روشی جدید در بازشناسی مقاوم گفتار مبتنی بر دادگان مفقود با استفاده از شبکه عصبی دوسویه

Performance of speech recognition systems is greatly reduced when speech corrupted by noise. One common method for robust speech recognition systems is missing feature methods. In this way, the components in time - frequency representation of signal (Spectrogram) that present low signal to noise ratio (SNR), are tagged as missing and deleted then replaced by remained components and statistical information of clean speech. In this article a new approach of missing features method based on compensation are proposed. A Bidirectional Neural Network (BNN) was developed and implemented in order to modify unreliable components in input feature vectors and improve the overall recognition accuracy. Distorted components in feature vectors were estimated in accordance with the latent knowledge in the hidden layer of the neural network. This knowledge is obtained by training with clean and noisy speech, simultaneously and is mostly induced by reliable and less influenced components by the irrelevant variations in speech signal. In this approach, there is no need to identify missing components that is a challenging issue in the field of robust speech recognition based on missing feature method because reconstruction is done on all components (whether reliable or unreliable), in order to become more similar to the clean speech component. This point is a very significant advantage that has been achieved in this article. Comparing the results of these two methods shows that using Missing feature methods, 4.2% improvement were obtained in the accuracy of speech recognition for noisy signal by SNR=0dB, whereas improvement value increased to 8.5%, using bidirectional neural network for the same signal to noise ratio.