JP2705062B2 - Split labeling device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a time-series division labeling apparatus for feature vectors representing audio data. [Prior Art] FIG. 6 is a functional block diagram of a conventional divided labeling apparatus, in which a divided labeling unit (3) converts a plurality of learning speech data stored in a file apparatus (1) into these. For each of the learning speech data, the divided labels are assigned according to the label sequence stored in the file device (2), and the division result is assigned to the file device (4) and the label is assigned to the file device (5). Store another statistic. Since the statistic for each label obtained in this way is used for template matching in a speech recognition device or the like, it is necessary to obtain a statistic for each label with as little bias as possible.
For that purpose, it is necessary to label a large amount of audio data including many vocal environments. [Problems to be Solved by the Invention] In the conventional apparatus having the above-described configuration, in use, it is necessary to create a label string to be divided and attached to each of the learning speech data, and thus a large amount of In order to label audio data in a divided manner, a large number of labels must be created as a label sequence, which is not suitable for dividing and labeling a large amount of audio data. Also, when adding a new label type or changing the configuration of a label sequence for the purpose of reflecting the influence of the context of the utterance on the label sequence, etc., the same It is not easy to make corrections in such a way that errors do not occur. In order to eliminate such a drawback, the present invention provides a conventional device of this type with a label sequence generating section for generating a label sequence corresponding to the phoneme symbol sequence representing the utterance content of voice data by a rule. The purpose of the present invention is to make it easy to generate a large number of label strings and to easily change the label strings by changing the rules without error. [Embodiment of the Invention] FIG. 1 is a functional block diagram of an embodiment of the present invention, and (6) is a file device for storing a phoneme symbol string representing the utterance content of learning speech data, (7) ) Is a label string generation unit that generates a label string from the phoneme symbol string on the file device (6) and stores the label string in the file device (2). FIG. 2 illustrates a part of the phoneme symbol string stored in the file device (6).
This is a number for associating the learning speech data stored in the file device (1) with the label string stored in the file device (2), and is associated with the same number. The phoneme symbols constituting these phoneme symbol strings include those shown in FIG. The utterance content of any learning speech data is represented by a combination of these phoneme symbols. The label string generation unit (7) generates a label string from each of the phoneme symbol strings stored in the file device (6) and stores the label string in the file device (2). By the way, the phoneme symbol string as described above is an abstract representation of the utterance content and has a complicated correspondence with the label string attached to the physical voice pattern. Must be converted by a multi-step rewrite rule group. FIG. 3 shows an example of the configuration of such a label string generation unit. The phoneme symbol string (701) to be converted is (7041)
Refer to the rewrite rule group (704) composed of the rewrite rules of (1) to (7046), and replace the subsymbol string that matches the symbol string on the left side of these rules with the symbol string written on the right side of the rule. It is converted into a label string (702) by the rule applying unit (703) which is repeatedly performed. FIG. 4 shows a phoneme symbol string partially exemplified in FIG. 2 and a label string obtained by passing through the above-described label string generation unit with a corresponding number attached to the left side. As shown in this figure,
In general, a label sequence is considerably more complex than a phoneme symbol sequence, and the device according to the present invention has a larger amount of data than a conventional device which has to create and provide a label sequence for each voice data. It can be said that it is suitable for dividing and labeling audio data. On the other hand, as for the increase of the label type and the change of the label configuration, only the rewriting rule group needs to be changed, and the number of rules of the rewriting rule group (704) used in the label sequence generation unit (7) is subject to the division labeling. Since the number is usually sufficiently smaller than the number of voice data, it can be said that the apparatus according to the present invention is superior to the conventional apparatus that changes the label string itself. [Effects of the Invention] As described above, in the divided labeling apparatus according to the present invention, a label sequence generation unit that generates a label sequence from a phoneme symbol sequence representing the utterance content of voice data according to a group of rewrite rules that can be changed. By eliminating the need to create and give a complex label sequence for each audio data, the addition of label types and changes in the label sequence configuration can be achieved by changing a small number of rewriting rules. There is an effect that it is easy to label a large number of different divided labels.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a functional block diagram of one embodiment of the present invention, FIG. 2 is a diagram showing an example of phoneme symbol string representation, and FIG. 3 is a label string generation in one embodiment of the present invention. FIG. 4 shows an example of a generated label sequence, FIG. 5 shows an example of phoneme symbols constituting a phoneme symbol sequence, and FIG. 6 shows a conventional divided labeling apparatus. 3 is a functional block diagram of FIG. In the figure, (1) is a file device, (2) is a file device, (3) is a divided label portion, (4) is a file device,
(5) File device, (6) File device, (7)
Is a label sequence generation unit, (701) is a phoneme symbol sequence, (702) is a label sequence, (703) is a rule application unit, and (704) is a rewrite rule group. In the drawings, the same or corresponding parts are denoted by the same reference numerals.

Claims (1)

(57) [Claims] Voice data storage means (1) for storing a time series of feature vectors representing learning voice data, and phoneme symbols corresponding to the learning voice data and storing phoneme symbol strings representing the utterance contents of the learning voice data A learning voice stored in the voice data storage means (1) based on a rewrite rule group which can be changed from a row storage means (6) and a phoneme symbol string stored in the phoneme symbol string storage means (6). A label string generating means (7, 2) for generating a label string attached to a physical voice pattern in a time series of a data feature vector; and a label string generated by the label string generating means (7, 2). The divided labeling means (3) for dividing the time series of the feature vector of the learning speech data stored in the speech data storage means (1) in accordance with the label and sequentially assigning each label of this label string to each divided section. Dividing labeling apparatus characterized in that it comprises and.