Identification of complex signals such as a phonetic sound in speech is accomplished by the combination of ratio values both between the amplitudes and frequencies of the resonances in the complex sound wave. The amplitude ratio between the signals detected from two resonances (e.g. formants) is derived as a null signal when they match a pair of signal-gain preadjustments during a given time period. This null-signal is also obtained when both input signals are absent during that time period. In order to avoid false null-signal indication, a sensing signal is derived from said input signals to indicate that the null-signal is not false.
An apparatus and method are provided for the recognition of speech produced by various vocal pitches capable of recognition and classification of speech articulation at a real time rate. The apparatus and method assume that articulation of a given sound in an individual's speech can be approximated as the output of a specific linear filter, corresponding to the condition of the individual's vocal tract at the time of articulation, in response to an input of one or more source impulses. The invention selects one of a library of sounds, in response to a speech waveform input, by means of a bank of vocal tract inverse filters, each of which is connected to the speech waveform input. Each vocal tract inverse filter has a complex Fourier transfer function that is the reciprocal of a particular vocal tract transfer function corresponding to a specific speech sound. Thus there is one vocal tract inverse filter for each speech sound as spoken by the particular individual. The assumed vocal tract filters are thus effectively in cascade with the vocal tract inverse filters of the invention so as to form an all-pass filter which can derive the original source waveform.
In the speech recognition apparatus disclosed herein, an audio signal is digitized and a succession of short-term power spectra are generated over a time interval corresponding to a spoken word. The short-term power spectra are frequency band equalized as a function of the peak amplitude occurring in each frequency band over the word interval. The changes in amplitude in each frequency band subjective time and then a limited number of frequency band equalized spectra are selected as representing equal intervals of subjective time so as to suppress variations in rate of articulation. The selected spectra are then non-linearly sealed in amplitude and transformed so as to maximize the separation between phonetically different sounds. By means of a maximum-likelihood method, the transformed selected spectra are compared with a data base representing a vocabulary to be recognized.
Recognition of human speech is carried out by storing a template for each unit of speech to produce a dictionary of stored words and phrases. A given speech signal is converted to produce a template which is compared to the stored template to find the closest comparison. The word or phrase corresponding to the identified template is produced and displayed to complete the recognition of the speech signal. The speech signal is processed to produce two separate frequency components. The frequency components are processed to produce a DC signal proportional to the frequency of the frequency component. The frequency components are also rectified to produce amplitude signals corresponding to the envelope of the frequency components. The products [F.sub.1 ][F.sub.2 ] and [A.sub.1 ][A.sub.2 ] and ratios F.sub.1 /F.sub.2 and A.sub.1 /A.sub.2 of the pairs of frequency and amplitude signals are produced to generate a plurality of relational signals which comprise the templates corresponding to each speech signal. In building the dictionary each speech sample is submitted a number of times to produce an average template value together with the variants for each data point.
An audio data input unit of an encoding device splits an audio data string into contiguous samples of audio data, and a transforming unit transforms the split audio data into spectral data in a frequency domain. A data dividing unit divides the spectral data into a lower frequency band and a higher frequency band at 11.025 kHz (f1) as a boundary. The spectral data in the lower frequency band is quantized and encoded by a first quantizing unit and an encoding unit. A second quantizing unit generates sub information indicating a characteristic of the spectral data in the higher frequency band, and a second encoding unit encodes the sub information. A stream output unit integrates the codes obtained by the first and second encoding units and outputs the integrated one. Here, f1 is a half or less of a sampling frequency f2 at which the audio data string is created.
