Linear Predictive Coding (or “LPC”) is a method of predicting a sample of a speech signal based on several previous samples. Similar to the method employed by the cepstrum, we can use the LPC coefficients to separate a speech signal into two parts: the transfer function (which contains the vocal quality) and the excitation (which contains the pitch and the sound).

We can predict that the nth sample in a sequence of speech samples is represented by the weighted sum of the p previous samples:

s ^ = ∑ k=1 p a k s[n−k] s ^ = ∑ k=1 p a k s[n−k]

The number of samples (p) is referred to as the “order” of the LPC. As p approaches infinity, we should be able to predict the nth sample exactly. However, p is usually on the order of ten to twenty, where it can provide an accurate enough representation with a limited cost of computation. The weights on the previous samples (ak) are chosen in order to minimize the squared error between the real sample and its predicted value. Thus, we want the error signal e(n), which is sometimes referred to as the LPC residual, to be as small as possible:

e[n]=s[n]− s ^ [n]=s[n]− ∑ k=1 p a k s[n−k] e[n]=s[n]− s ^ [n]=s[n]− ∑ k=1 p a k s[n−k]

We can take the z-transform of the above equation:

E(z)=S(z)− ∑ k=1 p a k S(z) z −k =S(z)[ 1− ∑ k=1 p a k z −k ]=S(z)A(z) E(z)=S(z)− ∑ k=1 p a k S(z) z −k =S(z)[ 1− ∑ k=1 p a k z −k ]=S(z)A(z)

Thus, we can represent the error signal E(z) as the product of our original speech signal S(z) and the transfer function A(z). A(z) represents an all-zero digital filter, where the ak coefficients correspond to the zeros in the filter’s z-plane. Similarly, we can represent our original speech signal S(z) as the product of the error signal E(z) and the transfer function 1 / A(z):

S(z)= E(z) A(z) S(z)= E(z) A(z)

The transfer function 1/A(z) represents an all-pole digital filter, where the ak coefficients correspond to the poles in the filter’s z-plane. Note that the roots of the A(z) polynomial must all lie within the unit circle to ensure stability of this filter.

The spectrum of the error signal E(z) will have a different structure depending on whether the sound it comes from is voiced or unvoiced. Voiced sounds are produced by vibrations of the vocal cords. Their spectrum is periodic with some fundamental frequency (which corresponds to the pitch). Examples of voiced sounds include all of the vowels. Unvoiced signals, however, do not have a fundamental frequency or a harmonic structure. Instead, they are just white noise.

In speech processing, computing the LPC coefficients of a signal gives us its ak values. From here, we can get the filter A(z) as described above. A(z) is the transfer function between the original signal s[n] and the excitation component e[n]. The transfer function of a speech signal is the part dealing with the voice quality: what distinguishes one person’s voice from another. The excitation component of a speech signal is the part dealing with the particular sounds and words that are produced. In the time domain, the excitation and transfer function are convolved to create the output voice signal. As shown in the figure below, we can put the original signal through the filter to get the excitation component. Putting the excitation component through the inverse filter (1 / A(z)) gives us the original signal back.

Figure 1: Using Linear Predictive Coding to separate the two parts of a speech signal: transfer function and excitation.

A Voice Conversion Algorithm

We can perform voice conversion by replacing the excitation component from the given speaker with a new one. Since we are still using the same transfer function A(z), the resulting speech sample will have the same voice quality as the original. However, since we are using a different excitation component, the resulting speech sample will have the same sounds as the new speaker.

In speech processing, a process called pre-emphasis is applied to the input signal before the LPC analysis. During the reconstruction following the LPC analysis, a de-emphasis process is applied to the signal to reverse the effects of pre-emphasis.

Pre- and de- emphasis are necessary because, in the spectrum of a human speech signal, the energy in the signal decreases as the frequency increases. Pre-emphasis increases the energy in parts of the signal by an amount inversely proportional to its frequency. Thus, as the frequency increases, pre-emphasis raises the energy of the speech signal by an increasing amount. This process therefore serves to flatten the signal so that the resulting spectrum consists of formants of similar heights. (Formants are the highly visible resonances or peaks in the spectrum of the speech signal, where most of the energy is concentrated.) The flatter spectrum allows the LPC analysis to more accurately model the speech segment. Without pre-emphasis, the linear prediction would incorrectly focus on the lower-frequency components of speech, losing important information about certain sounds.

Deng, Li and Douglas O”Shaughnessy. Speech Processing: A Dynamic and Optimization-Oriented Approach. Marcel Dekker, Inc: New York. 2003.

Gold, Ben and Nelson Morgan. Speech and Audio Signal Processing: Processing and Perception of Speech and Music. John Wiley and Sons, Inc: New York. 2000.

Lemmetty, Sami. Review of Speech Synthesis Technology. (Master’s Thesis: Helsinki University of Technology) March 1999. http://www.acoustics.hut.fi/~slemmett/dippa/thesis.pdf.

Markel, J.D. and A.H. Gray, Jr. Linear Predition of Speech. Springer-Verlag: Berlin. 1976.