Peter Veprek

Analysis, enhancement and evaluation of five pitch determination techniques

Abstract Speech classification into voiced and unvoiced (or silent) portions is important in many speech processing applications. In addition, segmentation of voiced speech into individual pitch epochs is necessary in several high quality speech synthesis and coding techniques. This paper introduces criteria for measuring the performance of automatic procedures performing this task against manually segmented and labeled data. First, five basic pitch determination algorithms (PDAs) (SIFT, comb filter energy maximization, spectrum decimation/accumulation, optimal temporal similarity and dyadic wavelet transform) are evaluated and their performance is analyzed. A set of enhancements is then developed and applied to the basic algorithms, which yields superior performance by virtually eliminating multiple and sub-multiple pitch assignment errors and reducing all other errors. Evaluation shows that the enhancements improved performance of all five PDAs with the improvement ranging from 3.5% for the comb filter energy maximization method to 8.3% for the dyadic wavelet transform method.

Compression of exception lexicons for small footprint grapheme-to-phoneme conversion

We present a method to reduce the memory footprint of a grapheme-to-phoneme conversion (G2P) module, without sacrificing accuracy. Since the G2P module is typically not 100% correct, it is common to augment the system with an exception lexicon - a list of words which the G2P does not handle correctly (and for which we require correct pronunciations), along with their corrected pronunciation. Since the size of the exception lexicon is one of the major limiting factors in reducing the overall size of the G2P module, we try to compress the exception lexicon. We suggest a novel compression method which is closely tied to the G2P conversion method. The idea behind this compression is that, even for words which are not transduced correctly, the decision trees generate a phonetic transcription which is close to the correct one. Therefore, it is sufficient to store only the correction in the exception lexicon. The correction information is represented in terms of corrections to the transduction process; it is thus able to take advantage of the knowledge gained from the training data regarding the probabilities of different corrections, and is used to obtain more efficient compression. An experiment showed that, by using this method, an exception pronunciation can be represented, on average, with less than 4 bits (a compression factor of 7, compared to the baseline representation).