Assessment of speech communication interference effects under small sample conditions

Aiming at the difficulty of data collection in the actual speech communication countermeasure and the lack of research on the disturbed speech in the strong jamming environment, in this paper, we study the estimation method based on the difference of domain distribution and category for small samples of speech under strong interference. By aligning the edge distribution and conditional distribution of the source domain and the target domain, the purpose of improving the evaluation accuracy in the case of small samples is achieved. The method first converts the interfering speech into a mel cepstrum, then uses a convolutional neural network to automatically extract features, and then evaluates small sample data by aligning the marginal distributions of the two domains. The experimental results show that the small sample evaluation model based on domain distribution differences can reach 82% on the target domain data set and 79% on the source domain data set, the small-sample evaluation model based on category differences can achieve an evaluation accuracy of 87% on the source domain data set, and the target domain data The set can reach 85%. Compared with the evaluation model on small samples, the accuracy rate is increased by 29%, which effectively solves the problem of evaluating the effect of voice communication interference in the case of small samples.