Understanding I-vectors and X-vectors in Speaker Recognition (Full Transcript)

Speaker 1: Hello, this is Daniel Povey, and today we're asking him what's the difference between I-vectors and X-vectors? Okay, so I-vectors and X-vectors are both concepts from speaker recognition, meaning like speaker identification. So it's basically a fixed dimensional vector of, let's say, the dimension 256 or 512 or something like that. It's supposed to represent the information about the speaker, but the original thing about I-vectors was that you extract an I-vector from just a recording, and it contains information about both the speaker and the kind of recording conditions. And then you use other methods to separate those two sources of variation, like PLDA and stuff. But for reality purposes, we mostly use I-vectors for a very basic form of speaker adaptation, so that when we train a neural network, we input the I-vector as a kind of extra input to the neural network, and it helps it to adapt. And actually, for the most part, it just acts, it has a similar effect to just like mean normalization or something like that, because it can use the I-vector to figure out what's roughly the mean of the input features. So actually, in the end, I kind of regretted putting the I-vector stuff in, because you can get most of the improvement just from giving it the mean of the features up till the present point. So anyway, so that's what I-vectors are. Now X-vectors is a kind of a neural net version of I-vectors, where you basically train a neural net to discriminate between speakers, and inside the neural net, there's some kind of embedding layer that's just before the classifier, and you call that the X-vector. So you can extract, basically, it's a way of extracting a fixed dimensional feature from an utterance. Now, the thing with both I-vectors and X-vectors is that to train the classifier effectively, to train the system that extracts the I-vector or the X-vector, you need a very huge amount of data. So for I-vectors, ideally, you want 1,000 hours or something, if it's for speaker identification purposes, and for X-vectors, ideally, you want something like 10,000 hours, which is a bit ridiculous. Now, for speech recognition, it's not as critical. So it's fine if you have just 10 hours or 100 hours, because we're not really using it for speaker identification. We're just using it for a basic form of adaptation. So it's not so critical. OK. So does Kali use X-vectors at all? Well, there are speaker recognition recipes in Kali. Like if you look at SRE16, things like that. That's not for speech recognition, though, because there's no advantage of X-vectors over I-vectors for its application to speech recognition. We're just using it, like I said, for basic adaptation. And we don't really need all of that discriminating power of X-vectors. So answer is, we're using it only for speaker recognition. Thank you. Thank you. Bye. Bye.