Sei sulla pagina 1di 3

Coursework 3

Informatics 2B

Vowel Recognition
Songfang Huang and Steve Renals Out: 28 February 2008 Submit by: 15:00 14 March 2008

Introduction

This coursework involves a simplied speech recognition task. The aim is to develop a classier that is able to correctly classify spoken vowels. You will use a dataset that contains a number of utterances of vowels extracted from words. The speech was recorded from a number of speakers. This is a speaker independent speech recognition task: the vowels in the test set were spoken by speakers who were not in the training set. This makes it a relatively hard task, particularly since the training set is rather small. The eleven vowels (indexed by integers from 0 to 10) and the words in which the vowel sounds were recorded are given in Table 1. Vowel i E a o u e Word heed head hard hoard whod heard Vowel I A Y O U Word hid had hud hod hood

Table 1: Words used in recording the eleven vowels. Your task involves designing and running a set of experiments to investigate the performance of four pattern recognition approaches on this vowel recognition problem. The approaches that you should investigate are: Multivariate Gaussian density estimation k-nearest neighbour Single-layer network with softmax outputs Multi-layer network with softmax outputs The use of MATLAB and the Netlab toolbox (http://www.ncrg.aston.ac.uk/netlab/, which is installed on DICE) is strongly recommended for this coursework.

Coursework 3

Informatics 2B

Data

The vowel data used in this coursework was collected by David Deterding [Deterding89], who recorded examples of the eleven steady state vowels of English spoken by fteen speakers. Each speaker uttered each of the eleven vowels six times. Each vowel utterance, is represented by a 10-dimensional oating point feature vector. (The features were derived from linear predictive coecients (LPC), but you do not need to understand exactly how the feature vectors were obtained.) For more detailed information about the Deterding vowel database, please see http://archive.ics.uci.edu/ml/ datasets/Connectionist+Bench+(Vowel+Recognition+-+Deterding+Data). For this coursework there are two data les: vowel-train.csv containing vowels from ve male and ve female speakers for training (660 examples in total), and vowel-test.csv containing the other three male and two female speakers for testing (330 examples in total). The data in the les is in comma separated value (.csv) format. Each line holds one vowel sample. The rst 10 comma separated values hold the oating point feature vector, while the 11th value on each line holds the class index (ranging from 0 to 10) of the vowel. In MATLAB you can use the csvread function to easily read in CSV les. You can obtain the vowel data les from:
http://www.inf.ed.ac.uk/teaching/courses/inf2b/coursework/cwk3.html

To unpack the data, run the following command on DICE: tar zxvf inf2b-cwk3-data08.tar.gz

Approach

You should rst train a multidimensional Gaussian classier. You should separate the training data into individual classes and estimate the parameters of a multivariate Gaussian for each class from the training data, and evaluate these Gaussians on test data For the next part of the coursework, you should use a k-nearest neighbour classication approach. It will be useful to look at classication accuracy with dierent values of k. Note that the Netlab knn function is rather inecient! For the last part of the experiments, you should train single-layer and multi-layer networks for vowel recognition using gradient descent: both the single-layer network and the multi-layer network (multilayer perceptron) should use softmax outputs. You should observe behaviour with dierent learning rates and comment on it. For the multi-layer perceptron, you can also experiment with dierent numbers of hidden units. For each case you should compute the accuracy on the test set and analyze the results. It is strongly recommended that use the Netlab library for these experiments. You should work through MATLAB lab 3 (http://www.inf.ed.ac.uk/teaching/courses/inf2b/labs/matlab-lab3. pdf) before attempting the coursework.

Coursework 3

Informatics 2B

Tasks
1. Classify the vowels using a multivariate Gaussian classier. (a) Fit a multivariate Gaussian, using a full covariance matrix, to the training data for each vowel. Make sure you use only the training data at this stage. Then, assuming that the prior probabilities of each class are equal, classify the test vowels. (10 marks) (b) Repeat the experiment using spherical and diagonal covariance matrices for each class. Compare the classication results with the previous experiment. (5 marks) 2. Classify the same vowels using a k-Nearest Neighbour classier. Observe the eect of varying dierent values of k. (15 marks) 3. Train a single layer network with softmax outputs. Vary the learning rate and observe the eect on training time and classication accuracy. (15 marks) 4. Train a multi-layer perceptron network with softmax outputs. Vary the number of hidden units and the learning rate, and observe the eect on training time and classication accuracy. (15 marks) 5. Write it up, clearly and concisely in a 4-page report. (a) Describe the classication techniques used. (10 marks) (b) Compare the results of the classication methods in terms of accuracy and in terms of amount of computation. Analyse the eects of varying parameters. (20 marks) (c) What do you conclude? Which method is best and why? (10 marks)

Submission

Your report should be no longer than 4 pages. It should be formatted as 2 columns per page, and no A smaller than 10 point font. Suitable templates for this format are available in Word or LTEX: Word template:
http://www.inf.ed.ac.uk/teaching/courses/inf2b/coursework/is2005 word template.zip
A LTEX style le and template:

http://www.inf.ed.ac.uk/teaching/courses/inf2b/coursework/is2005 latex template.zip

You are required to submit your Matlab code, clear and suitably commented, and a pdf le containing your report. Collect everything together and submit as either a zip or a tar archive, using the following submission command: submit inf2 inf2b b3 filename-of-your-submission-archive The deadline for submissions is 3PM on Friday 14 March 2008. If you have any questions regarding this coursework, please post them to the inf2b newsgroup eduni. inf.course.inf2b, or mail s.f.huang@ed.ac.uk.

Potrebbero piacerti anche