This work has successfully achieved a robust, fast and accurate pitch detection system called the Harmonic Grouping Pitch Detection. This system is able to perform pitch detection on a wide variety of signals such as speech, signing, whistling and musical instruments. On a 1.5Ghz AMD processor, the running time is 10x faster than real time. Using the CSTR database and the GPE measure of accuracy we have shown that the accuracy of Harmonic Grouping pitch detection is higher than other common systems.; The front-end of Harmonic Grouping pitch detection has been designed to match the front-end of state of the art speech recognition systems. Therefore, the computation requirements such as windowing and FFT calculation can be shared if the two systems are combined into a single application. This feature makes Harmonic Grouping an ideal choice for utilizing pitch information in a speech recognition system. Finally, two methods for utilizing pitch information in speech front-ends are presented to improve the recognition accuracy. These methods are: "Pitch-dependent models" and "Harmonic Density Normalization (HDN)". These methods can be utilized together in a speech recognition system and are shown to improve the recognition accuracy.
展开▼
