Enhancing Arabic ASR in Noisy and Transcoding EVS Conditions: A Multimodal Deep Learning Study

In this paper, we investigate the impact of speech transcoding and noise on the performance of Arabic automatic speech recognition (ASR) systems based on deep learning.We apply Non-negative Matrix Factorization (NMF) as a denoising preprocessing step to enhance robustness to noise.Three deep architectures-CNN-LSTM, LSTM, and DNN-are evaluated using fused acoustic features including MFCCs, Mel-spectrograms, and Gabor filter representations.Experiments are conducted under four signal-to-noise ratio (SNR) conditions (-5 dB, 0 dB, 5 dB, and 10 dB) on both transcoded and nontranscoded speech.Results show that the CNN-LSTM model achieves the highest accuracy of 87% at 10 dB SNR on clean (non-transcoded) speech using multimodal features.However, speech recognition performance degrades by 2-4% when using the Enhanced Voice Services (EVS) codec, especially in highnoise environments.Specifically, accuracy drops from 65.00% to 61.43% at -5 dB SNR, and from 87.00% to 84.00% at 10 dB SNR due to transcoding.These findings highlight the negative impact of mobile codec compression on ASR systems, particularly under low-SNR conditions.Our study confirms the effectiveness and stability of NMF-based feature fusion and denoising in improving recognition, offering insights into deploying Arabic ASR in real-world scenarios such as mobile and VoIP communications.