JP2007233148A - Utterance section detection device and utterance section detection program - Google Patents

Abstract

Speech segments are detected quickly and with high accuracy.
In an utterance interval detection device for detecting an utterance interval from input speech, an acoustic analysis means for converting the input speech into an acoustic feature, an acoustic feature obtained by the acoustic analysis, and a preset acoustic A continuous speech recognition means for sequentially calculating a cumulative likelihood in each subword in synchronization with the input speech using a subword network comprising a model and a language model; and a speech start point in the input speech from the cumulative likelihood in each subword And the utterance section detecting means for sequentially detecting the utterance end, the above-mentioned problems are solved.
[Selection] Figure 1

Description

  The present invention relates to an utterance interval detection device and an utterance interval detection program, and more particularly to an utterance interval detection device and an utterance interval detection program for quickly and efficiently detecting an utterance interval for speech.

  In speech recognition used for caption production and metadata production of broadcast programs, it is important to improve speech detection performance in noisy environments and conversations, and speech recognition performance mixed with male and female speakers. Therefore, conventionally, various methods for detecting an utterance section from words, voices, and the like have been proposed. For example, conventional speech segment detection methods include a method using short-time power (for example, see Non-Patent Document 1 and Patent Document 1), a method based on a phoneme recognition result (for example, see Non-Patent Document 2), A method using the likelihood at the time of recognition (for example, see Patent Document 2) and a method based on a local speech / non-speech likelihood ratio (for example, see Patent Document 3) are known.

  Here, the method using the short-time power provides a short-time power threshold for speech and a short-time power threshold for non-speech, and when the short-time power of the input speech exceeds the speech threshold, Is the beginning of speech, and the end of speech when the short-time power of the input speech falls below the non-speech threshold. The two threshold values are dynamically changed according to the fluctuation of the short-time power of the input speech. Therefore, it is intended to reduce the influence of noise and the like.

Also, the method based on the phoneme recognition result performs continuous speech recognition in units of phonemes and identifies a portion recognized as non-speech as an utterance start / end. The technique using the likelihood at the time of recognition detects an utterance section by detecting a pause during utterance. Furthermore, the method based on the likelihood ratio of local speech / non-speech is to determine speech / non-speech independently in a short speech interval.
P. Renevey, et al. , “Entropy Based Voice Activity Detection in Vary Noise Conditions”, Eurospeech-2001, pp. 199-001. 1887-1890, 2001. JP 2005-31632 A F. Kubala, et al. "The 1996 BBN Byblos HUB-4 Transcription System", DARPA Speech Recognition Works, pp. 90-93, 1997. Japanese Patent Laid-Open No. 9-258765 Japanese Patent No. 3105465

  However, in the utterance detection method described above, the method using the power for a short time is very simple and widely used. However, even if there is no noise in the speech, the beginning of the utterance is sufficient. In many cases, the beginning of words such as “Japan” and “Hokkaido” that do not increase in power are missed, and the speech detection performance of such low S / N ratio speech is not practically sufficient.

  Moreover, although the method based on the phoneme recognition result has no problem in the off-line processing, the acquisition of the phoneme recognition result has a large time delay from the input speech, and is not suitable for the on-line processing.

  In addition, the method using the likelihood at the time of recognition has no problem because the utterance end is a pose itself, but the utterance start end is not determined until the utterance end point or the utterance end pose is detected. In an utterance such as reading a document that does not appear, a time delay from the input voice becomes a problem.

  Furthermore, the local speech / non-speech likelihood ratio method is to judge speech / non-speech independently in a short speech interval. Empirical smoothing processing such as value processing is required, and optimization of the speech section detection under various acoustic environments is not easy.

  The present invention has been made in view of the above-described problems, and an object thereof is to provide an utterance section detection device and an utterance section detection program for detecting an utterance section quickly and with high accuracy.

  In order to solve the above problems, the present invention employs means for solving the problems having the following characteristics.

  The invention described in claim 1 is an utterance section detection device for detecting a utterance section from input speech, an acoustic analysis means for converting the input speech into an acoustic feature quantity, and an acoustic feature quantity obtained by the acoustic analysis means, A continuous speech recognition means for sequentially calculating a cumulative likelihood in each subword in synchronization with the input speech using a preset subword network consisting of an acoustic model and a language model, and a cumulative likelihood in each subword It has an utterance section detecting means for sequentially detecting an utterance start end and an utterance end in the input voice.

