Linear Prediction and Cross Synthesis

Overview

Linear prediction is a method used to estimate a time-varying filter, often as a model of a vocal tract. Musical applications of linear prediction substitute various signals as excitation sources for the time-varying filter.

This mini-project will give you chance to experiment with different excitation sources and linear prediction model parameters. In addition, you will learn about cross-synthesis. Finally, you will find out how to calculate the excitation sequence associated with the set of filter coefficients.

Implement your project in HTML. Please install your presentation in a subdirectory called ‘miniproj8’ in the public area of your AFS account.

Deliverables

Sheet of paper containing the URL to the top level page of your project. Be sure that your HTML project includes the following:

Part 1 – Excitation Sources

lp2.m and blp.m contain the code I used in class to demonstrate linear prediction. Experiment with the following types of excitation signals:

1. white noise (see randn)

2. square wave (see square)

3. sawtooth wave (see sawtooth)

4. bandlimited pulse

5. sinewave

Present your results. Include your soundfile, spectrogram, signal type and relevant parameters, and brief discussion.

Part 2 – Block Size and Filter Order

Vary the block size and filter order parameters. Study the effect of each parameter, and discuss your results.

Part 3 – Cross-Synthesis

“Cross-synthesis” denotes the case where the spectral envelope of one signal (e.g., speech) is applied to another signal (e.g., a musical instrument). Find or create a speech signal, and use it to generate a time-varying filter. Find or create a music signal, and use it as the excitation. The sound files should have the same sampling frequency. You can trim the length of one of the files to ensure that both files have the same number of samples.

Repeat for a second set of signals. You might try cross-synthesizing two different speech signals or two different music signals.

Show your results, particularly the spectrograms of the two original signals and the spectrogram of the output signal.