Assumptions and calculations

1. DC Shift

---

DC shift is found out by taking the average of the sample data where microphone is turned off.
(I have considered the DC shift to be 0 if the average is less than 0.0005, since this value is too small.)

2. Normalization Factor

---

Normalization factor is found out by dividing the average of the two extreme values by 5000. For example - if the maximum value is 12500 and the minimum value is -12600 then normalization factor will be 5000/((12500+12600)/2) i.e. 0.398406 .
(To make things simpler, I am taking only one digit after decimal point. That means 0.398406 becomes 0.3 . It doesn't affect the values much. It also ensures that the range remains 5000 to -5000 since we are reducing the value from 0.398406 to 0.3 .)

3. Stable frames searching

---

While reading the input file, the point at which the amplitude value goes beyond the range -1500 to 1500 (i.e. either smaller than -1500 or larger than 1500) next 2000 amplitude values are skipped and next 320x5 values are taken. This process is giving good results. (100% accuracy for all vowels except /i/ which has 80% accuracy)

Functions and their Functionalities

1. float findDCShift(char * fileName)

---

This function is finding out the DC shift from <fileName> file.

2. float getNormalizationFactor(char * fileName, float range)

---

This function finds out the normalization factor for the given speech data.

3. void writeToFile(double * arr, char * fileName, int size)

---

This function is writing the data of arr into fileName file.

4. double * getRValues(double * x, int sampleSize, int p)

---

This function is finding out the R values using the auto corelation formula.

5. double * getAValues(double * R, int sampleSize, int p)

---

This function is finding out the A values using Levinson Durbin's Algorithm.

6. double * getCValues(double * A, int sampleSize, int p, double r0)

---

This function is calculating C_i values for 1 <= m <= p.

7. int _tmain(int argc, _TCHAR* argv[])

---

This is the main function. Here other functions are called sequentially as per the requirement. First the validation is done for the given data an it is written into a file for reference. Then training is done for first 10 samples followed by testing for the next 10 samples. Tokhura distance is taken into consideration for finding out the vowel spoken.

How to create the input files

1. Speech sample

---

This can be recorded by speaking the vowels and can be named as per requirement. It is better to follow some convention so that the string concatenation features can be used.

Testcases description

The convention followed is - a few fixed digits followed by vowel name followed by count e.g. 214101037_a_04. So for each iteration only the vowel name and count will be changed. This format can be given in the array at line number 193 to 195. All the other informations such as sample size, p, q, number of alphabets, number of speech samples etc can be changed as per requirement from line number 185 to 216 in the main cpp file. The variables on the next lines shouldn't be altered.

At line numbers 287-289, 348, 371-373, 424 must be changed according to the file name convention. These lines replace the vowel and the file count for each iteration.

Instructions to execute the code

1. The project should be opened in visual studio.

2. The variables which should be decided before execution are listed separately within main function at line number - 185 to 216.

3. If the naming convention of the files are being changed, then proper changes to some variables should be made.

4. To see the console window, a breakpoint should be put at line number - 478. For this, the shortcut is to place the cursor at 478 and press f9 key. (Fn+f9 for some systems)

5. f7 can be pressed for building the project and f5 should be pressed for execution.

6. Output can be observed in the console window.

7. The output file containing the same information as the console window will be in the same directory where the program is present.

Want to improve the project?

Fork this repository. Make the required changes and create a pull request.

• This project is giving 80%+ accuracy for others' speech samples also ;)