Speaker Identification is known as the technology that enables users to access a device by speaking into the microphone and detecting the present talker among a group of speakers. The deformation of the incoming voice signal by external sounds greatly degrades the quality of the speaker identification systems. Noise could impact the efficiency of speaker recognition as well as cause the system to not function appropriately. Therefore, speech enhancement is critical for improving the performance of speaker recognition systems under challenging conditions. The spectral Subtraction method is among the most common approaches offered for audio enhancement since it is simple to apply and requires minimal computation in signal processing. The suggested system combines three primary modules: background noise reduction, feature extraction, and sound classification. First of all, the speech enhancement approach has been used to remove the additive noise. Secondly, a combined strategy of feature extraction that uses Mel Frequency Cepstral Coefficients (MFCC) as a feature extractor to be integrated with a Mel filter bank as a single package has been used. Furthermore, by using extracted features and a deep learning algorithm, this method can recognize the identity of the speaker. A convolutional neural network (CNN) for speech modeling that demonstrates very positive results in classifying participants was applied. This architecture has been built in a text-independent configuration. The dataset was made containing 50 speakers, each speaker has 20 voice samples. The accuracy and precision parameters were used to check the feasibility of this model. Our successful hybrid approach attained an accuracy and precision of 98.46%.
Recursive least square (RLS) is an effective noise cancellation algorithm, compared to other adaptive algorithms. Its performance depends on the prior knowledge of the noise reference signal available as one of the inputs to the algorithm. However, having noise reference information in advance is not always possible. For example, if the microphone is fixed at one position but the speaker position is not fixed, it is difficult to record the noise reference signal. To address this issue, a novel method is proposed that uses a coherent to-diffuse power ratio (CDR) and RLS algorithm to suppress reverberation. First, the CDR method processes the reverberant signal, and the result is fed to one of the inputs of the RLS algorithm to further remove reverberation. The proposed algorithm aims to investigate its performance in a distance-talking environment, where noise has a greater impact, and reverberation plays a vital role in degrading speech quality and intelligibility. The performance of the proposed method was compared to the CDR method and the microphone-recorded unprocessed signal. The evaluation criteria included PESQ, fwSNRseg, CD, and LLR. The proposed method yielded better results than the CDR method.
3 October 2026 in Circuits, Systems, and Signal Processing