  According to the first aspect of the present invention, it is possible to detect an utterance section quickly and with high accuracy. Therefore, it is possible to automatically remove non-speech sections unnecessary for speech recognition, such as silence, noise, music, etc., from the input speech and extract only the speech sections to be recognized. Thereby, the amount of speech recognition processing is reduced and the recognition performance is improved.

  The invention described in claim 2 includes a subword acoustic model having one or a plurality of speaker clusters that express acoustic features of speech and non-speech sounds, and a subword language that expresses transitions between subword acoustic models. And subword network integration means for integrating the subword network using a model.

  According to the second aspect of the present invention, it is possible to generate a highly accurate subword network corresponding to the content of the input voice.

  According to a third aspect of the present invention, the subword / network integration means includes a transition from the utterance section detection start state to an acoustic model corresponding to non-speech of all speaker clusters, and the non-speech acoustic model to each Transition according to the subword language model to the acoustic model corresponding to the speech of the speaker cluster, transition from the non-speech acoustic model to the utterance interval detection start state to absorb the non-speech over a certain length of time, each speaker Transitions according to the subword language model between the acoustic models corresponding to the speech of the cluster, transitions with a penalty from the acoustic model corresponding to the speech of each speaker cluster to the acoustic model corresponding to the speech of a different speaker cluster, each speaker Corresponding to non-speech from acoustic model corresponding to cluster speech A transition according to a subword language model to an acoustic model, a transition to an utterance interval detection end state according to the utterance end detection condition, and a transition from the utterance interval detection end state to the utterance interval detection start state, A subword network that enables at least one transition is configured.

  According to the third aspect of the present invention, the accuracy of the subword can be improved by performing the respective state transitions.

  In the invention described in claim 4, the continuous speech recognition means detects a speech start edge over a certain length of time after transitioning from a speech section detection start state in the subword network to an acoustic model corresponding to non-speech or speech. If the condition is not satisfied, the speech recognition unit returns to the speech segment detection start state from the non-speech acoustic model, and at the same time, the results of speech recognition such as cumulative likelihood in all acoustic models are cleared and the speech segment detection start time is updated. Then, continuous speech recognition in units of subwords is started again.

  According to the fourth aspect of the present invention, it is possible to absorb a long non-speech until the start of the utterance is detected. Therefore, it is possible to detect the utterance start end with high accuracy.

  In the invention described in claim 5, when detecting the utterance start edge, the utterance interval detection means supports speech of all speaker clusters for all input speech from the utterance interval detection start time to the current time. Of the acoustic models, the ratio between the maximum cumulative likelihood and the cumulative likelihood of the acoustic model corresponding to the non-speech of the same speaker cluster following the speech section detection start state is sequentially calculated in synchronization with the input speech, Based on the calculated ratio value and a preset threshold, a time that is a certain length of time from the end time of the non-speech acoustic model at the start of the subword sequence indicating the maximum cumulative likelihood is detected as the start of speech It is characterized by doing.

  According to the fifth aspect of the present invention, it is possible to detect the utterance start end quickly and with high accuracy.

  According to a sixth aspect of the present invention, when detecting the utterance end, the utterance section detecting means corresponds to speech of all speaker clusters for all input speech from the utterance section detection start time to the current time. Sequential calculation of the ratio of the maximum cumulative likelihood of the acoustic models corresponding to non-speech following the acoustic model and the maximum cumulative likelihood of the acoustic model corresponding to speech of the same speaker cluster in synchronization with the input speech When the calculated ratio value exceeds a preset threshold for a certain time length or more, a time that is a certain time length backward from the time when the ratio starts to be exceeded is detected as an utterance end point.

  According to the sixth aspect of the present invention, the utterance end can be detected quickly and with high accuracy.

  The invention described in claim 7 is characterized in that the utterance section detecting means outputs the voice of the utterance section from the input voice based on time information of the utterance start end and the utterance end.

  According to the seventh aspect of the present invention, it is possible to output the voice of the utterance section quickly and with high accuracy based on the time information of the utterance start end and the utterance end.

