KR20060065262A - Storage method for nametag in speaker dependent speech recognition system - Google Patents

Abstract

According to the present invention, in the speaker-dependent speech recognition system, in particular, an end point of an input speaker-dependent speech can be detected and a speech waveform can be stored as much as possible by eliminating unnecessary silence existing between a recording time of a name tag and a speech utterance time. In the speaker-dependent speech recognition system according to the present invention, the name tag storing method detects an end point of an input voice signal when a speaker voice is input, separates silence and voice, and detects a voice section of the voice signal. The voice start time point is compared with a preset reference time, and as a result of the comparison, the voice start time point is shifted forward by a predetermined multiple of the reference time according to the difference between the voice start time point and the reference time point, and the shifted time interval The speech waveform from which the silence is removed is stored. Through this, it is possible to prevent waste of memory by unnecessary mute when the name tag is stored in the speaker-dependent speech recognition system and to improve the system response characteristics.

Speaker dependent speech recognition, name tag, mute

Description

STORAGE METHOD FOR NAMETAG IN SPEAKER DEPENDENT SPEECH RECOGNITION SYSTEM}

1 is a block diagram showing a speaker-dependent speech recognition navigation system according to the present invention.

2 is a block diagram showing a speech processing process for speaker-dependent speech recognition according to the present invention.

3 is a flowchart illustrating a method of storing a name tag in a speaker dependent speech recognition system according to an exemplary embodiment of the present invention.

Figure 4 (a) is a first voice utterance form according to the present invention, (b) is a waveform diagram showing an example of the name tag storage for the first voice utterance form.

Figure 5 (a) is a first voice utterance form according to the present invention, (b) is a waveform diagram showing an example of storing the name tag of the second voice utterance form.

<Description of the symbols for the main parts of the drawings>

11 ... GPS satellite 13 ... Receiver

21.Remote control transmitter 22 ... Remote control receiver

30.Microphone 40 ... Preprocessing part

50 ... Voice processing unit 51 ... Endpoint detection unit

60 ... Memory 70 ... CD-ROM driver

80 ... Display section 90 ... Control section

According to the present invention, in the speaker-dependent speech recognition system, in particular, when detecting an end point of the speaker-dependent speech and storing a name tag, it is possible to store a speech waveform that can remove unnecessary silence between recording time and speech utterance as much as possible. The present invention relates to a method of storing name tags in a speaker-dependent speech recognition system.

Voice information processing technology refers to a technology that analyzes voice signals spoken by a person, grasps the meaning inherent in it, or finds out the identity of a speaker, and a technology that generates information in the form of text into a human voice. do.

In other words, it means a technology that enables the use of voice as a means of communication between people and machines. Voice is the most natural communication tool for human beings, which is an important position in the interface between human and machine, and the necessity of voice information processing technology is emerging to realize this.

Types of speech recognition systems are classified into speaker-independent systems and speaker-dependent systems according to makeup.

The first speaker dependent system is a system for recognizing the voice of a specific speaker, and is typically a voice dialing system used in mobile phones. In speaker-dependent systems, a pattern matching technique is generally used that stores and registers a user's voice and compares the pattern of the input voice with the stored voice prior to using the system.

For example, a voice phone card serviced by Sprint stores up to 30 names, and when a call is made, a person's name is spoken, and the person's phone number is automatically found.

The second speaker independence system is for recognizing the voice of an unspecified majority speaker, and there is no need for the user to register the voice before the operation of the system as in the speaker dependent system. The speaker-independent system collects the voices of the majority speakers to learn statistical models and recognizes them using the learned models. Therefore, the characteristic characteristics of each speaker disappear, and the characteristics common to each speaker are highlighted.

For example, AT & T has applied a toll free telephone and performs a function of recognizing whether the recipient will accept the toll telephone.

Among the speech recognition systems, speaker-dependent speech recognition has a higher recognition rate than speaker-independent speech recognition, which is advantageous for practical use. If the same amount of learning data is used for the same vocabulary, the speaker-dependent system generally outperforms the speaker-independent system.

Conventional speaker dependent speech recognition is recognized by recording its own voice and extracting voice feature parameters from the voice. This is commonly used for applications that call by name. In addition to the voice feature parameters of the recorded file, the voice waveform itself must be stored in non-volatile memory by some compression method.

