surindersinghssj/gurbani-asr-ctc-aligned

Gurbani ASR CTC-Aligned Dataset

A forced-aligned corpus of Gurbani kirtan audio segments with CTC-refined boundaries using Meta's MMS (Massively Multilingual Speech) forced aligner. This is the CTC-refined companion to gurbani-asr-whisper-aligned.

Both datasets share identical segments and ground-truth labels. The difference is in boundary precision: this dataset refines Whisper's attention-based word boundaries using CTC acoustic alignment, producing tighter audio cuts.

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 timestamps → MMS CTC forced alignment refinement

How CTC Refinement Works

This dataset starts from the same Whisper timestamp alignment as the companion whisper-aligned dataset, then adds a CTC refinement step:

1. Whisper alignment (same as companion dataset): Whisper large-v2 provides word-level timestamps, matched to canonical STTM tuks via matra-normalised F1 scoring
2. CTC refinement: For each Whisper-matched segment:
   • Extract audio with ±0.5s padding around Whisper boundaries
   • Run MMS forced aligner (CTC-based) on the padded audio
   • Compute token-level frame alignments against the canonical text
   • Adjust start/end times based on acoustic CTC boundaries
   • Calculate alignment confidence score and boundary shift
3. Fallback: If CTC alignment fails for a segment, Whisper boundaries are kept (pipeline marked as whisper_v2_gpu+mms_ctc_fallback)

Why CTC Refinement?

Whisper's attention-based timestamps can have ±100–200ms jitter at word boundaries. CTC forced alignment uses frame-level acoustic evidence to snap boundaries to actual speech onset/offset, producing cleaner training segments with less silence padding and fewer cut-off phonemes.

Features

All features from the whisper-aligned dataset, plus:

Feature	Type	Description
audio	Audio	FLAC segment with CTC-refined boundaries (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)
mms_text	string	Decoded text from MMS alignment (approximate for Gurmukhi)
tuk_index	int32	Index of canonical line in source shabad
start	float32	CTC-refined segment start time (seconds)
end	float32	CTC-refined 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)
ctc_alignment_score	float32	Token alignment confidence from MMS CTC
boundary_shift_ms	float32	Mean absolute shift from Whisper boundaries (ms)
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+mms_ctc" or "whisper_v2_gpu+mms_ctc_fallback"
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

CTC-Specific Quality Metrics

CTC Alignment Score: Token-level alignment confidence from MMS forced aligner.

• Higher values indicate stronger acoustic evidence for the boundary placement

Boundary Shift (ms): Average absolute difference between Whisper and CTC boundaries (mean of start shift + end shift).

• < 100 ms: CTC confirms Whisper boundaries (good agreement)
• 100–300 ms: Moderate correction (typical for word-boundary refinement)

• 300 ms: Large correction (may indicate Whisper boundary error or CTC quirk — inspect manually)

MMS Text: Decoded text from MMS CTC alignment. Since MMS was primarily trained on Latin scripts, its Gurmukhi decoding is approximate. Use sentence (STTM canonical) as the true label — mms_text is for diagnostics only.

Companion Dataset

The Whisper-aligned variant (identical segments, Whisper attention boundaries instead of CTC-refined) is at surindersinghssj/gurbani-asr-whisper-aligned.

Intended Use

• Fine-tuning Whisper or similar ASR models with tighter acoustic boundaries
• Comparing CTC vs attention-based alignment for Gurbani ASR training
• Curriculum learning (clean studio → noisy live gurdwara audio)
• Vocabulary-constrained Gurbani decoding (STTM as closed vocabulary)

Attribution

• Audio: SikhNet — original artists retain copyright
• Text labels: SikhiToTheMax database (public domain)
• MMS model: Meta AI (Massively Multilingual Speech)
• Alignment pipeline: Surt (Gurbani ASR v3)

License

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