  The invention described in claim 8 is an utterance period detection program for causing a computer to execute an utterance period detection process for detecting an utterance period from an input voice, and an acoustic analysis process for converting the input voice into an acoustic feature amount; Continuous speech that sequentially calculates the cumulative likelihood in each subword in synchronization with the input speech using the acoustic feature obtained by the acoustic analysis processing and a subword network composed of a preset acoustic model and language model The computer is caused to execute a recognition process and an utterance section detection process for sequentially detecting the utterance start end and the utterance end from the cumulative likelihood in each subword.

  According to the invention described in claim 8, it is possible to detect the utterance section quickly and with high accuracy. Further, by installing the execution program in the computer, it is possible to easily detect the utterance section.

  According to the present invention, it is possible to detect an utterance section quickly and with high accuracy.

<Outline of the present invention>
The present invention relates to an utterance interval detection technique for automatically and quickly detecting an utterance interval of a human voice spoken under various acoustic environments from speech. Specifically, subword acoustic models and subword language models of multiple speaker clusters are integrated to form a subword network, and continuous speech recognition in units of subwords (eg, phonemes, syllables, triphones, etc.) for input speech. During execution, the cumulative likelihood in each subword corresponding to speech and non-speech is calculated and compared in synchronization with the input speech, so that the speech start and speech end can be detected with high accuracy with a small delay time.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments in which an utterance section detection apparatus and an utterance section detection program according to the present invention having the above-described features are preferably described in detail with reference to the drawings.

<Speaking section detection device: device configuration>
FIG. 1 is a diagram illustrating a configuration example of an utterance section detection device according to the present invention. The utterance section detection apparatus 10 shown in FIG. 1 is configured to include a subword / network integration means 11, an acoustic analysis means 12, a continuous speech recognition means 13, and an utterance section detection apparatus 14.

  The subword / network integration unit 11 generates a subword network 23 by using the subword acoustic model 21 of one or a plurality of speaker clusters and a preset subword language model 22, and transmits the subword network 23 to the continuous speech recognition unit 13. Output.

  Here, for example, when the number of speaker clusters is 2, the subword acoustic model 21 is a speaker cluster A male, speaker cluster B female, or speaker cluster A wideband speech, speaker cluster B narrowband. As the speech or the like, the subword can be arbitrarily set such as a phoneme or syllable that is dependent on the acoustic environment or independent of the acoustic environment. The number of speaker clusters in the subword acoustic model may be three or more, or may be singular.

  The subword language model 22 can arbitrarily set an existing chain probability model such as a phoneme chain probability model or a syllable chain probability model. The subword network 23 will be described later.

In addition, the acoustic analysis unit 12 inputs the input voice 24 to be utterance detection target, converts it into an acoustic feature value 25, and outputs it. The acoustic feature quantity 25 has the same configuration as the acoustic feature quantity used for learning the subword acoustic model 21, and can be, for example, a cepstrum representing frequency characteristics, short-time power, dynamic feature quantities thereof, or the like. . Here, in the following description, the column of the acoustic feature value 25 from the start point detection start time τ of the utterance to the current time ^{t is} assumed to be x _τ ^t .

The continuous speech recognition means 13 performs a state transition in accordance with the subword network 23 in synchronization with the input of the acoustic feature 25, and the sequence x _τ of the acoustic feature 25 from the utterance start detection start time τ to the current time t. A time-synchronized beam search speech recognition method using, for example, a hidden Markov model (for example, Seiichi Nakagawa, “Speech recognition by a probability model”) using a hidden Markov model, for example, a sequence of a plurality of subwords that may correspond to ^t. , In accordance with the Institute of Electronics, Information and Communication Engineers, pp. 44-46, 1988, etc.). In addition, the recognition method of the subword string and the cumulative likelihood 26 thereof in the continuous speech recognition means 13 will be described later.

  The utterance section detection unit 14 detects one or a plurality of utterance start points and utterance end points in the input speech 24 based on the subword cumulative likelihood 26 obtained by the continuous speech recognition unit 13. Specifically, the utterance section detection unit 14 outputs the utterance start time 27 and the utterance end time 28 corresponding to the time (time record) given to the input voice 24. Further, the utterance section detecting means 14 may output the utterance section voice 29 corresponding to the utterance start time 27 and the utterance end time 28. With the configuration of the utterance section detection device 10 described above, the utterance section can be detected quickly and with high accuracy.

  In the utterance section detection apparatus 10 described above, the subword network 23 is generated from the subword acoustic model 21 and the subword language model 22 of the speaker cluster by the subword / network integration unit 11. Instead, the subword network 23 may be generated in advance and stored in the continuous speech recognition means 13 or other storage means (not shown).

