Understanding and Utilizing Biased Language Models (Full Transcript)

Speaker 1: Hello, this is Daniel Povey, and today we're going to ask him what are biased language models? Okay, a biased language model is a language model that's mostly estimated from the specific utterance or recording that you're trying to recognize. So it's something that you can estimate when you have the transcript available. And you normally do it for data cleanup or alignment purposes. So the idea is if someone gives you a transcript, and you're not sure if it's correct, or you're not sure if it's the transcript for that utterance, then you build a biased language model on that transcript that mostly has probability mass just for that sequence. And you do data alignment with that graph from that language model, and you see whether it recognizes the same utterance, you know, you look to see if that same sequence is the re, or maybe you cut out parts where it didn't align, because those are probably wrong. Follow up question, do you build biased language models per sentence? I mean, often you would, you normally you would build them at the level of however you got the transcript. So if you got the transcripts in, let's say, one file that covers the whole recording, then you normally build a biased language model at that level. Or if you got them for individual segments of the recording, then you get them per segment. Often these things don't necessarily correspond to what we would think of as a sentence. Okay, thank you. Okay, bye.