laion/tts-realspeech-dpo-en-de
Viewer • Updated • 5.91M • 425
Four EN/DE preference corpora with three-Whisper sidecars. Pitch-shift pairs are quality-confounded; read each dataset card.
Note Real-speech short/long negatives; 1,947,272 chosen-clip Whisper rows.
Note Emotion, truncation and continuation preferences; 2,146,797 chosen-clip Whisper rows.
Note CFG and phase-2 contrastive preferences; 550,612 indexed source-clip Whisper rows. Avoid transcript leakage in text-free pairs.
Note 66,805 experimental pitch-shift pairs with matched Whisper annotations; mostly quality-confounded, not a pure speaker-identity benchmark.