JPS6132680B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a speech recognition method that recognizes unknown speech by measuring the degree of similarity between the feature vector series of unknown speech and the feature vector series of standard speech stored in advance. This relates to a method that takes into account nonlinear expansion and contraction. In this specification, a feature vector refers to a plurality of audio features in one time window, and a feature vector series refers to a time series of feature vectors over a certain length of time, and refers to the order of feature vectors with respect to the time axis. A word is used as the sample position. As a specific representative example, when audio is analyzed using l filters and features are extracted by sampling at certain time intervals, the l filter outputs at a certain sample position are called feature vectors. , a time series of continuous feature vectors is called a feature vector sequence. One of the factors that worsens the recognition rate of speech recognition is variation in speaking speed, and some kind of time axis normalization is generally required in similarity measurement.
Dynamic programming is known as a typical example of time axis normalization. However, although this method restricts the calculation route by a certain recurrence formula,
Variation in speaking rate is essentially calculated by adding a relatively large number (e.g. 30
Since the method measures similarity by associating sample positions (%), it requires an extremely large amount of calculation processing time. In addition, in this method, the time axis normalization regarding the speaking rate is performed based on the similarity between feature vectors, so the time axis may be expanded or compressed nonlinearly beyond normal speaking rate fluctuations, and as a result, ,
There are cases where the degree of similarity with other standard speech that is not the target is increased, which may actually worsen the recognition rate. An object of the present invention is to perform similarity measurement for nonlinear expansion/contraction corresponding to variations in speaking rate in a short processing time, and this can be done by providing information on nonlinear expansion/contraction for each standard voice. This is what we achieved. The present invention is characterized in that non-linear fluctuations in speaking speed are achieved in areas where fluctuations in speaking speed are extremely small, such as in articulatory parts,
This is based on the recognition that this is due to the coexistence of parts where the voicing rate fluctuates greatly, such as the steady vowel part, and the similarity is measured based on the same time scale for the former part, while the similarity is measured based on linear expansion/contraction support for the latter part. This method measures the degree of similarity by measuring the degree of similarity, and allows the degree of similarity to be measured in response to non-linear fluctuations in the speaking rate as a whole. To this end, the present invention includes a memory area in which a standard voice is expressed as a feature vector series and a sample position indicating a series of portions with a low speaking rate (specific subsequences) within the feature vector series. means for extracting a feature vector sequence of the form from the unknown speech. Also, the sample positions of the standard voice and the unknown voice are made to correspond linearly, a certain range including the position corresponding to the first sample position of the specific subsequence is determined, and each position is set as the first sample position and the number of samples is the same as that of the specific subsequence. comprising a first step of determining a continuous feature sample as a corresponding subsequence candidate;
Also, measuring the similarity between the specific partial sequence and each candidate,
A second step is provided in which a candidate giving the maximum similarity is determined as a corresponding subsequence, and a corresponding similarity is detected as a similarity of a specific subsequence. In addition, the similarity between the residual sequence obtained by removing the specific subsequence from the feature vector sequence of standard speech and the remaining sequence obtained by removing the corresponding subsequence from the feature vector sequence of unknown speech is measured using linear expansion/contraction correspondence, and the similarity is calculated for each This method includes a third step of detecting the similarity of the remaining sequence, and recognizes unknown speech from the integrated similarity of the specific partial sequence and the remaining sequence, and will be described in detail below. For example, a standard speech expressed by information including the following feature vector series and sample positions of a specific subsequence is prepared in advance. Feature vector series X; x ₁ , x ₂ , ..., x _i , x _i+
1 ,...,x _+k ,...,x _n . Number of feature vectors in feature vector series X;
m. Starting sample position of specific subsequence x _i ~ _{x i + k} ; T
_i . The number of consecutive feature vectors in the specific subsequences x _i to x _i+k ; k+1. If there are multiple specific subsequences, prepare one for each. The specific subsequence is a part where the variation in speaking rate is small. Generally, when a certain word is uttered, articulatory connections occur between adjacent phonemes, and this articulatory part contains important information for speech recognition, and the variation in speech rate is extremely small in this articulatory part. This articulatory part can be easily detected, for example by visually observing a sonogram or by detecting formant transients. One or more articulatory parts are designated for each standard voice as a specific subsequence, but the designation method is not necessarily uniform. For example, if a relatively large number of feature vectors are detected in an articulatory part, it is sufficient to make this part a specific subsequence. Also, for example, for the word "nana", if we consider the case where the feature vector is extracted by sampling at 10 msec intervals, only 1 to 2 points can be extracted in the first articulatory part.
Since it is a feature vector and lacks stability in similarity measurement as described later, it is better to designate a specific subsequence by including a part of the stationary part. Furthermore, as will be described later, when the articulatory part including particularly important features is targeted, it may be specified redundantly. Also, of course, for standard speech where it is not possible to know parts where the variation in speaking speed is small,
A specific subsequence is not specified, and the similarity is measured using another method not directly related to the present invention. In order to determine a corresponding subsequence in the feature vector sequence of unknown speech, a plurality of corresponding subsequence candidates are set by correlating sample positions by linear expansion and contraction. Even if the corresponding subsequence corresponds to a toning part, its position cannot generally be specified because of the time expansion and contraction of the vowel stationary parts and stop parts before and after it. However, as long as the duration of the word speech is less than several seconds, estimation can be made within a relatively narrow range. For example, the duration of words such as "ichi", "nii", "san", "Tokyo", "Yokohama", etc. is at most 400 m.
sec to 500 msec in length, and when such a short word is sampled at a 10 msec period, the variation in the sample position of the leading feature vector of the articulatory part is about 4 to 10 samples. Therefore, if the feature vector series and sample positions of the unknown voice are as follows, the sample positions T ₁ to _{T n} of the standard voice and the sample positions T ₁ to _{T o} of the unknown voice correspond linearly. The specific subsequence x _i to x _i+k including the position T _j corresponding to the first sample position T _i and corresponding to a plurality of sample positions T _j-5 to _{T j+4} before and after it are unknown. The corresponding subsequence can be well estimated by finding the sample positions T ₁ to _{T o} of the audio and selecting these as the leading sample position. The number of corresponding subsequence candidates is required to be about 4 to 10, and the number can be determined for each standard voice, but since it is about 10 at most, it may be uniform. Unknown speech feature vector sequence Y; y ₁ , y ₂ , ...,
y _i ,..., y _o . Unknown voice sample position; T ₁ , T ₂ , ..., T _j , ...
... _To . That is, a candidate whose leading position is the sample position T _j of the following first equation as shown in FIG. 1 is determined.