<Subword network 23>
Here, the above-described subword network will be specifically described. FIG. 2 is a diagram illustrating an example of a subword network when the number of speaker clusters is two.

  The subword network 23 having two speaker clusters shown in FIG. 2 includes an utterance detection start state 31, a non-speech acoustic model 32 of the speaker cluster A corresponding to the utterance start point, and a speech acoustic model of the speaker cluster A. 33, non-speech acoustic model 34 of speaker cluster A corresponding to the end of speech, non-speech acoustic model 35 of speaker cluster B corresponding to the start of speech, speech acoustic model 36 of speaker cluster B, and end of speech Can be configured to have a non-speech acoustic model 37 of the speaker cluster B corresponding to

  Here, for example, a hidden Markov model can be used as the acoustic model. The non-speech acoustic model is learned in advance from speech such as silence, noise, music, etc. other than speech, and the speech acoustic model is speech. It is assumed that learning is performed in advance in units of subwords such as phonemes such as vowels and consonants and syllables.

  In FIG. 2, it is possible to transition from the speech detection start state 31 to the non-speech acoustic model 32 of the speaker cluster A and the non-speech acoustic model 35 of the speaker cluster B without restriction immediately after the start of the speech segment detection (FIG. 2). Arrow * 2 in 2).

  In addition, a transition can be made between the non-speech acoustic models 32 and 34 of the speaker cluster A and the speech acoustic model 33 of the speaker cluster A according to the subword language model 22 (arrow * 2 in the figure).

  Similarly, a transition can be made between the non-speech acoustic models 35 and 37 of the speaker cluster B and the speech acoustic model 36 of the speaker cluster B according to the subword language model 22 (arrow * 2 in FIG. 2). .

  Further, the utterance start detection condition is not satisfied from the non-speech acoustic model 32 of the speaker cluster A and the non-speech acoustic model 35 of the speaker cluster B to the utterance detection start state 31 for a predetermined time length. Transition to the case (arrow * 3 in FIG. 2).

  Further, the speech acoustic model 33 of the speaker cluster A and the speech acoustic model 36 of the speaker cluster B can transition to different speaker clusters with a predetermined penalty (arrow * 4 in FIG. 2).

  Further, transition from the non-speech acoustic model 34 of the speaker cluster A and the non-speech acoustic model 37 of the speaker cluster B to the utterance detection end state 38 can be made according to the utterance end detection condition (arrow * in FIG. 2). 5). Furthermore, it is possible to transition from the utterance detection end state 38 to the utterance detection start state 31 immediately after the utterance end detection without restriction for the next utterance (arrow * 6 in FIG. 2).

  Note that the non-speech acoustic model 32 of the speaker cluster A and the non-speech acoustic model 35 of the speaker cluster B can be collectively configured as one non-speech acoustic model. Similarly, the non-speech acoustic model 34 of the speaker cluster A and the non-speech acoustic model 37 of the speaker cluster B can be collectively configured as one non-speech acoustic model.

  Here, the non-speech acoustic models 32 and 34 of the speaker cluster A are expressed as different states, but their statistical properties may be exactly the same. Similarly, the non-speech acoustic models 35 and 37 of the speaker cluster B are expressed as different states, but their statistical properties may be exactly the same.

  The subword network integration means 11 in the present invention can integrate the subword network 23 using at least one of the transitions described above in one or a plurality of speaker clusters.

<Subword sequence and their cumulative likelihood 26>
Next, a method for recognizing sub-word strings and their cumulative likelihood 26 in the continuous speech recognition means 13 will be described in detail. FIG. 3 is a diagram illustrating an example of speech recognition at the beginning of utterance. FIG. 4 is a diagram showing an example of speech recognition at the utterance end.

For example, when the number of speaker clusters in the subword acoustic model 21 is 2, and the time-synchronized beam search speech recognition process is performed, the non-speech acoustic model of the speaker cluster S∈ {A, B} is set to sil _S , If the speech acoustic model of the user cluster S is ph _{S, i} (where i indicates a subword number such as a phoneme), there is a possibility of corresponding to an acoustic feature 25 as shown in FIG. For a plurality of subword strings, the logarithmic value of the cumulative likelihood of the maximum likelihood subword string is sequentially obtained by the following equation (1).

