1. Introduction
In this homework, you are going to implement Structured SVM as a more general purpose decoder. In homework 1, we only consider frame to phone label probabilities. In homework 3, we use Structured SVM to model the phone transition probabilities, and see the whole utterance as one training data, instead of frame by frame training. Note this is different from homework 1 where homework 1 is a classification problem (output phone label), homework 3 is structure learning problem (output phone label sequence).
2. Problem Description
Goal
: Given input feature of each time frame, utilizes structured SVM to predict the whole utterance.
Input
: The raw wave files, the MFCC features, the FBANK features or any other derivations. (Exactly the same as homework 1)
Output
: 39 phone label sequence.
Firstly, the output label sequence is mapped to 39 phone labels (replace by a-z and A-M 39 English characters). Secondly, it should reduce replicated phones. Thirdly, the front end and back end silence is removed from the label sequence. Details descriptions are in the
[sildes](https://docs.google.com/presentation/d/1sXXfcODQ_jkDaxPDRyqpZ7ELBscnFlgV9Kl0qdrk9BA/edit?usp=sharing)
.
3. Implementation
You can use any methods with structured learning algorithm, e.g., HMM, structured SVM, CRF. You are encouraged to use any result or model from homework 1. Any other toolkit or library is required to ask TA for sure.