Cross-sentence Zero-shot Speech Synthesis on LibriSpeech clean (test)
1.9WERHVoicebox
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VoiceboxTraining Data Hours=60K2024.07 | 1.9 | — | 66.2 | |
| MELLETraining Data Hours=50K2024.07 | 2.1 | 1.47 | 62.5 | |
| MELLE-R2Training Data Hours=50K, reduction factor=22024.07 | 2.14 | 1.5 | 60.8 | |
| Ground Truth2024.07 | 2.15 | 1.61 | 77.9 | |
| MELLE-R3Training Data Hours=50K, reduction factor=32024.07 | 2.19 | 1.51 | 57 | |
| Ground Truth (mel-spectrogram)2024.07 | 2.24 | 1.64 | 73.2 | |
| MELLE-R4Training Data Hours=50K, reduction factor=42024.07 | 2.3 | 1.56 | 53.2 | |
| Ground Truth (EnCodec, 8 codebooks)2024.07 | 2.33 | 1.65 | 71 | |
| VALL-E 2Training Data Hours=50K2024.07 | 2.44 | 1.5 | 64.3 | |
| MELLE-R5Training Data Hours=50K, reduction factor=52024.07 | 2.72 | 1.96 | 50.6 | |
| RALL-ETraining Data Hours=44K2024.07 | 2.8 | 2.5 | 49 | |
| MELLE-limitedTraining Data Hours=9602024.07 | 2.8 | 2.21 | 59.1 | |
| VALL-E RTraining Data Hours=9602024.07 | 3.97 | 3.18 | 36.5 | |
| CLaM-TTSTraining Data Hours=55K2024.07 | 5.11 | — | 49.5 | |
| VALL-ETraining Data Hours=60K2024.07 | 5.9 | — | 58 | |
| ELLA-VTraining Data Hours=9602024.07 | 8.9 | 7.15 | 30.7 |