Jafar Howe: applying machine learning to analyze and detect different language accents

Aug. 1, 2022

Jafar Howe, 22, Class of 2023


Howe is a computer science major who is pursuing the undergraduate certificate from the Center for Statistics and Machine Learning (CSML) and another in engineering and management systems.

Undergraduate Work:

Howe is an international student from Trinidad and Tobago and his first language is English. His speech is accented with a Caribbean lilt, which has made it difficult for himself to be understood by voice-related software.

“When I first arrived in America and got an iPhone, Siri had trouble picking up what I had to say. I had to speak differently in order to be understood,” he said. “What I have found is that it’s not just Siri, but other voice-related software programs usually have trouble deciphering or processing different accents.”

This stumbling block poses a problem, Howe said. English has become the most spoken language in the world, resulting in many different accented versions of English.

“Voice-related software should be as inclusive as possible and recognize and accept these various accents instead of requiring users to code switch, like myself,” he said.

Code switching is the act of tailoring your accent/speech according to your audience, a common practice among people whose accent deviates from standard American English, Howe said.

“Essentially, the models that we have currently are mostly trained on American accents, which makes it much more difficult for people who have different accents to interact with voice assistants such as Siri,” he said.

For his independent project for CSML, Howe decided to develop an English language accent classifier, basically a tool to accurately detect different English accents. Howe did this by converting speech samples into a time-frequency spectrogram, a visual representation of audio. The spectrograms, which show changes in frequencies over time, served as inputs for a convolutional neural network called AlexNet. After training the neural network on a dataset of over 2,100 speakers from over 150 countries, Howe evaluated his accent classifier and found it achieved 83% accuracy.

In the future, he would like to add more non-native English accents to the training data set, Howe said.

“This was a fun project to do because I not only got to increase my experience using neural networks, but I also had the opportunity to work on something that was personally meaningful to me,” Howe said.

In addition to his academic work, Howe did a stint as a software engineer intern at Bloomberg. This summer, he is doing another software engineer internship at MongoDB, a software company known for its database by the same name.

When he graduates next year, Howe said he does not have exact plans yet, but he most likely will be entering the software engineering industry. Eventually, he would like to create a startup.

Extracurricular Activities:

Howe is a member of Dorobucci Dance, an African dance company on campus.

For Fun:

Howe enjoys dancing, working out, table tennis, watching TV, and hanging out with his friends.