Yoshiaki Kitazume

Speech recognition method and device

Speech sound recognition is made using a reduced number of speech parameter elements, e.g., five correlation coefficients rather than sixteen spectral coefficients. The five correlation coefficients are derived from comparison of the spectral coefficients of unknown or standard sounds against the spectral coefficients of five highly-separable vowel-like sounds. Then, unknown-sound correlation coefficients are compared with standard-sound coefficients for recognition.

Speech detecting method

Speech signal presence is decided if total signal power is above a first threshold, and if either low or high frequency components exceed thresholds as a large fraction of the total power. Total power is calculated as the zero-order auto-correlation coefficient, and fractional power of frequency components is calculated as the first-order partial auto-correlation coefficient.

Method and apparatus for registering standard pattern for speech recognition

In order to produce and store a pattern of reference words for use in a speaker-dependent speech recognition system, the system prompts the operator of the system to speak standard words in a predetermined sequence. For this purpose, a prestored standard word is spoken by the system with a predetermined length, power and rhythm, and the operator then repeats the standard words while attempting to simulate the same predetermined length, power and rhythm. The standard word repeated by the operator is detected and processed to determine whether it meets a certain resemblance criteria with respect to the standard word as spoken by the system. If the standard word repeated by the operator does not meet the resemblance criteria, the system repeats the same standard word to prompt the operator to try again; and, if the standard word repeated by the operator meets the resemblance criteria, it is stored as a reference word. This operation is repeated for each of the sequence of prestored standard words.

LSI implementation of a pattern matching algorithm for speech recognition

Many speech recognition systems contain the foUowing functional blocks: a voice input circuit, a feature extractor (analyzer), a unit for calculating the distance between input and standard patterns at every frame, a memory for storing standard patterns, a unit for matching whole word patterns (a pattern matching circuit), and a final decision and system control circuit. The pattern matching circuit is independent of the recognition algorithm and requires many conventional integrated circuits for implementation. Therefore, we decided to develop a custom integrated circuit for the pattern matching function. Due to its original architecture, our integrated circuit is able to execute in real time a continuous nonlinear matching process and is able to handle large volumes of data. This integrated circuit has been designed to be suitable for both isolated word recognition and connected speech recognition. The circuit has been tested in a pilot isolated-word and connected-speech recognition system. In this paper, the authors describe this integrated circuits specifications, architecture, and performance, as well as its application in the model system.

Conceptual system design for continuous speech recognition LSI

Speech recognition equipment is basically composed of these function blocks: a distinctive parameter extractor (analyzer); a calculator for the distance between input speech and standard patterns in each analyzing frame; a standard pattern memory; a calculator, to compare patterns, as a whole, using a DP matching process; a final judgement apparatus; and a system control. It is advisable that function blocks which are less effected by algorithmic change be designed as custom LSIs. Other function blocks, which are more effected, should be designed using general purpose signal processing LSIs and MSIs. A continuous speech recognition system has been designed from this philosophical base. In this system, a high speed signal processer (HSP) is adopted as an analyzer. The whole matching calculation block is designed as a custom LSI.

System design for large-vocabulary speech recognition LSI

Speech recognition systems generally depend upon pattern matching, thus the pattern matching circuit is indispensable in these systems. Hitachi designed an NL-LSI [1] with an original architecture and useful pattern matching function two years ago. The new LSI enhances the NL-LSIs functions and performance. The vocabulary size of this LSI is 512 words, eight times that of the NL-LSI. This LSI is able to calculate the Chebyshev norm and matching value. And this LSI has been designed to be suitable for both isolated word recognition and connected-speech recognition. In this paper, the authors describe this LSIs specifications, architecture and performance.