[Table] To determine the corresponding subsequence, T _j-5 to _{T j+4}
Measure the similarity between the specific subsequence and 10 corresponding subsequence candidates consisting of the same number of continuous feature vectors as that of the specific subsequence with the leading sample position as the first sample position, and select the one that gives the maximum similarity. Determine the corresponding subsequence. Similarity can be measured by feature distances such as absolute distances and squared distances, and when the feature distances are expressed as d( ), for example, the difference between a candidate whose first sample position is sample position T _j and a specific subsequence is Distance D _j
becomes as follows. D _j = d (x _i , y _j ) + d (x _{i +1} , y _{j +1} )...d (x _i + k, y _{j + k} ) - 2nd formula The distance to each candidate is D _j-5 ~ Assuming that D _j+4 is the smallest candidate among them, that is, the candidate with the maximum similarity is y _ja + y _ja+1 , ..., y _{ja + k} , this is determined as the corresponding subsequence, and the distance at that time is D ₁
is determined as the similarity of the specific subsequence, and if there is no other specific subsequence, the process moves on to measuring the similarity of the remaining feature vector sequences. Note that if the number of feature vectors in a specific subsequence is extremely small, the determination of a corresponding subsequence becomes unstable, so it is better to make the specific subsequence a certain length, including feature vectors such as vowel stationary parts. FIG. 1 shows the linear correspondence of the residual feature vector series. Similarity measurements for specific subsequences were performed on the same time scale, but generally differ for the remaining sequences. Therefore, the similarity is measured by linearly expanding and contracting the sample positions of each residual sequence of the standard speech excluding the specific subsequence and the sample positions of each remaining sequence of the unknown speech excluding the corresponding subsequence. For example, in FIG. 1, in the correspondence between the feature vectors x _i+k+1 to x _n of the last residual sequence of standard speech and the last residual sequence y _ja+k+1 to y _o of unknown speech, samples Positions Tu and Tv correspond to feature vectors x _u and y _v according to the following relationship. Tv=T _o −T _ja+k+1 / _Tn −T _i+k+1 (T _u −T _i+k+1 )+T _ja+k+1 −3rd formula However, T _u =T _i+k+1 , T _{i+k+ 2} , ..., T _n The distance d (x _u , y _v ) between the pair of feature vectors x _u , y _v thus made to correspond by nonlinear expansion/contraction is calculated, and the sum of these is the distance D ₂ of the final residual sequence. ,
In other words, it is the degree of similarity. The degree of similarity between the standard voice and the unknown voice is determined by dividing the sum of the degree of similarity in the specific partial sequence and the degree of similarity in the remaining sequence by the number of samples m of the standard voice. Note that when a certain standard speech has a plurality of specific subsequences, the specific subsequences may be overlapped so that a specific feature vector belongs to both. This case is equivalent to giving weight to that partial time series, and is effective when you want to emphasize that part. However, when determining the similarity of standard speech, it is necessary to add the length of the overlapping part and take a proper look, so it is stored as a normalized number separately from the number of samples, and the similarity regarding the specific subsequence/remaining sequence is stored separately from the number of samples. The total sum is divided by this normalized number to obtain the degree of similarity, that is, the distance to the standard speech. Note that in the present invention, it is not necessarily necessary to store the total number of samples of standard speech in advance; in this case, on the contrary, the feature vector T in the residual sequence
The correspondence between _u and _Tv is determined by identifying the sample position in the unknown voice, and correspondingly determining the sample position in the standard voice. The degree of similarity of the standard speech is determined by dividing the degree. As is clear from the above explanation, since the present invention is essentially linear matching, the processing speed can be configured with a sufficiently high speed and simple circuit compared to dynamic programming, etc. In recognizing words and phrases, by using partial time series that are overlapped (or even completely matched) for different feature sequences, that part is emphasized, making identification easier. There is. Furthermore, the present invention can automatically set a specific subsequence by using a transient part detection means, and can therefore be applied to a registered word recognition device, but it is not applicable to a speaker-independent recognition device. Since it is easy to apply various restrictions to the memory pattern of standard speech, it is most suitable for speaker-independent recognition.