更に、始端の非スピーチ音響モデルの累積尤度の対数値を以下に示す（２）式により逐次求める。

Further, the logarithmic value of the cumulative likelihood of the non-speech acoustic model at the beginning is sequentially obtained by the following equation (2).

また、発話終端では、図４に示すような発話の始端検出開始時刻τから現時刻ｔまでの音響特徴量２５の列ｘ_τ ^tに対応する可能性のある複数のサブワード列に対して、全話者クラスタのスピーチに対応する音響モデルに後続し、非スピーチに対応する音響モデルのうち、最大の累積尤度の対数値を以下に示す（３）式により逐次求める。

Further, at the end of the utterance, all the subword strings that may correspond to the string x _τ ^t of the acoustic feature value 25 from the utterance start detection start time τ to the current time t as shown in FIG. Subsequent to the acoustic model corresponding to the speech of the speaker cluster, the logarithmic value of the maximum cumulative likelihood among the acoustic models corresponding to the non-speech is sequentially obtained by the following equation (3).

更に、同じ話者クラスタのスピーチに対応する音響モデルの最大の累積尤度の対数値を以下に示す（４）式により逐次求める。

Further, the logarithmic value of the maximum cumulative likelihood of the acoustic model corresponding to the speech of the same speaker cluster is sequentially obtained by the following equation (4).

なお、連続音声認識中は、話者クラスタ間のサブワード音響モデルの遷移を許可するものとし、話者クラスタ間のサブワード音響モデルの遷移を許可する場合、一定のペナルティのスコアをサブワード累積尤度の対数値に付加する。上述した処理を行うことで、連続音声認識手段１３は高精度なサブワード累積尤度２６を出力することができる。

During continuous speech recognition, subword acoustic model transitions between speaker clusters are allowed, and when subword acoustic model transitions between speaker clusters are allowed, a score of a certain penalty is assigned to the subword cumulative likelihood. Append to logarithmic value. By performing the processing described above, the continuous speech recognition unit 13 can output the subword cumulative likelihood 26 with high accuracy.

Note that the continuous speech recognition means 13 will be described later, which is set in advance after a transition from an utterance section detection start state in the subword network 23 to an acoustic model corresponding to non-speech or speech over a certain time length t _idle. When the utterance start edge detection condition is not satisfied, the speech recognition detection results such as the cumulative likelihood in all acoustic models are cleared (reset) at the same time as returning to the utterance section detection start state from the non-speech acoustic model. The section detection start time τ is updated to the current time t, and continuous speech recognition in units of subwords is started again. Thereby, it is possible to absorb a long non-speech until the start of the utterance is detected. Therefore, it is possible to detect the utterance start end with high accuracy.

<Speech section detection means 14>
Next, the utterance section detection unit 14 will be specifically described. The utterance section detection means 14 sets a threshold value θ _{start at} which the difference between the logarithmic value L ₁ of the cumulative likelihood of the maximum likelihood subword sequence and the logarithmic value L ₂ of the cumulative likelihood of the non-speech acoustic model at the _start is constant at the utterance _start . When this is exceeded, that is, when (L ₁ −L ₂ )> θ _start , this is used as the utterance _start edge detection condition, and the non-speech acoustic model at the start of the subword string indicating the maximum cumulative likelihood as shown in FIG. An utterance _start time 27 is a time that is a predetermined time length t _start from the end time.

Note that the time length t _start is preferably about 200 msec in the case of a general voice speed for reading a news manuscript, for example, but is not limited to this in the present invention.

On the other hand, in the speech termination, the logarithmic value L ₃ of the maximum cumulative likelihood of the maximum likelihood word string termination is non-speech acoustic models, accumulation of maximum likelihood subword sequence to terminate the speech acoustic models of the speaker cluster when the difference between the logarithmic value _{L 4} of the likelihood is that a certain threshold theta case of a continuously exceeds the time length _{t end1 end the,} i.e. _{t end1} continued _{(L 3 -L 4)> θ} end, which as the speech termination detection condition, as shown in FIG. 4, the predetermined length of time _{t end2 (t end2 <t end1} ) min time speech ending time 28 going back relative to the time length _{t end1} from the current time t .

Note that the time length t _end1 is a reference for the utterance end detection condition, and thus becomes longer than the actual utterance end time. Therefore, by setting a time length t _end2 that satisfies the relationship of t _end2 <t _end1 , it is possible to detect a time closer to the utterance termination part. Here, the time length _tend2 is preferably about 200 msec in the case of a general voice speed for reading a news manuscript, for example, but is not limited to this in the present invention.

