained Gurbani decoding (STTM as closed vocabulary)
• Shabad search: live audio → matched shabad via BM25 index

Companion Dataset

A CTC-refined variant using Meta's MMS forced aligner is available at surindersinghssj/gurbani-asr-ctc-aligned, with additional columns for CTC alignment scores and boundary shift metrics.

Attribution

• Audio: SikhNet — original artists retain copyright
• Text labels: SikhiToTheMax database (public domain)
• Alignment pipeline: Surt (Gurbani ASR v3)

License

CC-BY-4.0. Audio sourced from SikhNet for non-commercial research and education.