Dorsa Zeinali Email & Phone Number

Boston, Massachusetts, United States

Pittsburgh, Pennsylvania, United States

• Independently learned background on speech recognition through thorough research o Supplemented by Prof. Shinji Watanabe’s course: “Speech Recognition and Understanding”
• Learned the ESPnet speech recognition framework and used it to create a speech recognition model for Bengali, which was then added to the toolkit on GitHub
• Worked on a project for creating a benchmark and a tool for language-universal alignments, which can also be used for data selection and cleaning o The data available was 2TB of audios and transcriptions from YouTube.
• Continued work on project started at the JSALT workshop o Trained several models on the synthetic data, and achieved improvement in performance vs training solely on monolingual data o Created different kinds of.

Baltimore, Maryland, United States

• Participated in a 2-week summer school on speech and language technology
• Took the lead on a project which aimed to improve automatic speech recognition on code-switched speech when only monolingual training data is available by generating synthetic code-switched training examples
• Researched background in the topic and collaborated with others on the team to come up with a method of generation
• Fully implemented the pipeline to efficiently generate audio examples
• Learned valuable skills such as thorough literature review, independent problem-solving, and how to approach problems methodically and scientifically
• Presented the progress of the project weekly to the rest of workshop and presented the results of the project during the live-broadcasted JSALT 2022 Closing Presentations

Pittsburgh, Pennsylvania, United States

• Created custom machine learning models to predict opioid cravings from wearable data
• Created a set of gradient-boosted-tree based classifiers to clean mislabeled data
• Identified optimal features from timeseries data using a modified Markov Chain Monte Carlo,accounting temporal dependencies
• Increased performance of classifiers by 5-10% using a combination of hyperparameter tuning, data cleaning, and optimal feature selection

Pittsburgh, Pennsylvania, United States

• Using Python, fine-tuned a Factorized-Multimodal-Transformer to classify didactic interactions between parents and children inorder to eventually detect signs of mental illness
• Used forced-alignment tools to align audio data and their transcripts
• Extracted features from the transcripts in order to feed into the model
• Through preparing for the project, obtained valuable knowledge about state of the art Deep Learning methods and models

Chicago, Illinois, United States

• Using Python, manually classified numerous movie reviews into positive and negative for training
• First used a Naive-Bayes Classifier on the unlabeled with a given bank of positive and negative words.
• Later used a CNN and a RNN to classify the unlabeled data to account for the significance of the order of the words.

Evanston, Illinois, United States

• Visited multiple dermatology practices and photographed images of skin cancers.
• Worked with a post-doc on an image recognition program to detect skin cancers based on image data.

Fairfax, Virginia, United States

• Used Autodesk Maya and Matlab to model extraocular muscles in three dimensions.
• Worked with a PhD student to develop a model for a biomechanical simulator of strabismus.

Bachelor Of Science - Bs, Computer Science

Activities and Societies: Alpha Chi Omega

High School Diploma

Education record

Doctor Of Philosophy - Phd, Computer Science