Datasets:
Audio and text columns are misaligned - most rows pair the wrong sentence (verified by transcription)
We trained a Piper TTS voice on this dataset and got unintelligible output despite healthy training metrics, so we audited the pairing directly: we transcribed all 6,892 clips with Whisper large-v3 and compared each transcript against the texts shipped in the dataset.
Findings:
- Only 1,323 of 6,892 rows (19%) have audio that matches the text in their own row. Those correct rows sit in the first ~44% of the file; after that, essentially no row matches.
- For the remaining ~81%, the correct sentence is not merely in a different row - it does not appear anywhere in the dataset's text column. Example: distinctive words from a story the female speaker reads (e.g. ਕਿਸਮਤ) occur in 0 of the 6,892 provided texts.
- Concrete example, second row of the train split as distributed: the text column reads "ਇਕ ਤਾਂ ਨਵੀਂ ਇਸ ਗੱਲੋਂ ਕਿ ਉਹ ਆਪਣੇ ਖ਼ਾਵੰਦ ਦੀ ਦੂਸਰੀ ਬੀਵੀ ਹੈ..." while the audio in the same row says "ਕਾਰਣ ਐ ਕਿ ਮੈਂ ਤਾਂ ਹਾੜੇ ਨਾਲੋਂ ਵਕਰੀ ਕਿਸਮਤ ਦੇ ਵਡਿਆਈ ਗਈ ਹਾਂ" - a sentence from a different story.
This looks like the audio and text lists came from two different subsets of the IndicTTS corpus when the parquet was built. The audio itself is fine - it's the pairing that's broken.
Could you publish a corrected version, or the original filename-to-transcript mapping from the IndicTTS release? We can share our full per-row transcription file if it helps verify or rebuild the pairing.
Update: verified pairing is 35.2% (not 19.2%), sources identified
Updating my original report with a correction and new findings.
Correction: the 1,323 figure understated the pairing. The first matcher punished spelling differences in nukta/bindi diacritics, which Whisper renders inconsistently across Devanagari and Gurmukhi. Re-running with diacritic-insensitive normalization gives 2,425 of 6,892 rows (35.2%) verified as same-index pairs (female 1,204, male 1,221), spot-checked by reading the pairs. A cross-index pass over the remaining 4,467 clips found zero further matches, so the unmatched sentences are absent from the text column rather than displaced within it. After a 1-10 s duration filter the verified set is 2,355 clips (1,175 female / 1,180 male, about 1.9 h per speaker).
Sources identified for the sentence pool, in case it helps a future corrected release: the text column opens with Amrita Pritam's "Jangli Booti"; the dominant block (roughly rows 900-3445) is Balraj Sahni's autobiography "Meri Filmi Aatmakatha" (content-verified against its English translation); the history block in the audio (roughly rows 1000-2000) is adapted from Ganda Singh's "Sardar Jassa Singh Ahluwalia".
The request from the original post stands: a corrected release, or the original filename-to-transcript mapping, would fix this completely. If anyone holds the Punjabi edition of the Sahni autobiography (Language Department Punjab, 1970), I would also welcome a pointer - it is digitized at the Panjab Digital Library (BK-070966) but custodian-restricted.