Takashi Ariyoshi

Noise eliminating apparatus and speech recognition apparatus using the same

A noise eliminating apparatus includes a first feature extracting block for individually extracting a feature quantity X(f) (f =1, 2, . . . , n) of an input speech containing a noise for n channels (n is an integer), and a second feature extracting block for individually extracting a feature quantity N(f) of the noise for the n channels. The apparatus also includes a noise eliminating block for deriving a first noise eliminating variable k(f) and a second noise eliminating variable d(f) from the feature quantity X(f) of the input speech containing the noise provided for each of the n channels and the feature quantity N(f) of the noise provided for each of the n channels and for generating a feature quantity S(f) of a noise-eliminated input speech in accordance with the following formula: s(f)=X(f)-k(f).N(f)-d(f).

Speech recognition processors using fuzzy pattern matching

A set of LSIs for speech recognition using binary time spectrum pattern (BTSP) and pattern matching based on fuzzy theory (fuzzy pattern matching) is described. The set offers both speaker-dependent and -independent voice-recognition capabilities by changing the external program ROM. In the speaker-dependent case it has up to 1000 in the vocabulary (recognition rate 95%), while in the independent case it has a 120-word vocabulary (recognition rate 92%). The set consists of two processors: one extracts voice features and generates BTSPs, the other has a 16-bit CPU core for recognition and a unit for calculating similarity. The LSI set needs external ROM and RAM. The former stores recognition, control, and other application programs while latter stores voice templates.<<ETX>>