dipit099/bengali-tts-excluded-v3

Bengali TTS Excluded Dataset (v3)

Quality-excluded subset of rwd51/bengali-tts-combined.

Exclusion Criteria

Rows excluded due to poor transcription quality:

• has_english=False AND (WER > 0.5 OR CER > 0.3)
• ~21,580 expected from CSV; actual ~23,520+ rows (includes duplicates from pipeline resume)

Columns

Column	Description
uuid	Unique identifier
speaker	Speaker name
video_id	Video ID
chunk_file	Chunk filename
audio_file	Audio filename
duration	Duration in seconds
transcription	Original transcription
audio	Audio (WAV, 16-bit PCM, mono, 22050 Hz)
gemini_transcription	Gemini transcription text
whisper_transcription	Whisper ASR output
wer	Word Error Rate
cer	Character Error Rate
has_english	Whether text contains English
decision	KEEP / EXCLUDE

Audio Column

The audio column uses the HuggingFace Audio feature type with struct format. All 13 shards have been converted to use the proper Audio struct format.

Splits

• train: Full dataset (~23K rows, 13 shards)
• preview: Small sample (20 rows) for dataset viewer compatibility

Source

Excluded from rwd51/bengali-tts-combined (397,437 rows) using pre-computed WER/CER scores. Companion KEEP dataset: dipit099/bengali-tts-filtered-v3