Innovative Research
Progress Report and Upcoming Research Project Plan
April 4, 2022
This section includes an outline of our progress so far. Included here is challenges and adjustments made to our initial project proposal along with collected information and explanations of implemented Digital Signal Processing Tools. Also included is a list of future ideas and tasks to be completed to reach our goal of implementation for the Identification Device.
Progress Report Items
This section includes an outline of our progress so far. Included here is challenges and adjustments made to our initial project proposal along with collected information and explanations of implemented Digital Signal Processing Tools.
Data Collection Portion
Progress Summary and Encountered Challenges
Upcoming Tasks
Implemented DSP Tools
Progress Summary
Includes Encounter Challenges and Actions for Overcoming
During our initial weeks working on the project, we have finalized formatting our data set, along with progress on filter design. Finalization of our data set began with selecting the species of dolphins and whales our algorithm would focus on identifying. Then carefully selected raw audio signals of each species' voice were downloaded and imported into our audio library. Using these signals, we were able to begin the filtering process to determine the optimal filter for removing the echo of the underwater effect. Firstly, we applied the main filters covered in class which include: Lowpass, Highpass, Bandpass, Bandstop, and Notch Filters. Each of these filters were applied to an audio signal to determine which would provide the most effective outcome. Challenges were faced when carrying out the filter applications because the frequency ranges of the different mammals were very similar. We also faced great difficulty in eliminating water ripples and the underwater echoes. That being said, we shifted our focus a bit to using a Digital Thresholding Technique. This technique allowed us to eliminate lower frequencies entirely which would account for those of the water. However, with our applied Threshold of value 1.5, we are also sacrificing lower frequencies of the mammal voices.
In addition, there has also been major progress on the implementation of the classifying algorithm for our device. We are conducting two trials which include training for both filtered and unfiltered data. The outcomes from these trials will provide us with an accuracy percentage which can determine which filter was most effective for our data sets. Regarding challenges for the classifier, there were initial difficulties selecting which features to move forward with our algorithm training. There were initial challenges deciphering useful information from the larger audio signals. This prompted us to carry out a standardization process with the data sets, to transform all of the audio signals into a useful form. This transformation is included in our “cutter” phase of the system diagram.
The cutter phase begins by reading in multiple samples for each of the ten selected species. From here, the program resamples each of those to get a common frequency of 44kHz. After the resampling, the signals are buffered into lengths of .5 seconds to increase the data set available to train the data. This allowed us to condense large audio signals into a smaller set of distinguishable features. Each of the samples in this new set then go through an evaluation stage to determine the usefulness of the audio. Any signal containing over 10% of zero padded values, are then discarded from the data set. This elimination avoids biasing the classifier by removing repetitive data. This will prove to be more helpful later on when we begin the testing process of our algorithm. Moving forward, our major task will be the developing and testing of the machine learning classifier algorithm.
DSP Tools
We decided to pass all of our audio data through a filter to reduce the background noise introduced by water waves, splashes, and bubbles. We also targeted the propeller hum from ships and artifacts in the recordings to apply to our overarching idea of the mammal identifier device. This allowed the filtered audio data to focus more on the vocals of the mammals. The thought was that this specified filtering would help the classifier achieve better predictions and accuracy results.
Initially we attempted applying the standard threshold filter utilized from lecture. The filtered audio output from this filter sounded decently good. The background noises sounded reduced and the mammal vocals were still audible. However, this filter had the complete opposite effect we were looking to apply. The threshold filter actually cut off the higher frequencies which reduced the mammal vocals and made the background noise more prominent. A visual representation of this is represented in Figure 1.
Figure 1: Spectrogram of Pilot Whale with Applied Standard Threshold
This plot compares the audio signal after applying the standard threshold filter. The top portion shows the spectrogram of the unfiltered signal. The bottom portion demonstrates the spectrogram of the filtered signal. As you can see, the standard threshold limit that is cut off from all the frequencies above this value. This filter removed the frequencies of the vocals and left in the lower frequencies filled both the background noise and vocals.
The first digital signal processing tool utilized for our project was the spectrogram. We plotted spectrograms for every species in our dataset using the training audio. The spectrogram allowed us to visualize the frequencies of the vocals produced by each mammal with respect to time. We could see when a mammal would make a noise and the pattern of the noise. For example, the pilot whale would make sounds that looked like “chirps”, which started at a lower frequency and increased. This is visually represented in Figure 2. The spectrogram also showed the frequencies of the unwanted background noise. We noticed most of the background noise was located in the low frequencies below 2kHz and most of the mammal vocals were above that designated threshold. Depending on the behavior of the classifier with the filtered data, adjustment of this selected frequency might be necessary to increase accuracy ratings. Our shortcomings with the standard threshold applications caused us to explore the filters and the filter designer in MATLAB.
Figure 2: Spectrogram of Original Pilot Whale Audio Signal
The top figure plots the amplitude of the audio signal versus time. As you can see, there are three segments that show higher amplitudes. These segments correspond to when the pilot whale vocalizes. The bottom plot displays the spectrogram of the audio signal.
When beginning the filtering process, the goal was to apply a filter in the hopes of reducing the unwanted background noise to emphasize the vocals of the mammals. Utilizing the information we obtained from the spectrograms, we thought about applying a high-pass filter. The high-pass filter would allow for frequencies above 2kHz to pass through and block the frequencies below. Next we thought about applying a bandstop filter. This filter would allow us to isolate the high frequency vocals of mammals like dolphins, but also maintain lower frequency vocals of mammals like whales. The middle frequencies contain unnecessary information and unwanted background noise and therefore could be eliminated. Currently, we settled on using a high-pass filter to avoid filtering our data too much. Other filters, such as the bandstop filter, might get explored more during the training and testing process of our classifier.
When beginning the implementation of the high-pass filter, we utilized the highpass command in MATLAB. This function allowed us to filter out the part of the signal specified at the cut-off frequency provided. Since our scenario was very dependent on the voices of the mammals, we were motivated to utilize the Filter Designer. This app allowed us to graphically specify the parameters of a filter and visualize it. It also generated the code which created the filter which we were able to apply to our signal. The plotted spectrogram of the filtered audio signal was then compared to the unfiltered signal in Figure 3.
Figure 3: Spectrogram of Pilot Whale with Applied Highpass
This plot is similar to Figure 2, but just has an additional plot of the spectrogram for the applied highpass filtered audio signal at the very bottom of the figure.
Upcoming Project Tasks
Task 1
The next task for the project will be beginning the training process for the algorithm. This process will include determining the different features of the audio to compare which will yield us with the highest accuracy of classifying a mammal species.
Task 2
Based on the results from the algorithm training, the next task will be determining the optimal filter design. Using the selection of narrowed down filters and applying them to the system, the most effective filter will be selected based on the highest accuracy of the classifier.
Task 3
The final task for our project will be to test the classifier on randomized audio signals from our data collection. When inputting these signals, the system should determine the species based on the frequency values of the signal.