  Thereby, the processing amount of voice recognition can be reduced. Also, the recognition performance can be improved. Therefore, it is possible to quickly and accurately detect the utterance section from the input voice.

<Execution program>
Here, the utterance section detection device 10 described above can perform the utterance section detection processing according to the present invention using the above-described dedicated device configuration or the like, but can execute a process in each configuration on a computer. , And the program is installed in a general-purpose personal computer, server, or the like, for example, so that the speech segment detection processing according to the present invention can be realized.

<Hardware configuration>
Here, a hardware configuration example of a computer capable of executing the speech section detection processing according to the present invention will be described with reference to the drawings. FIG. 5 is a diagram illustrating an example of a hardware configuration capable of realizing the speech segment detection processing according to the present invention.

  5 includes an input device 41, an output device 42, a drive device 43, an auxiliary storage device 44, a memory device 45, a CPU (Central Processing Unit) 46 for performing various controls, and a network connection device. 47, which are connected to each other by a system bus B.

  The input device 41 includes a keyboard and a pointing device such as a mouse operated by a user, a voice input device, and the like, and inputs various operation signals and voice signals such as a program execution instruction from the user. The output device 42 has a display, a speaker, and the like that display various windows and data necessary for operating the computer main body for performing processing in the present invention. Display or audio output is possible.

  Here, in the present invention, the execution program installed in the computer main body is provided by, for example, the recording medium 48 such as a CD-ROM. The recording medium 48 on which the program is recorded can be set in the drive device 43, and the execution program included in the recording medium 48 is installed in the auxiliary storage device 44 from the recording medium 48 via the drive device 43.

  Further, the drive device 43 can record the execution program according to the present invention on the recording medium 48. Thereby, using the recording medium 48, it can be easily installed in a plurality of other computers, and the speech segment detection processing can be easily realized.

  The auxiliary storage device 44 is a storage means such as a hard disk, and can store an execution program according to the present invention, a control program provided in a computer, and the like, and can perform input / output as necessary. The auxiliary storage device 44 also includes the subword acoustic model 21, the subword language model 22, the subword network 23, the input speech 24, the acoustic feature 25, the subword cumulative likelihood 26, the utterance start time 27, the utterance end time 28, Also, it can be used as a storage means for storing the speech section voice 29 and the like.

  The CPU 46 performs various calculations and data input / output with each hardware component based on a control program such as an OS (Operating System) and an execution program read from the auxiliary storage device 44 and stored in the memory device 45. Each process in the utterance section detection process can be realized by controlling the process of the entire computer. Various information necessary during the execution of the program can be acquired from the auxiliary storage device 44 and can also be stored.

  The network connection device 47 obtains an execution program from another terminal connected to the communication network or executes the program by connecting to a communication network such as a telephone line or a LAN (Local Area Network) cable. The execution result obtained in this way or the execution program in the present invention can be provided to other terminals or the like.

  With the hardware configuration as described above, the above-described speech segment detection processing can be realized at low cost without requiring a special device configuration. Further, by installing the program, it is possible to easily realize the speech segment detection process.

<Speech section detection processing procedure>
Next, a speech segment detection processing procedure using the execution program (speech segment detection program) according to the present invention will be described with reference to a flowchart. FIG. 6 is a flowchart illustrating an example of an utterance section detection processing procedure. In the utterance section detection processing procedure shown in FIG. 6, a detection target parameter is provided to clarify whether the detection target is the utterance start end or the utterance end. Further, in the following description, it is assumed that either “starting end” or “end” is set as the detection target parameter, but the present invention is not limited to this.

  In FIG. 6, first, immediately after the start of the program, the subword network is integrated using the subword acoustic model and subword language model of a plurality of speaker clusters (S01). Is set (S02). In addition, the process so far may be processed previously as pre-processing.

  Next, it is determined whether or not there is a voice input (S03), and when a voice is input (YES in S03), a short time of, for example, about 25 milliseconds necessary for calculating the acoustic feature amount for one frame is required. The voice of the section is digitally input (S04), and the input voice is analyzed (S05). Next, for each acoustic feature obtained in the process of S04, each cumulative likelihood is calculated on the subword network obtained in the process of S01 (S06).