In addition, the stored compressed speech waveform can be used not only to confirm the speech recognition result, but also to allow the user to listen to the recorded speech waveform to easily identify which names are registered. This can easily be done using a speech synthesizer (TTS), but it is absolutely necessary in a system that does not use a speech synthesizer because there must be an element to replace PLAYBACK.

However, in the method of storing a voice in a conventional speaker dependent speech recognition system, the speaker voice storage (ie, nametag) is fixed to a predetermined time (for example, 3 to 5 seconds) and then the voice is accepted. However, since the voice is recorded for a certain time, in the embedded solution (EMBEDDED SOLUTION), unnecessary memory causes a waste of memory when viewed from the system optimization point of view.

For example, if you store about 100 nametags, you should have a memory to store about 100 5 seconds. In addition, even when compressing a speech waveform stored using an actual encoding technique, unnecessary silences are put in the front and back of the actual speech, thereby causing a problem of unnecessary storage of redundant memory.

A first object of the present invention is to remove unnecessary silence from a speech waveform used as a playback for confirming whether a registered person's name, telephone number, or the like is properly recognized in a speaker dependent speech recognition system.

A second object of the present invention is to reduce the memory size by storing the audio waveform by eliminating the silence section occurring between the prompt time and the voice ignition time for playback.

A third object of the present invention is to detect an end point of an input voice and compare the reference time with the end point, thereby eliminating unnecessary silence in the name tag.

Voice storage method in a speaker-dependent speech recognition system according to the present invention for achieving the above object,

In the name tag storage method of the speaker-dependent speech recognition system,

Detecting an end point of the input voice signal and separating the silence and the voice when the speaker voice is input;

Detecting a voice section of the voice signal and comparing a voice start time point with a preset reference time;

As a result of the comparison, shifting the voice start time forward by a predetermined multiple of the reference time according to the difference between the voice start time and the reference time, and storing the voice waveform from which the silence is removed for the shifted period. It is characterized by including.

Preferably, as a result of the comparison, when the voice start time is greater than a preset reference time, the voice start time is shifted forward by a multiple of the reference time, and the voice waveform from which the silence is removed for the shifted period is stored. It features.

Preferably, when the voice start time is less than a preset reference time, the voice start time is shifted forward by 1/2 times the reference time, and the voice waveform from which the silence is removed by the shifted interval is stored. Characterized in that.

Preferably, the section in which the silence is removed is characterized in that silence is removed by a predetermined multiple of the reference time from the prompt start point of the name tag.

A speaker dependent speech recognition navigation system according to the present invention will be described below.

1 is a speaker dependent speech recognition navigation system according to the present invention.

Referring to FIG. 1, a receiver calculating a longitude and latitude of a point from a GPS antenna 12 receiving a position signal from three or more GPS satellites 11 and a position signal received by the GPS antenna 12. 13, a remote controller transmitter 21 for remotely controlling the navigation system, a remote controller receiver 22 for receiving a navigation system control signal transmitted by the remote controller transmitter 21, and a microphone 30 for inputting voice. ), A pre-processing unit 40 (hereinafter referred to as an analog / digital conversion unit) for converting the voice input to the microphone 30 into digital signal processing, and a digital output of the analog / digital conversion unit 40. A voice processing unit 50 for extracting the data, a memory 60 for storing a user database such as a voice recognition name tag, and a CD-ROM driver 7 for reading data from the compact disk 71 containing the map data. 0), a display unit 80 for displaying information, and a controller 90 for controlling the system and removing the silence of a predetermined section of the voice recognition name tag and storing it in the memory 60.

The speaker dependent speech recognition system according to the present invention is a system for recognizing only a specific speaker's voice, and can be learned by one person's voice signal so that the person's voice can increase the recognition rate.

Speaker-dependent speech recognition records your voice (phone number, name tag, etc.), extracts voice feature parameters from the voice and stores it in non-volatile memory. The voice feature parameters are extracted in real time and recognized using previously stored parameters.

Here, if the system does not have a speech synthesizer (TTS), there is no way to properly measure whether the recognition is correct or not. Therefore, not only the voice feature parameter but also the voice waveform itself is stored in the memory to provide feedback in order to confirm whether the user is properly recognized after recognition. In other words, if the voice information "Hong Gil-dong" is stored and registered in the memory and a voice name "Hong Gil-Dong" is input from the user, the system will answer "Hong Gil-Dong?" Will give feedback.

