Yinan Zhou

• Scraped data scientist opening information within Canada from glassdoor.ca using Selenium, and extracted features and built a dataset.
• Conducted Exploratory Data Analysis (EDA) and created data visualizations with Seaborn to assess market demand for data-related positions, and identify essential skills for professional development strategies.
• Built machine learning models with custom loss function to estimate data scientist salary, and packaged it into a client facing API using Flask.

• Scraped rumratings.com using BeautifulSoup and saq.com using Selenium to construct a comprehensive rum dataset, and performed data cleaning and preprocessing.
• Conducted an in-depth analysis of the rum market, identified market gaps, and predicted potential opportunities for specific rum varieties to meet consumer demand.
• Developed a web-based interactive dashboard using Plotly and Dash to visualize the findings.

• Built a new dataset of 6,500 ancient Chinese poetries and corresponding meolodies using OMR techniques, whose lyrics are poetries from more recent times.
• Finetuned the template-to-melody module in the Telemelody model with the new dataset.
• Regenerated melodies using the finetuned model to restore the melody of poetry whose melody has been lost.

This project explores the performance of multimodal learning on Music Emotion Recognition. Unimodal and bimodal methods on audio and lyrical features were built and compared. Both middle fusion and late fusion were applied for bimodal methods. This result demonstrates that the combination of two feature domains can improve the MER performance.

This project implements a block-based dereverberation algorithm. The hyperparameter settings are explored on a singing voice recording. Moreover, the algorithm is also implemented on a short clip of a symphonic music piece by Mozart. It is found that this algorithm has limited effect on signals with high reverberation. The hyperparameter setting for high reverberation extraction is also discussed.

This project implements both Karplus-Strong algorithm and digital waveguide model to synthesize ukulele sound. A system is then built to play MIDI files with the resulting sound of two models.

This project introduces the basic concepts of cepstrum and cepstrogram. Their implementations in pitch estimation, spectral envelope estimation, and audio feature extraction are also covered.

• Used web crawler to collect MIDI files on YouTube with music emotion labels.
• Extracted MFCC, chords, and other features from music fragments.
• Chose and constructed appropriate neural networks.
• Trained the model to get the output value; calculated error and updated model parameters to reduce error.

• Preprocessed motion and speech data using techniques including wavelet filtering, pre-emphasis, windowing, and Fourier transform.
• Performed feature extraction, filtering, and dimensionality reduction on motion and speech data using quartile box charts and PCA algorithms.
• Classified data using KNN, GMM, and decision trees.