  Here, it is determined whether or not “starting end” is set to a parameter set in advance as a detection target (S07), and when “starting end” is set (YES in S07), the utterance start end time is output. (S08), and voice output is started (S09). Further, “end” is set in the parameter to be detected (S10), the process returns to S03, and the same processing is continued thereafter.

  Further, in the process of S07, when “starting end” is not set in the detection target parameter (NO in S07), it is determined that the detection target is “end”, and the time of the utterance end is output (S11). Also, the output of the voice is stopped (S12).

  Next, it is determined whether or not to continue the utterance section detection process (S13). If it is continued (YES in S13), “starting end” is set as the parameter to be detected (S14), and the process returns to S03. The same process is continued.

  If there is no voice input in the process of S03 (NO in S03), or if the speech section detection process is not continued in the process of S13 (NO in S13), the process ends.

  As described above, an utterance section for speech can be detected quickly and with high accuracy by the utterance section detection process using the utterance section detection program. Further, by installing the program, it is possible to easily realize the speech segment detection process.

  In the utterance section detection processing, the utterance start time and utterance end time are output (S08, S11), and the voice of the utterance section is further output (S09, S12). At least one of the speech start time, speech end time, and speech in the speech section may be output.

  As described above, according to the present invention, it is possible to detect an utterance section for speech quickly and with high accuracy. Specifically, the present invention integrates a subword acoustic model and a subword language model of a plurality of speaker clusters for an acoustic feature amount composed of short-time power and frequency characteristics and dynamic feature amounts thereof. A highly accurate and simple subword network is constructed, and the cumulative likelihood in each acoustic model corresponding to speech and non-speech is calculated and compared in synchronization with the input speech during execution of continuous speech recognition in units of subwords for the input speech. By doing so, it becomes possible to automatically detect an utterance section of a human voice in the input voice with high accuracy and with a small delay time even in various acoustic environments where background noise exists.

  Therefore, by using the present invention for speech recognition preprocessing, non-speech sections unnecessary for speech recognition, such as silence, noise, and music, are automatically removed from the input speech, and only the speech sections to be recognized are extracted. Can do. Thereby, the amount of speech recognition processing is reduced and the recognition performance is improved.

  Further, by using the present invention for speech compression pre-processing, it is possible to selectively apply an optimum compression method to each of the speech period and the non-speech period, thereby improving the compression efficiency. In addition, by using the present invention for automatic labeling of a speech database, labeling of speech sections and non-speech sections and division into files can be automated, and work efficiency can be improved. Also, by using the present invention for voice transcription text creation support, it is possible to extract only the speech section from the voice and automatically add time information of each utterance in the voice, thereby improving work efficiency. Can do.

  Furthermore, by using the present invention for a recording apparatus, it is possible to record only a speech section, and it is possible to save a recording medium such as a tape or a memory.

  In other words, the present invention is applied to technologies in various fields using speech recognition and language processing, such as subtitle production of broadcast programs, voice dialogue systems, voice word processors, automatic creation of meeting minutes, and control of devices by voice. be able to.

  The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to such specific embodiments, and various modifications, within the scope of the gist of the present invention described in the claims, It can be changed.

It is a figure which shows one structural example of the utterance area detection apparatus in this invention. It is a figure which shows an example of a subword network when the number of speaker clusters is two. It is a figure which shows an example of the speech recognition in the utterance start end. It is a figure which shows an example of the speech recognition in the utterance termination | terminus. It is a figure which shows an example of the hardware constitutions which can implement | achieve the speech area detection process in this invention. It is a flowchart which shows an example of an utterance area detection process procedure.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speaking section detection apparatus 11 Subword network integration means 12 Acoustic analysis means 13 Continuous speech recognition means 14 Speaking section detection apparatus 21 Subword acoustic model 22 Subword language model 23 Subword network 24 Input speech 25 Acoustic feature amount 26 Subword sequence and them 27 utterance start time 28 utterance end time 29 utterance interval sound 31 utterance detection start state 32 non-speech acoustic model of speaker cluster A corresponding to utterance start end 33 speech acoustic model of speaker cluster A 34 equivalent to utterance end Non-speech acoustic model of speaker cluster A 35 Non-speech acoustic model of speaker cluster B corresponding to the beginning of speech 36 Speech speech model of speaker cluster B 37 Non-speech acoustic model of speaker cluster B corresponding to the end of speech 38 Speech test Output end state 41 Input device 42 Output device 43 Drive device 44 Auxiliary storage device 45 Memory device 46 CPU
47 Network connection device 48 Recording medium

