Abstract:Existing methods still face challenges in terms of generalization ability in cross-domain tasks and capturing the precise mapping between speech temporal features and global scores. To address these issues, this paper proposes a cross-domain objective speech quality evaluation method that integrates multi-source adaptation with temporal-global score mapping. The method utilizes the distortion prior of the target domain to perform adaptive multi-source selection, matching the data distributions of source and target domains, thus reducing negative transfer from anomalous source domains. On this basis, by combining the local perceptual capability of convolutional neural networks (CNN) and the long-term dependency modeling ability of bi-directional long short-term memory (BLSTM), a CNN-BLSTM-based objective speech quality scoring model is proposed. Additionally, a hybrid loss function combining frame-level prediction error with global error is introduced to improve the model’s robustness to frame-level quality fluctuations and enhance the stability of predictions. Experimental results show that the proposed method outperforms current mainstream NI-SQE approaches using only about half the training data, achieving average SRCC and PLCC improvements ranging from 2.21% to 35.62% and from 2.61% to 39.22%, respectively. This demonstrates not only its strong cross-domain generalization ability but also its feasibility for practical engineering deployment.