In addition, when the voice waveform itself is stored, the embedded solution having limited memory resources occupies a lot of unnecessary memory, and thus is compressed and stored.

1 is a navigation system to which a speaker-dependent speech recognition system according to the present invention is applied.

Referring to FIG. 1, a user inputs a name tag to be input by voice through a microphone 30 (S10). A user inputs a voice through the microphone 30 and the input voice is converted into a digital signal by an analog to digital converter 40.

The analog / digital converter (ADC) 40 converts the input voice into digital and undergoes a pre-processing process. The preprocessing process converts an analog signal, a voice signal input through the microphone 30 into a digital signal, at a sampling frequency of about 10 kHz by analog / digital conversion into a digital signal, thereby converting the signal to noise ratio (S / N). In order to improve the ratio, a process such as pre-emphasis is performed.

The analog / digital converter 40 pre-processes the input voice and outputs it to the voice processor 50. When a voice is input, the voice processor 50 performs endpoint detection for distinguishing between the silence and the voice of the voice signal input through the endpoint detection unit 51 as shown in FIG. 2. The voice processor 50 detects the feature parameter of the voice that has passed through the endpoint detector 51 using the voice feature detector (not shown). This serves as an element for obtaining the speech feature parameter.

In other words, if unnecessary silence is included in the speech section, the time required for the speech recognition waveform is increased to distinguish between silence and speech through end point detection. When detecting the end point, the end point is detected by estimating the noise level.

In this way, the voice endpoint detecting unit 51 detects the silence and the voice of the voice section separately and transmits the timing information to the control unit 90 together. The controller 90 performs the silence removing operation of shifting forward with respect to the voice start time by using the voice timing information and the silence timing information.

In other words, the controller 90 shifts the voice start time forward to remove an arbitrary silence time using the reference time according to the timing information of the voice start time, and then stores the voice waveform itself.

3 is a flowchart illustrating a voice storage method in a speaker-dependent speech recognition system according to an exemplary embodiment of the present invention.

Referring to FIG. 3, when the speaker voice is input after the prompt for the use of the name tag (S11), the silence and the voice of the voice signal are detected by detecting the end point of the voice (S13). Detecting a voice section of the voice signal (S15), and comparing the voice start time point with a preset reference time to check whether it is within the reference time (S17), if the reference time is exceeded, the voice start time point corresponding to the multiple of the reference time By shifting forward (S19), the silent section corresponding to the shifted section is removed, and the speech waveform from which the silence is removed is stored in the memory (S21).

If the voice start time is less than the reference time (S23), by shifting the voice start time forward by a section corresponding to 1/2 of the reference time with respect to the voice start time (S25), 1/2 times the reference time The speech waveform from which the silence of the section corresponding to is removed is stored (S27).

That is, in FIG. 3, after setting at least one reference time in advance, it is determined whether a voice ignition time exists before / after the reference time, and according to the result, the voice ignition time is a predetermined multiple of the reference time, that is, one day of the reference time. By shifting the forward section by a multiple of one-half or a half of the reference time, the unnecessary silent section can be eliminated as much as possible. In addition, by performing a comparison between the speech utterance time point and the reference time and removing the silence at least one or more times, the name tag from which the silence is removed is stored as much as possible.

In general, when the voice length of the recorded name is spoken, the voice is uttered after a certain time. This is because the time before the actual system sends a prompt to the user to guide the user's voice and the user actually speaks the voice (eg, nametag) appears in several time zones, but most of the recorded voices are prompted after the prompt. The point in time (that is, the start point of the voice) appears after a certain time or more. Accordingly, the size of the voice waveform can be reduced by removing the silent section of the predetermined section (or time) from the start point of the prompt.

With reference to FIGS. 4 and 5, the silence of the voice signal in the present invention is removed, and two types of utterances are classified and the silence sections are removed.

Referring to Figure 4 (a), after the actual voice recording T1 (approximately 0.85 seconds) after the appearance of the voice appears, Figure 5 (a) is after the actual voice recording starts T3 (about 2 seconds) After), the voice is spoken.

