Speech recognition system which avoids ambiguity when matching frequency spectra by employing an additional verbal feature

• US 4,882,755 A
• Filed: 08/11/1987
• Issued: 11/21/1989
• Est. Priority Date: 08/21/1986
• Status: Expired due to Fees

• Current Assignee: OKI Electric Industry Company Limited
• Original Assignee: OKI Electric Industry Company Limited
• Inventors: Yamada, Youichi, Takahashi, Keiko
• Primary Examiner(s): Harcom, Gary V.
• Application Number:
  US07/084,107
• Time in Patent Office: 833 Days
• Field of Search: 381/41-50, 364/513.5
• US Class Current: 704/239
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence

Abstract

In a speech recognition system using normalized spectrum matching, a second feature pattern is calculated and compared with reference patterns. The similarity obtained as a result of the comparison is used to determine overall similarity, from which the recognition is made. The second feature pattern can be a spectrum variation pattern, a level decrease pattern, or a spectrum relative value pattern.

Claims

1. A speech recognition system comprising

a frequency analyzer which performs frequency analysis of the input speech into a number of channels, performs logarithmic conversion and extracts a frequency spectrum,

a voiced interval detector which detects voiced intervals on the basis of the said frequency spectrum,

a spectrum normalizer which determines a least square fit line for the frequency spectrum and normalizes the frequency spectrum with reference to the least square fit line to provide a normalized frequency spectrum pattern,

a spectrum reference pattern memory in which spectrum reference patterns are stored in advance,

a spectrum similarity calculator which calculates the similarity between the said normalized spectrum pattern and a spectrum reference pattern for each of a plurality of recognition categories, and

an identifier which, as the result of the spectrum similarity calculation, outputs the name of the recognition category which has the highest similarity,

the speech recognition system further comprising:

(a) a second feature pattern calculator which calculates a second feature pattern, said second feature pattern calculator (a) including a spectrum variation pattern calculator which, for each frame in a voiced interval, calculates for each channel, as the second feature pattern, a spectrum variation pattern which quantifies the degree and the direction of the transition between channels of the normalized spectrum with the advance of time around the said frame, and

(b) a second feature reference pattern memory in which second feature reference patterns have previously been stored, said second feature reference pattern memory (b) including a spectrum variation reference pattern memory in which spectrum variation reference patterns, as said second feature reference patterns, have previously been stored, and

(c) a second feature similarity calculator which calculates the similarity between the said second feature pattern and the second feature reference patterns with respect to each recognition category, said second feature similarity calculator (c) including a spectrum variation similarity calculator which calculates the similarity between the said spectrum variation pattern and the spectrum variation reference patterns with respect to each of the recognition categories, and in that

(d) in the identifier, the overall similarity is calculated for each of the recognition categories by reference to both the similarity of the spectrum and the similarity of the second feature, and the category giving the largest overall similarity is output as the recognition result.

2. A speech recognition system according to claim 1, wherein the said spectrum variation similarity calculator comprises

(c1) means for calculating the interframe distance in which, for each frame of a voiced interval, the distance between the said frame and the frame preceding the said frame is calculated,

(c2) means for determining a maximum variation frame in which a frame is determined to be a maximum variation fame if the interframe distance within a voiced interval is, in the said calculation, found to be maximum, and

(c3) means for extracting spectrum variation patterns in which, for a frame which has been determined to be the said maximum variation frame, the normalized spectrum values are calculated in each of four regions for each channel in the said frame, namely,

(c31) a region where the frame number is on the negative side with respect to the said frame and the channel number is on the negative side with respect to the said channel,

(c32) a region where the frame number is on the positive side with respect to the said frame and, the channel number is on the positive side with respect to the said channel,

(c33) a region where the frame number is on the negative side with respect to the said frame and in which the channel number is on the positive side with respect to the said channel,

(c34) a region where the frame number is on the positive side with respect to the said frame and in which the channel number is on the negative side with respect to the said channel, and

values showing relationships between the normalized spectrum values in the regions (c31) and (c32), and the normalized spectrum values in the regions (c33) and (c34) are taken to be the spectrum variation pattern for the said frame and said channel, and

the values for the spectrum variation pattern for frames which were not determined to be the said maximum variation frame are taken to be 0 for every channel.

3. A speech recognition system according to claim 1, wherein the overall similarity is calculated by adding the similarity for the spectrum and the similarity for the second feature.

