Many of my friends know that I have previously studied computer vision, communication channel sensing, and process mining, whereas very few of them know that my journey into artificial intelligence actually started with speech processing.
Near the end of 2007, when I was still an undergraduate student in Xi'an, I enrolled in an elective subject called "Speech Signal Processing," which taught the basics of speech coding and speech recognition. I have forgotten much of the speech coding material, but for speech recognition, I learned a very significant concept called a "instantiated stationary process." This concept set the foundation for using sliding windows in speech processing; within a small temporal window, a Fourier transform can be applied, which is known as the "Short-Time Fourier Transform." Then, the teacher introduced us to a mathematical model called the "Hidden Markov Model," which was probably the first model I ever studied in machine learning.
In 2009, when I started a master's degree course at QUT in Australia, I undertook a research project in the then Microsoft eResearch Centre with Paul. The team led by Paul was conducting analysis on animal sounds to understand the trends of eco-systems in Queensland. Though it was not for human speech, I gained valuable experience in audio processing.
Despite the fact that I later mainly conducted research in computer vision, the research group I was in still had some researchers working on speech and speaker recognition, and Sridha was the best among them.
One day, I discussed a problem with Sridha that I encountered in Gaussian Mixture Models (GMMs). It was about the covariance matrices becoming singular after some iterations during the training process.
"Yes, GMMs have this issue. This is a known limitation of the model, so I don't think it is your problem," he said.
"So you mean I shouldn't use it?" I asked.
"You can use it," Sridha said, “Every model in this area (machine learning) has its drawbacks, but there are always methods to solve them and apply them to our applications."
"I heard that Ahilan also used this model in speaker recognition. Is it because a speech signal is essentially a one-dimensional signal, and thus the covariance matrix is simpler?" I asked.
"No. Features for speech can also be high-dimensional vectors. There are many ways to deal with it," he said.
I then explained my thoughts on how to tackle it to him. "Do you believe it will work?" I asked.
"You can try," he said. "In our area, only experimental results can tell you if a method is working or not. If it is a method you thought of, you can write a paper if it works. If you fail in the experiments, at least you know it won't work."
"Yes, GMMs have this issue. This is a known limitation of the model, so I don't think it is your problem," he said.
"So you mean I shouldn't use it?" I asked.
"You can use it," Sridha said, “Every model in this area (machine learning) has its drawbacks, but there are always methods to solve them and apply them to our applications."
"I heard that Ahilan also used this model in speaker recognition. Is it because a speech signal is essentially a one-dimensional signal, and thus the covariance matrix is simpler?" I asked.
"No. Features for speech can also be high-dimensional vectors. There are many ways to deal with it," he said.
I then explained my thoughts on how to tackle it to him. "Do you believe it will work?" I asked.
"You can try," he said. "In our area, only experimental results can tell you if a method is working or not. If it is a method you thought of, you can write a paper if it works. If you fail in the experiments, at least you know it won't work."
Of course, the experimental results failed; otherwise, I would have written a paper on it. I knew my very good friend Ahilan was using a so-called "universal background model" based on GMMs for speaker recognition. That was why I often discussed his project with him. Soon, I found that the GMMs were trained on unlabeled data, and the datasets were massive. At that time, my understanding of machine learning had become much better. I had learned the concept of overfitting—a problem that occurs when a model has too many parameters to be trained on a small dataset. Then I realized that for GMMs, the majority of the parameters reside in the covariance matrices. This led me to think that what I had encountered was, in fact, a form of overfitting problem. For Ahilan's speaker recognition project, the labeled dataset was small, making the GMMs hard to train. However, a large scale of unlabeled data could be used to train the parameters to some extent. Then, there was a step called "enrollment" to finally adapt the model's parameters to the labeled data. I merrily went to chat with Ahilan about this, feeling as if I had discovered a treasure box on a remote island.
In 2014, I worked for half a year at a company where part of my work involved the development of speech recognition. I gained a lot of knowledge through this project. I found that accents in the training data significantly influenced the performance of the applications.
Recently, my father complained that Baidu Translate kept failing to translate his speech into English.
I said, "That's because your Mandarin is not standard or accurate."
I said, "That's because your Mandarin is not standard or accurate."
No comments:
Post a Comment