Through the experimental results, it can be seen that by analyzing several waveforms, the point of rapid speech utterance rarely ignites 0.5 seconds after the start of the voice recording, which is until the user speaks after the system's voice recognition prompt. It can be seen that there is a silence gap of at least about 0.5 seconds.

As shown in FIGS. 4A and 5A, only voices are extracted and stored in a memory, that is, a nonvolatile memory area, in a manner consistent with various utterances of users, thereby reducing the memory size. That is, by removing and storing the silent sections appearing before the voice uttering points P1 and P2, the memory size can be reduced and the silent part disappears, so that the system response speed can be increased during playback.

To this end, the method of extracting speech roughly detects speech and silence sections using speech endpoints. If the voice end point is incorrectly detected, the voice is cut off, and thus the detection is rough.

After extracting the part judged to be the voice, the voice start time point P1 is compared with the reference time point R1 by comparing the voice start time point P1 with the reference time time R1 without storing the voice on the basis of the time point. If it is small, it shifts by the front part of the voice signal by the amount (0.5 seconds) corresponding to half (1/2) of the reference time (eg 1 second). That is, the silence section of the voice signal of FIG. 4A that is about half of the reference time (for example, 0.5 seconds) is removed, so that the voice start time point T1 '= T1-R1 / 2 as shown in FIG. ), The voice start time point P1 'is stored in the memory.

If the voice start time point P2 is less than the reference time R1 as shown in FIG. 5A, silence as much as the reference time may be removed.

That is, after comparing the voice start time point P2 and the reference time R1, when the voice start time point P2 is greater than the reference time, the silent section corresponding to the reference time R1 is removed. Since the voice start time point P2 occurs after about 2 seconds in FIG. 5A, the silent section corresponding to the reference time R1 is removed and T3 as shown in FIG. 5B. '= T3-R1 saves the audio signal waveform from which the silent section is removed to memory.

In addition, at least one mute removal operation may be performed on all speech signals. If the speech is started two times, if the speech start time is about 2 seconds, the silence section is removed for about 1 second, and then again the silence section (about 0.5 second) may be performed.

If the recorded voice has a voice start time within 0.5 seconds, an error message is sent to the user to request re-recording.

In other words, the voice start time is 0.8 seconds in FIG. 4A, but the silence start time of 0.5 seconds is removed in FIG. 4B so that the voice start time is about 0.3 seconds. In addition, as shown in FIG. 5A, the voice start time appears after 2 seconds, but silence of about 1 second is removed as shown in FIG. 5B, and the voice start time appears in about 1 second. Accordingly, the silent section is removed from the original audio signal, so that the voice response characteristics appear quickly during playback.

So far, the present invention has been described with reference to the preferred embodiments, and those skilled in the art to which the present invention pertains to the detailed description of the present invention and other forms of embodiments within the essential technical scope of the present invention. Could be implemented. Here, the essential technical scope of the present invention is shown in the claims, and all differences within the equivalent range will be construed as being included in the present invention.

As described above, according to the voice storage method of the speaker-dependent speech recognition system according to the present invention, when the user stores a name tag to be used as a playback, the memory size can be reduced by removing the silent section. Not only that, but it also has the effect of speeding up system response.

Claims (4)

In the name tag storage method of the speaker-dependent speech recognition system, Detecting an end point of the input voice signal and separating the silence and the voice when the speaker voice is input; Detecting a voice section of the voice signal and comparing a voice start time point with a preset reference time; As a result of the comparison, shifting the voice start time forward by a predetermined multiple of the reference time according to a difference between the voice start time and the reference time, and storing the voice waveform from which the silence is removed for the shifted period. Speech storage method in a speaker-dependent speech recognition system comprising a. The method of claim 1, As a result of the comparison, when the voice start time is greater than a preset reference time, the voice start time is shifted forward by a multiple of the reference time, and the voice waveform from which the silence is removed for the shifted period is stored. Speech storage method in speaker dependent speech recognition system. The method of claim 1, As a result of the comparison, when the voice start time is smaller than the preset reference time, the voice start time is shifted forward by 1/2 times the reference time, and the voice waveform from which the silence is removed for the shifted period is stored. A speech storage method in a speaker dependent speech recognition system characterized by the above-mentioned. The method of claim 1, The silence storage method of the speaker-dependent speech recognition system, characterized in that the silence is removed as much as a predetermined multiple of the reference time from the start point of the prompt of the name tag.

Images