4. A speech recognition system comprising

a frequency analyzer which performs frequency analysis of the input speech into a number of channels, performs logarithmic conversion and extracts a frequency spectrum,

a voiced interval detector which detects voiced intervals on the basis of the said frequency spectrum,

a spectrum normalizer which determines a least square fit line for the frequency spectrum and normalizes the frequency spectrum with reference to the least square fit line to provide a normalized frequency spectrum pattern,

a spectrum reference pattern memory in which spectrum reference patterns are stored in advance,

a spectrum similarity calculator which calculates the similarity between the said normalized spectrum pattern and a spectrum reference pattern for each of a plurality of recognition categories, and

an identifier which, as the result of the spectrum similarity calculation, outputs the name of the recognition category which has the highest similarity,

the speech recognition system further comprising:

(a) a second feature pattern calculator which calculates a second feature pattern, said second feature pattern calculator (a) including a level decrease pattern calculator which, for each frame in a voiced interval, determines whether the particular frame is a voiceless frame in accordance with the level of input speech for the frame with respect to the maximum value of the input speech level, and calculates, as said second feature pattern, a level decrease pattern for the voiceless frame which quantitizes the level decrease relative to the maximum value of the input speech level in the particular voiceless frame,

(b) a second feature reference pattern memory in which second feature reference patterns have previously been stored, said second feature reference pattern memory (b) including a level decrease reference pattern memory in which level decrease reference patterns have previously been stored, and

(c) a second feature similarity calculator which calculates the similarity between the said second feature pattern and the second feature reference patterns with respect to each recognition category, said second feature similarity calculator (c) including a level decrease similarity calculator which calculates the similarity between the said level decrease pattern and the level decrease reference patterns with respect to each of the recognition categories, and in that

(d) in the identifier, the overall similarity is calculated for each of the recognition categories by reference to both the similarity of the spectrum and the similarity of the second feature, and the category giving the largest overall similarity is output as the recognition result.

5. A speech recognition system according to claim 4, wherein

said level decrease pattern calculator comprises

a voiced interval frame identifier which calculates, for each of the frames during speech input, the input speech level in the particular frame, and repeatedly performs, from the speech start point frame to the speech end point frame, the process of identifying the frame in question to be a voiceless frame if the input speech level is not more than 1/N of the input speech level maximum value among the frames from the speech start point frame to the frame in question, N being a predetermined positive integer, and

a level decrease pattern extractor taking, after detecting the speech end point, for each voiceless frame and for each channel, a value obtained by normalizing, by the input speech level maximum value, a value that in turn is obtained by subtracting, from the said input speech level maximum value, the frequency spectrum value at the voiceless frame and at the particular channel, to be the level decrease at the particular voiceless frame and at the particular channel, and for each frame that has not been identified to be a voiceless frame, "0" to be the level decrease at all the channels and at the particular frame.

6. A speech recognition system according to claim 4, wherein the overall similarity is calculated by adding the similarity for the spectrum and the similarity for the second feature.

7. A speech recognition system comprising

a frequency analyzer which performs frequency analysis of the input speech into a number of channels, performs logarithmic conversion and extracts a frequency spectrum,

a voiced interval detector which detects voiced intervals on the basis of the said frequency spectrum,

a spectrum normalizer which determines a least square fit line for the frequency spectrum and normalizes the frequency spectrum with reference to the least square fit line to provide a normalized frequency spectrum pattern,

a spectrum reference pattern memory in which spectrum reference patterns are stored in advance,

a spectrum similarity calculator which calculates the similarity between the said normalized spectrum pattern and a spectrum reference pattern for each of a plurality of recognition categories, and

an identifier which, as the result of the spectrum similarity calculation, outputs the name of the recognition category which has the highest similarity,

the speech recognition system further comprising:

(a) a second feature pattern calculator which calculates a second feature pattern, said second feature pattern calculator (a) including a spectrum relative value calculator which, for each channel, calculates as the spectrum relative value pattern, the relative value of the normalized spectrum in a voiced interval with respect to a normalized spectrum average value in each frame,

(b) a second feature reference pattern memory in which second feature reference patterns have previously been stored, said second feature reference pattern memory (b) including a spectrum relative value reference pattern memory storing spectrum relative value reference patterns, and

(c) a second feature similarity calculator which calculates the similarity between the said second feature pattern and the second feature reference patterns with respect to each recognition category, said second variable similarity calculator (c) including a spectrum relative value similarity calculator which calculates the similarity between the spectrum relative value pattern and the spectrum relative value reference patterns for each of the recognition categories, and in that

(d) in the identifier, the overall similarity is calculated for each of the recognition categories by reference to both the similarity of the spectrum and the similarity of the second feature, and the category giving the largest overall similarity is output as the recognition result.

8. A speech recognition system according to claim 7, wherein the said spectrum relative value calculator comprises

means for calculating the normalized spectrum average value from the speech start point frame to the speech end point frame at every channel, and

means for calculating the spectrum relative value by subtracting the normalized spectrum average value from the normalized spectrum for the particular channel, for all the frames from the speech start point frame to the speech end point frame,

wherein the spectrum relative value pattern is obtained by determining the spectrum average value for all the channels that have been frequency-analyzed.

9. A speech recognition system according to claim 7, wherein the overall similarity is calculated by adding the similarity for the spectrum and the similarity for the second feature.