List your company

Emotion and speaker adaptation on Chinese speech data (test)

85.7SS (%)CSP-FT
73.32476.53779.7582.963Jan 24, 2025

Evaluation results

MethodLinks
CSP-FTFine-tuning strategy=Characteristic-Specific Partial Fine-Tuning (Proposed), Base model architecture=Fun-CosyVoice3.02025.0185.794.11.2
Full FTFine-tuning strategy=Full Fine-tuning, Base model architecture=Fun-CosyVoice3.02025.0185.193.93.9
LoRA FTFine-tuning strategy=LoRA, Base model architecture=Fun-CosyVoice3.02025.0184.792.82.1
OriginFine-tuning strategy=Zero-shot (None), Base model architecture=Fun-CosyVoice3.02025.0183.590.51.2
CSP-FTFine-tuning strategy=Characteristic-Specific Partial Fine-Tuning (Proposed), Base model architecture=CosyVoice2025.0181.291.92.5
Full FTFine-tuning strategy=Full Fine-tuning, Base model architecture=CosyVoice2025.018191.23
LoRA FTFine-tuning strategy=LoRA, Base model architecture=CosyVoice2025.0180.289.92.7
OriginFine-tuning strategy=Zero-shot (None), Base model architecture=CosyVoice2025.0179.888.62.5
CSP-FTFine-tuning strategy=Characteristic-Specific Partial Fine-Tuning (Proposed), Base model architecture=VALLE-X2025.0178876
Full FTFine-tuning strategy=Full Fine-tuning, Base model architecture=VALLE-X2025.0177.786.17.6
LoRA FTFine-tuning strategy=LoRA, Base model architecture=VALLE-X2025.0176.984.27.4
OriginFine-tuning strategy=Zero-shot (None), Base model architecture=VALLE-X2025.0175.482.35.2
Full FTFine-tuning strategy=Full Fine-tuning, Base model architecture=GPT-SoVITS2025.0175.480.93.3
CSP-FTFine-tuning strategy=Characteristic-Specific Partial Fine-Tuning (Proposed), Base model architecture=GPT-SoVITS2025.0175.2812.5
LoRA FTFine-tuning strategy=LoRA, Base model architecture=GPT-SoVITS2025.0174.878.53.4
OriginFine-tuning strategy=Zero-shot (None), Base model architecture=GPT-SoVITS2025.0173.870.32.2