Mel-Frequency Cepstral Coefficients is an algorithm which focuses on accurately determine what sound comes out. “The shape of the vocal tract manifests itself in the envelope of the short time power spectrum, and the job of MFCCs is to accurately represent this envelope.” [1]

Linear Predictive Coding is an algorithm that is chosen mostly in audio signal processing for illustrating the spectral envelope of a digital signal of speech in compressed form, using the information of a linear predictive model.

Perceptual Linear Predictive approximates the auditory spectrum of speech by an all-pole model. This method has bark filter bank. Bark scale mainly reflects subjective loudness perception in auditory system. It is known that this method has lower computation speed than MFCC.

Linear Predictive Cepstral Coefficients are cepstral coefficients derived from Linear Predictive Coding. LPCCs are used to capture emotion-specific information because LPCC method has a model of human vocal tract [2]. However, LPCC have low noise resistance [3].

The Gammatone Frequency Cepstral Coefficients is based on gammatone filter bank, which models the basilar membrane as a series of overlapping band pass filters [4].

Cochlear Filter Cepstral Coefficients is an auditory-based algorithm for extraction and measured regarding to a newly developed auditory-based time-frequency transform named Auditory Transform (AT) and a set of simulations from the processing functions in the cochlea.

I-vector is a front-end factors analysis which uses separate speakers and channel dependent subspaces. Its speech segment is expressed in a low-dimensional “identity vector”.

Gaussian mixture model recognizes the speaker on the basis of log probability. It recalculates the log probability of voice vector and compares it to previously stored value. To attain a maximum likelihood estimate we utilize an iterative expectation-maximization (EM) algorithm.

Hidden Markov Model is a statistical model that has unobservable states. In the Hidden Markov Model, the states are not directly visible, but the output is visible. Each state has a probability distribution over the possible outputs.

Vector Quantization maps the vectors from a large set into groups. The Euclidian distance between the acoustic vectors of test speaker and the codewords in the codebook is computed. The speaker having the littlest Euclidian distance is chosen [5].

Dynamic Time Warping is a method for comparing two temporal sequences, its principle is basically dynamic programming. It can be used to eliminate the time difference caused by speaking speed between two speech patterns and compare the similarity [6].

We took a part of data from a public voice dataset called VoxCeleb as our dataset for testing. Our dataset has thousands of pieces of 3-6 seconds audio. In total, the audio is 5-hour-long. When we test, we make our system do speaker identification rather than speaker verification (we didn't specify the speaker to be verified and let the system identify the speaker who was talking). The accuracy was calculated and the duration was noted down.

There are some open source libraries and APIs about MFCC and GMM. We just used them to implement the two algorithms. For the rest of the algorithms, there is no libraries and APIs. We implemented them either by reading papers or by studying other similar github projects.

We tried our best when it comes to implementing the algorithms and for some of them we succeeded. We still need to do some work for some of them. In terms of comparison between frameworks, we expect PLP + VQ to have the highest accuracy according to what we researched.