nusgnojkrap / algorithm Goto Github PK

The preparation process for MFCC.

1. Preemphasis : y(t)=x(t)-αx(t-1)
   The preprocessing process that boosts the energy of high frequency components.

By reinforcing relatively low-energy high-frequency components, the raw voice signal has a relatively even energy distribution across the entire frequency range.

Prevention of normative problems that may occur during Fourier transform.

Signal-to-noise ratio (SNR) improvement.

2. Framing
   The process of splitting the voice signal into small pieces in a short time.

Cut the frame into a certain time unit (frame_size=25 ms) but overlap the frame_stride=10 ms.

3. Windowing
   It is a technique of smoothing boundaries by applying a specific function to each frame.
   w[n]=0.54-0.46cos(2πn/(N-1))

4. Filter Banks : Mel Scale
   In data (pow_frames) with energy information for each frequency domain, the low frequency domain can be viewed in relatively detail compared to the high frequency domain.

5. Log-Mel Spectrum : LOG-FILTER BANKS
   We are interested in extracting characteristics that are important for human speech recognition, so we perform log transformation on the mel spectrum.
   filter_banks = 20 * np.log10(filter_banks)

6. MFCCs

The efficiency of the DFT algorithm is very low at n^2.
To improve this, we tried using the FFT algorithm.
The algorithm can achieve the efficiency of nlogn.

While implementing the above FFT algorithm, signal data sometimes flies away from the recursive part, and as a result, data is not accumulated in the array, so we are trying to figure it out.

locate : algorithm\machineLearning\newSTT/jongfft.js
function fft

Found an error during fft recursive that cannot return the array.

Real part calculation error during complex number calculation