Speech Processing

Speech processing is one of the crucial type of recognition technique when it comes to AI and it is challenging because the vocabulary of the people differs from region to region. The speech processing involves the followings steps:

1. Analysis of the Input Signal

2. Feature Extraction from the analyzed signal

3. Classification

The types of feature extraction in speech recognition are:

1. Mel-Frequency Cepstrum Coefficient (MFCC)

2. Linear Predictive Coding

3. Linear Predictive Cepstral Coefficient (LPCC)

4. Linear Discriminant Analysis (LDA)

5. Perceptual Linear Prediction (PLP)

6. Discrete Wavelet Transforms (DWT)

Classification of the speech is done using the following approaches:

1. Pattern Recognition Approach

2. Acoustic Phonetic Approach

3. Artificial Intelligence Approach

In order to do the Sentimental Analysis of the speech signal which is one of the types of Pattern Recognition Approach the methods used are:

1. Hidden Markov Model.

2. Dynamic Time Warping.

3. Neural Networks.

Dynamic Time Warping:

In speech processing time alliance of separate expressions is the central part of the problem to find the distance measurement. If there’s a slight time elapse and the word will have a different meaning and leads to false detection of the expression. The DTW is one of the algorithms that will analyze the resemblance in the two speech sequences that vary with respect to time and speed.

A variety of segments of the speech expressions are expanded and compressed to expose the arrangement ending with the promising preeminent match between test and position vectors function by facet root. Following are the steps that are done in DTW.

1. Record, consider and make up a dictionary of few reference words.

2. Record analyzed term for familiarization with and parameterization.

3. Evaluating the distance with the reference and the test words.

4. Picking the reference word nearest to the test word.

Hidden Markov Model (HMM):

HMM provides an effortless and valuable formation of a model that is time-varying spectral vector sequences. HMM is a model where the state series are hidden, and observations are state probabilistic functions in which a set of states are related to the changes. The HMM consists of the following parameters a, b and π and N which is the model of the states and M is referred to as,

The HMM model can be shown as,

The steps involved in calculation of HMM are:

1. Forward Algorithm

2. Backward Algorithm

Neural Networks:

Neural network is a mathematical ordering which is skilled enough to find out a complex relationship with any complex nonlinear connection between the input and the output variables. These models are found to be more important and well organized, primarily used when there is trouble in process characteristics are more complicated to be described using mathematical expression. The information that is obtained from the ANN is well versed because the network tries to learn from the input and output that are fed to the system.