This document describes a project on speech emotion recognition. It includes an abstract, introduction, flowchart, code to run adaptive noise cancellation algorithms, and results. The code reads an audio file, applies Fourier transforms, filters the signal using butterworth filtering, and achieves 85% accuracy in removing hissing noise. Instructions are provided on running the code and interpreting the wave file parameters.
This document describes a project on speech emotion recognition. It includes an abstract, introduction, flowchart, code to run adaptive noise cancellation algorithms, and results. The code reads an audio file, applies Fourier transforms, filters the signal using butterworth filtering, and achieves 85% accuracy in removing hissing noise. Instructions are provided on running the code and interpreting the wave file parameters.
This document describes a project on speech emotion recognition. It includes an abstract, introduction, flowchart, code to run adaptive noise cancellation algorithms, and results. The code reads an audio file, applies Fourier transforms, filters the signal using butterworth filtering, and achieves 85% accuracy in removing hissing noise. Instructions are provided on running the code and interpreting the wave file parameters.
This document describes a project on speech emotion recognition. It includes an abstract, introduction, flowchart, code to run adaptive noise cancellation algorithms, and results. The code reads an audio file, applies Fourier transforms, filters the signal using butterworth filtering, and achieves 85% accuracy in removing hissing noise. Instructions are provided on running the code and interpreting the wave file parameters.
1. Abstract 2. Introduction 3. Flow chart 4. Running code 5. Training and testing input 6. Accuracy achieved 7. Instructions for running the code ABSTRACT Virtually all audio recordings contain some amount of noise. This noise may join audio signal during recording process or due to long media storage. To produce best quality audio recordings these unwanted audio noises must be removed to the greatest extent possible. Noise cancellation of audio signal is key challenge problem in Audio Signal Processing. Since noise is random process and varying every instant of time, noise is estimated at every instant to cancel from the original signal. There are many schemes for noise cancellation but most effective scheme to accomplish noise cancellation is to use adaptive filter. Active Noise Cancellation (ANC) is achieved by introducing “anti- noise” wave through an appropriate array of secondary sources. These secondary sources are interconnected through an electronic system using a specific signal processing algorithm for the particular cancellation scheme. In this paper, the three conventional adaptive algorithms; RLS(Recursive Least Square), LMS(Least Mean Square) and NLMS(Normalized Least Mean Square) for ANC are analyzed based on single channel broadband feed forward. For obtaining faster convergence, Normalized Least Mean Square (NLMS) algorithm is modified and associated extended algorithm under Gaussian noise assumption. Simulation results indicate a higher quality of noise cancellation and more minimizing mean square error (MSE). INTRODUCTION Audio noise reduction system is that kind of system which is helpful to remove unwanted noise from speech signals. Audio noise reduction can be classified into two kinds. Complementary type and non-complementary type. One cannot effectively use speech communication in an environment in which the background noise level is too high. It interferes with the original message and corrupts the parameters of the message signal. Noise can be defined as an unwanted signal that interferes with the communication or measurement of another signal. A noise itself is an information-bearing signal that conveys information regarding the sources of the noise and the environment in which it propagates. The types and sources of noise and distortions are many and include: (i) electronic noise such as thermal noise and shot noise, (ii) acoustic noise emanating from moving, vibrating or colliding sources such as revolving machines, moving vehicles, keyboard clicks, wind and rain, (iii) electromagnetic noise that can interfere with the transmission and reception of voice, image and data over the radio-frequency spectrum, (iv) electrostatic noise generated by the presence of a voltage, (v) communication channel distortion and fading and (vi) quantization noise and lost data packets due to network congestion. FLOW CHART 4. Running code
from google.colab import drive
drive.mount('/content/drive') import numpy as np import IPython import wave import struct import matplotlib.pyplot as plt
# Above value is in complex number but we want absolute number
# This will give us the frequency we want magnitude = np.abs(frames_freq_domian) # Or ampliude ? phase = np.angle(frames_freq_domian) # Normally we are not interested in phase information, its only used in reconstruction.
print(magnitude.shape, phase.shape) print("The max frequency (highest magnitude) is {} Hz".format(np.where(magnitude == np.max(magnitude))[ 0][0]))
fig = plt.figure(figsize = (25, 6))
fig.suptitle('Original wav data')
ax1 = fig.add_subplot(1,3,1) ax1.set_title("Original audio wave / Spatial Domain") ax1.set_xlabel("Time(s)") ax1.set_ylabel("Amplitude (16 bit depth - Calulated above)") ax1.plot(frames_wave) ax2 = fig.add_subplot(1,3,2) ax2.set_title("Frequency by magnitude (Max at {} Hz) / Frequency Domain".format(np.where(magnitude == n p.max(magnitude))[0][0])) ax2.set_xlabel("Frequency (Hertz)") ax2.set_ylabel("Magnitude (normalized)") ax2.set_xlim(0, 44100) # we are not interested in rest ax2.plot(magnitude / nframes) # Normalizing magnitude ax3 = fig.add_subplot(1,3,3) ax3.set_title("[Unclipped]Frequency by magnitude (Max at {} Hz) / Frequency Domain".format(np.where(ma gnitude == np.max(magnitude))[0][0])) ax3.set_xlabel("Frequency (Hertz)") ax3.set_ylabel("Magnitude (normalized)") ax3.plot(magnitude / nframes) # Normalizing magnitude plt.show()
#Filtering the signal
from scipy import signal
from matplotlib import pyplot as plt def butter_pass_filter(data, cutoff, fs, order=5): nyq = 0.5 * fs # Nyquist frequency normal_cutoff = cutoff / nyq # A fraction b/w 0 and 1 of sampling rate print("normal_cutoff:", normal_cutoff, (data.shape[0] / 2) * normal_cutoff) # Tricky ? b, a = signal.butter(order, normal_cutoff, btype='high', analog=False) y = signal.filtfilt(b, a, data)
def _plot_graph(): # Get the filter coefficients so we can check its frequency response. # Plot the frequency response. w, h = signal.freqz(b, a, worN=8000) plt.subplot(2, 1, 1) plt.plot(0.5 *fs*w/np.pi, np.abs(h), 'b') plt.plot(cutoff, 0.5 * np.sqrt(2), 'ko') plt.axvline(cutoff, color='k') plt.xlim(0, 0.5*fs) plt.title("Filter Frequency Response") plt.xlabel('Frequency [Hz]') plt.grid() plt.show() _plot_graph() return y
# Filter requirements. order = 10 fs = framerate #* 6.28 # sample rate, Hz cutoff = 900 #* 6.28 # desired cutoff frequency of the filter, Hz