[Brief explanation of the drawing]

Figures 1 and 2 are explanatory diagrams related to the present invention, in which Figure 1 is an explanatory diagram of the corresponding sample position between a specific subsequence and each candidate, and Figure 2 is a diagram of similarity measurement between unknown speech and standard speech. FIG. 3 is an explanatory diagram of corresponding sample positions in FIG. x ₁ , x ₂ , ..., x _i , ..., x _n ; Standard speech feature vector sequence, y ₁ , y ₂ , ..., y _j , ... y _o ;
Feature vector sequence of unknown speech, T ₁ , T ₂ , ...,
T _i , ..., T _n ; Sample position of standard voice, T ₁ ,
T ₂ , ..., T _j , ...T _o ; Sample position of unknown voice.

Claims (1)

[Claims] 1. Specification consisting of a feature vector series consisting of feature vectors at a large number of sample positions and a plurality of continuous feature vectors in the series, using a feature vector of a voice corresponding to one sample position on the time axis. A speech recognition system comprising: a standard speech memory area expressing a standard speech using information including a sample position of a subsequence; and an extraction means for extracting a feature vector sequence in a format corresponding to the feature vector sequence of the standard speech from an unknown speech. In the apparatus, a sample position in the feature vector series of the standard speech and a sample position in the feature vector series of the unknown speech are made to correspond linearly, and a corresponding subsequence candidate in the unknown speech is determined based on the leading sample position of the specific subsequence. A plurality of leading sample positions are determined, and a plurality of successive corresponding subsequence candidates with the same number of feature vectors in the specific subsequence are unknown, using the feature vector of each of the determined leading sample positions as a leading feature vector. a first step of determining a voice feature vector sequence; a second step of measuring the degree of similarity between the specific subsequence and each of the candidates and determining the candidate that provides the maximum similarity as the corresponding subsequence; The residual sequence obtained by removing the specific subsequence from the standard speech feature vector sequence and the remaining sequence obtained by removing the corresponding subsequence detected in the second step from the unknown speech feature vector sequence are linearly calculated with respect to the sample position. and a third step of measuring the similarity between both residual sequences by expanding and contracting them into correspondence, and the maximum similarity in the second step and the third step.
A speech recognition method characterized by recognizing unknown speech based on similarity in stages. 2. The determination of each leading sample position of each corresponding partial sequence candidate in the first step determines a sample position on the unknown voice side corresponding to the leading sample position of the specific partial sequence and a specific number of sample positions in the vicinity thereof. A speech recognition method according to claim 1, characterized in that the speech recognition method is characterized in that: