Speaker recognition is a classification task to identify people from their voices. Bangla is one of the most spoken Indo-Europian languages which stands 5th among 100 most spoken languages. There are around 233.7 million people around the world who speak in Bangla. Deep learning has recently significantly improved speaker recognition. The majority of current research techniques are based on convolutional neural networks (CNN) or recurrent neural networks (RNN), as the voice signal can be converted to spectrogram as 2-D input or also as 1-D input for sequential modeling. In this work, we collected audio samples from 100 speakers of Bangla language. Furthermore, we converted those audio files into spectrograms to train a convolutional neural network model incorporating residual blocks, which can identify 100 different speakers. Moreover, we used the spectrograms produced by the Short-time Fourier Transform, Mel Spectrograms and MFCC Spectrograms to train the CNN model and evaluate the performance of the model with different spectrogram types as input for Bangla speaker identification. When the MFCC spectrogram is given as input, the highest accuracy of 0.98125 achieved the CNN model , the model taking spectrogram generated through STFT achieved an accuracy of 0.9625 and finally the model with Mel Spectrogram achieved the lowest accuracy of 0.91666.