A Comparisoat of Two Quanttization Techniques for Speech Spectral Parameters

FOR SPEECH SPECTRAL PARAMETERS John Leis Faculty of Engineering University of Southern Queensland Toowoomba, AUSTRALIA email: [email protected] Sridha Sridharan Signal Processing Research Centre Queensland University of Technology Brisbane, AUSTRALIA ABSTRACT The major contributor to the overall bit rate in low rate speech coders is the encoding of the short-term spectral parameters. Current schemes typically use 32-40 bits per coding frame to represent this information using scalar quantization techniques. Several researchers have recently investiged the use of vector quantization (VQ) for these parameters, and found that direct VQ requires a prohibitively large vector codebook. We report here on some experiments undertaken on a large number of speech frames involving the use of vector quantization for the line spectral frequencies. Speci cally, we set out to determine the in uence of the clustering algorithm used in the training process and the number of training vectors required to give good generalization and hence a robust codebook. 1. SPECTRUM CODING The well-known Linear Predictive Coding (LPC) model is normally used to remove short-term redundancies within a frame of speech. For the purposes of quantization (scalar or vector) the Line Spectrum Frequency (LSF, also referred to as Line Spectrum Pair or LSP) is used. Given the Linear Predictive Coding (LPC) model with coe cients ai, the LSF representation is found as in [3]. The spectral distortion is calculated using SD2 = 1 N N 1 Xn=0 20 log jS (n) j jŜ (n) j 2 (1) 2. RESULTS The algorithms compared were the Pairwise Nearest Neighbour (PNN) [1] and the Generalized Lloyd Algorithm (GLA) [2]. The speech was sampled at 16kHz and split into frames of 320 samples, giving 50 frames per second. A Hamming window was used with 10 LPC coe cients converted into 10 LSF's. The GLA and PNN training was implemented using the MATLAB system [7] and training of the VQ codebooks was conducted over approximately 32k speech frames of the \Train" section of the TIMIT DR-2 speech database [6]. Testing was carried out using the TIMIT DR-2 \Test" database using 10000 test vectors outside the training set. 9 bits, Av SD=3.3 dB 10 bits, Av SD=3.0 dB 11 bits, Av SD=2.8 dB 12 bits, Av SD=2.7 dB 13 bits, Av SD=2.5 dB