Evaluating YouTube's Beta Automatic Audio Transcription: A Critical Analysis (Full Transcript)

Speaker 1: This is the second of two tests that I am performing with YouTube captioning. This time, I am evaluating whether the beta version of the automatic audio transcription works well or not. I suspect that the answer is that it is performing significantly worse than the first test where I uploaded the transcript did. As I promised the viewers who watched the first test, I am going to try and explain why I think that this is true. When I first began using speech recognition programs, I assumed that they knew English grammar, that they were using rules of English grammar to try and interpret what it was that I said. It turns out that this is not true. In fact, they know math, not English. To put it another way, they are using a statistical model to try and predict which word or phrase you are going to say next. Statistical models in the context of speech recognition are going to work a lot better if you are using a speaker dependent approach. This is why you have to train Dragon NaturallySpeaking and Windows Speech Recognition. Because when you train the speaker, when it is dependent on the particular speaker, the statistical performance is going to be much higher than if it has no additional data on the speaker. What Google is trying to do is to create a speaker independent system. Speaker independent systems are not ready to be commercialized. If they were, there would already be companies out there selling speaker independent speech recognition programs. If there are any such companies, they remain hidden. The people who do Google's automatic speech recognition software are smart enough to know that speaker independent speech recognition is not ready to be commercialized. They know it's terrible. I suspect that what they are redoing with the beta version, especially since they specifically call it an experimental program, is using it to collect data. It's quite possible that the words I am speaking to you right now are being recorded in a database somewhere, so that Google can analyze that and use them to try and create a truly speaker independent speech recognition program.