Speech Quality Evaluation Using Latent Clustering and Masked Prediction

The patent replaces supervised learning-based MOS score estimation with an unsupervised anomaly detection approach. Instead of training models on large-scale speech-MOS pair datasets, the method uses normal speech data to train an autoencoder model that learns the distribution of normal speech. The quality evaluation is then performed by detecting anomalies between synthesized speech and the learned normal distribution, eliminating the need for large labeled datasets while maintaining objective evaluation capability.

The patent implements self-service by using the speech data itself to train the evaluation model. The autoencoder is trained solely on normal speech data without requiring external MOS labels. The model then uses its own learned representation to evaluate synthesized speech quality by measuring reconstruction errors, making the system self-sufficient and independent of large-scale annotated datasets.

The patent substitutes manual subjective evaluation with an automated unsupervised learning system. The autoencoder model processes speech data automatically, computing reconstruction errors as quality metrics without requiring human evaluators. This maintains objective measurement while dramatically improving evaluation efficiency and scalability.

The patent introduces latent representations as an intermediary between raw speech data and quality evaluation. The autoencoder compresses speech into latent space, and quality is assessed through reconstruction error in this intermediate representation. This intermediary mechanism enables automated objective evaluation that correlates with human perception without requiring direct human involvement in the evaluation process.