JPS58176699A - Voice standard pattern registration system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and in particular, to a speech recognition device for a specific speaker, which is suitable for registering standard patterns for recognition of large word patterns, ultimately arbitrary word types. Regarding pattern registration method.

Conventional large word recognition uses monosyllables as the standard pattern,
Discrete utterances are recognized, and in this aspect, articulatory coupling (the influence of preceding and following phonemes) is not taken into consideration.
The recognition rate obtained is low. On the other hand, it is possible to improve the recognition rate by registering a VCV <vowel-consonant-vowel) that takes into account articulatory combination as a standard pattern, but the problem in this case is that many VCv standard patterns (turns) This requires a lot of time to register.Especially when recognizing a specific speaker, the
It is necessary to utter the C■ pattern, and since a large number of meaningless VC■ patterns are uttered, it is naturally psychologically exhausting, and since the player cannot get into the rhythm, there is a possibility that each vCV may not be uttered clearly.

Therefore, the present invention has been made to solve the above-mentioned problems, and when using the VCV pattern, which is currently the best standard pattern to realize the recognition of large vocabulary words, In recognition, when the user utters a meaningful word or sentence,
v) (The purpose is to provide a method for automatically cutting out turns and registering standard patterns.

In the present invention, a continuous non-linear or linear matching circuit performs segmentation by comparing each area in an input cover turn with a turn. This is to realize a method of creating and registering a standard pattern using the function that allows you to ``fc'', which requires you to utter meaningless sounds.
Resolve quality issues (psychological fatigue, inability to clearly pronounce each pattern, etc.). f, that is, a single word sentence, etc., the utterer is
All you have to do is speak in a natural and meaningful way.

Hereinafter, the present invention will be described in detail based on Examples.

FIG. 1 shows a block configuration of a speech recognition device in which standard pattern registration is performed according to the present invention, FIG. 2 shows the principle of known continuous nonlinear matching, and FIG. 3 shows an example of standard pattern creation according to the present invention. show.

As can be understood from Figures 1 to 3, when registering a standard pattern, if you utter the words or sentences in the predetermined I1m order, the 9th toe will be as shown in Figure 3 (1). When you input the voice /AKA PE I RO/, the first
The audio is digitized by the audio input unit 1 shown in the figure.

The digital signal a is subjected to frequency analysis in the feature extraction section 2.

The analyzed result b (eg filter puncture output value) is entered into the feature parameter memory 3. As will be described later, the parameters stored in this memory are transferred to the standard pattern memory 6.

Further, the signal signal is input to the distance calculation unit 4, and the memory 5 stores a reference pattern created offline in advance.
The distance to the output of is calculated.

In this example, the reference pattern is V*V (vowel pair).
is used.

In this way, for the voice input uttered as shown in Fig. 3 (1), at each point in the input frame, as shown in Fig. 3 (2),
The standard pattern (A' = A umbrella A.

H' = A*E, C' = EI) is determined.

That is, the distance calculation unit 4 in FIG. 1 calculates the distance to all frames of the reference pattern at each point in time of the input frame, and sends this to the known continuous nonlinear matching unit 7.
5. In the continuous nonlinear matching unit 7, as shown in FIG. 2 (2),
Using the input distance d, the distance calculation unit 4 performs an overall comparison, and determines which reference pattern matches which part of the input sound field. In Fig. 2 (1), the reference pattern B'
The results are shown below.

As shown in Figure 2 (3), the contents of the inlet pattern are A,
, B, and C. For example, A is /AK
Assume that A/, BFi/ANE/, and C are /EI/.

Therefore, in response to the voice saying ABC...

The continuous nonlinear matching unit 7 generates reference patterns A' and B'.

Distance from CI I,..., i.e. d of 2nd-(2)
(t.

T) is input and the recurrence formula 4- shown below is computed to determine the position of each reference pattern in the input frame.

・・・・・・・・・・・・(υ The above formula () achieves the path selection method shown in Fig. 2, and matches the current distance d(t, r) with the past matching value. to select the optimal path.

In reality, it is necessary to normalize by path length, but the explanation is omitted here. The motion therapy of this continuous nonlinear matching unit 7 is based on the processing of equation (1) for each point of the input frame 2, and the distance #1ld(t,l
at any point in the input frame by determining the matching value only at the frame end of each reference pattern.

This is to know how best to match. These details are fully described in Japanese Patent Application No. 55-158296. In FIG. 2, optimal matching is performed with reference pattern B' at time t, and pattern B starting from t-TO is determined.

Based on this principle, the position t of the reference pattern existing in an arbitrary region of the input sound is determined.

The explanation will be given again with reference to FIGS. 1 and 3.

The voice uttered as /AKANEIRO/ is matched with the standard patterns /AmA/, /A*E/, /EI/, etc., and becomes % /AKA/, /ANE/, respectively.

The area where /EI/... exists is determined.

As a result, the determination unit 8 outputs the control signal f to the 1% characteristic parameter memory 3, and the characteristic parameter /AK
Data is transferred from the address of this memory 3 storing A/, /ANE/, /EI/ to a predetermined standard pattern memory 6 address.

In this way, by uttering words or sentences in a predetermined order, the first judgment unit specifies a predetermined reference pattern and performs continuous nonlinear matching, thereby segmenting the standard pattern included in the input. , standard patterns can be created, eliminating the need to utter meaningless patterns such as VCV.

In this embodiment, it is also possible to similarly achieve segmentation of each standard pattern in the input Ha by a vowel sequence using only the vowel v using the v*■ pattern as a reference. In this case, a continuous linear matching function is used.

Continuous nonlinear matching uses VCV (
Also includes V-V. Continuous linear matching is used for matching in units such as J, and continuous linear matching is used for matching vowels with a short number of frames.As described in Figure 2, both functions are used to match which reference pattern in which part of the input audio. determines the force that exists.

When the VCV is registered as a standard pattern, in the next phase it is stored on a floppy disk, etc., and when the user comes to the stage of recognition, the standard pattern is transferred from the floppy disk to the standard pattern memory, and the first In the system shown in the figure, the contents of the standard pattern memory 6 are sent to the distance calculation section 4. In this invention, it is important to create a reference pattern off-line, since there are 40 to 50 types of V umbrella V, % rui, and vowel patterns.
There is no problem because it is easy to collect patterns from multiple speakers and create an average pattern.

As explained above, according to the present invention, when using the Cv pattern, which is the best standard pattern currently considered, for large word recognition, the user can find a meaning in the recognition of a specific speaker. By uttering nine words and sentences, the system can automatically create a standard pattern, which reduces the burden on the user and is highly effective.

[Brief explanation of drawings]

FIG. 1 shows a block configuration of a speech recognition device implementing the present invention, and FIG. 2 shows a principle diagram of continuous nonlinear matching.
FIG. 3 shows an example of standard pattern creation in the present invention. 3...Characteristic parameter memory, 5...Reference pattern memory, 6...Standard pattern memory. Agent Patent Attorney Toshiyuki Usuda ■I (

Claims (1)

[Claims] 1. Continuous matching calculations are performed between a speech reference pattern created offline in advance and a series of feature parameters extracted from speech patterns uttered in a predetermined order so as to include the standard pattern to be registered, and the optimal matching interval is determined. A speech standard pattern registration method characterized by registering feature parameters in a section as a standard pattern in a memory each time it is detected.