Claims (8)

In the utterance section detection device for detecting the utterance section from the input voice,
Acoustic analysis means for converting the input speech into acoustic features;
Continuous speech that sequentially calculates the cumulative likelihood in each subword in synchronization with the input speech using the acoustic feature obtained by the acoustic analysis means and a subword network consisting of a preset acoustic model and language model Recognition means;
An utterance section detecting device comprising: an utterance section detecting means for sequentially detecting an utterance start end and an utterance end in the input speech from an accumulated likelihood in each subword.   The subword network using a subword acoustic model having one or a plurality of speaker clusters expressing acoustic features of speech and non-speech sounds, and a subword language model representing a transition between the subword acoustic models. The utterance section detecting device according to claim 1, further comprising subword / network integration means for integrating the utterances. The subword network integration means includes:
Transition from the speech section detection start state to an acoustic model corresponding to non-speech of all speaker clusters, according to a subword language model from the non-speech acoustic model to an acoustic model corresponding to speech of each speaker cluster Transitions, transitions from non-speech acoustic models back to the speech segment detection start state to absorb non-speech over a certain length of time, transitions according to subword language models between acoustic models corresponding to speech of each speaker cluster, Penalized transition from an acoustic model corresponding to the speech of each speaker cluster to an acoustic model corresponding to the speech of a different speaker cluster, an acoustic model corresponding to each non-speech from the acoustic model corresponding to the speech of each speaker cluster Transition according to the subword language model, utterance termination Forming a subword network that enables at least one of a transition from the utterance interval detection end state to the utterance interval detection start state according to the output condition, and a transition from the utterance interval detection end state to the utterance interval detection start state. The utterance section detection device according to claim 2, wherein The continuous speech recognition means includes
After transition from the speech section detection start state in the subword network to the acoustic model corresponding to non-speech or speech, the speech section detection is performed from the non-speech acoustic model when the speech start detection condition is not satisfied for a certain length of time. At the same time as returning to the start state, the intermediate results of speech recognition such as cumulative likelihood in all acoustic models are cleared, the speech segment detection start time is updated, and continuous speech recognition in units of subwords is started again. The utterance section detection apparatus according to any one of claims 1 to 3. The utterance section detecting means includes
When detecting the utterance start edge, the maximum cumulative likelihood and the utterance interval detection start state among the acoustic models corresponding to the speech of all speaker clusters for all input speech from the utterance interval detection start time to the current time Next, the ratio of the cumulative likelihood of the acoustic model corresponding to the non-speech of the same speaker cluster is sequentially calculated in synchronization with the input speech, and based on the calculated ratio value and a preset threshold value, 5. The time according to any one of claims 1 to 4, wherein a time that is a certain length of time from the end time of the non-speech acoustic model at the start of the subword string indicating the maximum cumulative likelihood is detected as the start of speech. The utterance section detection device described. The utterance section detecting means includes
When detecting the end of the utterance, for all input speech from the utterance interval detection start time to the current time, the largest cumulative among the acoustic models corresponding to non-speech following the acoustic model corresponding to speech of all speaker clusters The ratio between the likelihood and the maximum cumulative likelihood of the acoustic model corresponding to the speech of the same speaker cluster is calculated sequentially in synchronization with the input speech, and the calculated ratio value is preset over a certain length of time. 6. The utterance section detection device according to claim 1, wherein when the threshold value is exceeded, a time that is a certain length of time after the time when the threshold starts to be exceeded is detected as an utterance end point. 6. The utterance section detecting means includes
The utterance section detection device according to any one of claims 1 to 6, wherein a voice of an utterance section is output from the input voice based on time information of the utterance start end and the utterance end. In an utterance interval detection program for causing a computer to execute an utterance interval detection process for detecting an utterance interval from input speech,
An acoustic analysis process for converting the input speech into acoustic features;
Continuous speech that sequentially calculates the cumulative likelihood in each subword in synchronization with the input speech using the acoustic feature obtained by the acoustic analysis processing and a subword network composed of a preset acoustic model and language model Recognition processing,
An utterance interval detection program for causing a computer to execute an utterance interval detection process for sequentially detecting an utterance start end and an utterance end from an accumulated likelihood in each subword.

Images