# Get the filter coefficients so we can check its frequency response.
ax1 = fig.add_subplot(1,4,1) ax1.set_title("[After Filter] Original audio wave / Spatial Domain") ax1.set_xlabel("Time(s)") ax1.set_ylabel("Amplitude (16 bit depth - Calulated above)") ax1.plot(y) ax2 = fig.add_subplot(1,4,2) ax2.set_title("[Before Filter] Original audio wave / Spatial Domain") ax2.set_xlabel("Time(s)") ax2.set_ylabel("Amplitude (16 bit depth - Calulated above)") ax2.plot(frames_wave, 'r') m = np.abs(fftpack.fft(y)) ax3 = fig.add_subplot(1,4,3) ax3.set_title("[After Filter] Frequency by magnitude") ax3.set_xlabel("Frequency (Hertz)") ax3.set_ylabel("Magnitude (normalized)") ax3.set_xlim(0, 44100) # we are not interested in rest ax3.plot(np.abs(fftpack.fft(y)) / nframes) # ax2.plot(range(0, 676864), m, 'g-', label='dataa') ax4 = fig.add_subplot(1,4,4) ax4.set_title("[Before Filter] Frequency by magnitude") ax4.set_xlabel("Frequency (Hertz)") ax4.set_ylabel("Magnitude (normalized)") ax4.set_xlim(0, 44100) # we are not interested in rest # ax2.plot(magnitude * 2 / (16 * len(magnitude))) ax4.plot(magnitude / nframes, 'r') plt.show() IPython.display.Audio(data=y, rate=44100)
5. Training and testing input
We have taken an audio file as input. This audio file contains some noise in it. To remove that noise we perform following operation on it: I. Convert into wav file. II. Reading wav file. III. Applying fourier. IV. Obtain waveform V. Filtering the signal. VI. Writing filtered signal back to file.
6. Accuracy achieved Our code will work for any other audio file which contains noise inherently. In this code we achieved 85% of accuracy. It removes almost all the hissing noise present in it.
7. Instructions for running the code
The following are the instructions should take while using this code: Channels: A standard wave file consist of 2 channels wav_file.setnchannels(1) ensures that this is a mono-sound. For stereo use: wav_file.setnchannels(2) Each Channel sample is stored as a 16bit Signed number Hence can have range (-32768 : +32767 ) Sample rate: Audio signals are analog, but we want to represent them digitally. Meaning we want to discretize them in value and in time. The sample rate gives how many times per second we get a value. The unit is Hz. The sample rate needs to be at least double of the highest frequency in the original sound [Based on Nyquist theorm which states that to retain freq f we should sample at rate 2f] otherwise you get aliasing. Human hearing range goes from ~20Hz to ~20kHz, so you can cut off anything above 20kHZ. Meaning a sample rate of more than 40kHz does not make much sense Bit-depth: The higher the bit-depth, the more dynamic range can be captured. Dynamic range is the difference between the quietest and loudest volume of an instrument, part or piece of music. A typical value seems to be 16 bit or 24 bit. A bit-depth of 16 bit has a theoretical dynamic range of 96 dB, whereas 24 bit has a dynamic range of 144 dB (source). Subtype: PCM_16 means 16 bit depth, where PCM stands for Pulse-Code Modulation. Filtering the signal: The response of a filter for a single frequency can be expressed as a complex number such that it represent 2 parts of wave (Amplitude and Phase), where the angle is the phase response of the filter and the absolute value the magnitude response. amplitude*e^(i*phase) = amplitude*cos(phase)+i*amplitude*sin(phase) If we havenot limited the output to 44100 Hz in Frequency domain plot, you would have notice that first half and second half are almost same, complex conjugate (due to above and below sinwaves). Since human being can only hear upto 44100Hz Frequency we are not interested in above ranges. Sound can be expressed in terms of waves, which in turn have frequency components. If we mask some frequency, we can remove noise. Which frequency to mask is the matter of trial and error. Since we dont have past data to train on it cannot be a ML based task. In simplicity we want to allow a certain range of freuency and mask other, like appying a rectangular window over the whole signal. Now sinc func representation in frequency domain is a rectangle, we can use it as an ideal filter. We just need to place this over our signal in freq domain and multiply (Amplitude modulation), then take a IFFT and get filtered singal back. But based on nature of sinc it is not possible since it never reaches zero and if we have to use this we need to calculate value till infinity. So we can take a approximation here and settle with something in between. Now above is just a explanation and their are many filters and their types. scipy comes with lots of them already built in and their is no need to dive in the maths, we just need to use the functions exposed. Filters never add any new frequency components to the sound. They can only scale the amplitudes of already existing frequencies.
