JPH02184915A - Speech recognition device - Google Patents

Abstract

PURPOSE:To execute speech recognition processing by voice data without generating the omission of a leading part by storing input voice data just before a start command is supplied sequentially in a ring buffer, and performing the detection of a voice section and the extraction processing of the voice data by using the input voice data. CONSTITUTION:The input voice data just before a command to instruct the start of the speech recognition processing is inputted is stored in the ring buffer 3 sequentially extending over the several constant number of frames. Therefore, even when the leading part is omitted in the input voice data stored in a data buffer 4, the detection of the start of the voice section can be performed by supplementing an omitted part by the voice data stored in the ring buffer 3, and the data in the voice section can be extracted without defeating the data unexpectedly. In such a way, even when vocalization is started prior to the input of the command to start the speech recognition processing, the vocalization can surely be recognized and processed.

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial Application Field) The present invention relates to a speech recognition device that can reliably recognize input speech without dropping the beginning of input speech data.

(Prior Art) Recently, voice recognition processing has been attracting attention as one of the important technologies for realizing a highly natural human-machine interface, and various research and developments have been conducted. This type of speech recognition processing is
Basically, ∎ analyzes the input voice data collected from the microphone, ∎ detects the input voice section according to the analysis results, ∎ detects the characteristic data series (input voice pattern) of the input voice section, and This is done by comparing the standard pattern of the recognition target category, and (2) determining the matching result to identify the input voice. Therefore, the question is how to accurately detect the voice section from the input voice data, extract the input voice data (sequence of feature data) of that voice section, and use it for matching with dictionary patterns.
This is a major factor in determining its recognition performance.

However, in conventional speech recognition devices, the speech recognition processing operation is started upon receiving a command to start recognition processing from the host system (control unit), and after this operation starts, the recognition processing is executed on the input speech data. It becomes. However, it is generally very difficult to match the timing of giving the above command with the timing of the speaker's start of utterance, and usually the host system (control unit)
At regular intervals from then on, a voice request is prompted using means such as synthesized voice output or display display, and the voice uttered by the speaker is captured at the same timing and subjected to voice recognition processing.

Recently, a proximity sensor such as an optical sensor or an ultrasonic sensor has been installed near the microphone that captures the sound, and the above-mentioned commands are issued by sensing when the speaker is approaching the microphone and trying to speak. ing. According to a voice recognition device that uses such a proximity sensor to set the start timing of a voice recognition processing operation,
It is possible to realize a better man-machine O interface with high naturalness.

However, there is no guarantee that all speakers will start speaking after getting close to the microphone, and experienced speakers often start speaking while approaching the microphone, that is, before they are close enough to the microphone. Sometimes I end up doing it.

In such a case, the command to start speech recognition processing will be issued after the start of utterance, which will cause the beginning of the input speech to be omitted, resulting in failure to detect the start of the input speech section or erroneous recognition. It is undeniable that this is a factor in the vocalization of Various efforts have been made to improve the sensitivity of the proximity sensor, such as issuing a command at a slightly earlier timing, or issuing a request to speak after a predetermined period of time after detecting the approach of the speaker. , it is extremely difficult to set the timing due to individual differences among speakers, and the reality is that there is no essential solution.

(Problem to be Solved by the Invention) Conventionally, even when the approach of a speaker is detected and a command to start speech recognition processing is given, the speech starts before the recognition processing starts. There is a problem that erroneous recognition or recognition rejection is likely to occur due to the omission of the beginning part of the voice at the start of the voice recognition processing operation.

The present invention was made in consideration of these circumstances, and its purpose is to perform highly accurate speech recognition by reliably collecting input speech from the beginning without requiring complicated timing settings. An object of the present invention is to provide a highly practical speech recognition device with a simple configuration that allows for easy recognition.

[Structure of the Invention] (Means for Solving the Problems) The speech recognition device according to the present invention has a data buffer that stores input speech data after a command to start recognition processing is input, and a data buffer that stores input speech data after the command to start recognition processing is input. a ring buffer that sequentially stores a predetermined number of frames of input audio data up to immediately before the input audio data, and when the start end of the audio section of the input audio is not detected from the input audio data stored in the data buffer, The present invention is characterized in that voice section detection is performed by consecutively inputting voice data stored in the data buffer and voice data stored in the data buffer.

(Function) According to the device configured as described above, input voice data up to just before a command instructing the start of voice recognition processing is input is sequentially stored in the ring buffer over a fixed number of frames. Therefore, even if the beginning part of the input audio data stored in the data buffer is missing, the missing part can be compensated for using the audio data stored in the ring buffer and the beginning of the audio section can be detected. It is possible to extract the data of that voice section without missing anything.

As a result, even if speech is started prior to inputting a command to start recognition processing, the speech can be reliably recognized. Moreover, input speech can be easily and reliably recognized without requiring complicated timing adjustments as in the conventional method.

(Embodiment) Hereinafter, a speech recognition device according to an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of an embodiment of the apparatus, and l is a feature extraction unit that acoustically analyzes input speech collected through a microphone and extracts feature data thereof. This feature extraction part
is configured to perform, for example, bandpass filtering processing, LPC analysis, etc. at a predetermined frame period to sequentially improve the characteristic parameters of the input speech.

The human voice data (sequence of feature data) taken in through the feature extractor 1 is selectively input to the ring buffer 3 and the data volume buffer 4 via the input switch 2. This input switch 2 operates under the control of the control section 5, and normally supplies the input audio data analyzed by the feature extraction section 1 to the ring buffer 3.
The control unit 5 is configured to supply input audio data to the data buffer 4 from the time when a command to start the recognition processing operation is input to the control unit 5 .

Here, the ring buffer 3 has a storage capacity to sequentially store the input audio data over, for example, 50 frames, and when the amount of stored data reaches its limit, the oldest data is discarded and the latest data is stored. By storing , the latest 50 frames of input audio data are always stored, and the data storage operation continues until the input switch 2 is switched. In other words, the latest 50 frames of input voice data up to the time when the start of voice recognition processing is instructed by inputting a command to the control unit 5 are stored.

When the input switch 2 is switched to start storing the input audio data in the data buffer 3 in response to the input of a command to start the recognition processing operation, the input audio data stored in the ring buffer 2 up to the time of the command input is are sequentially read out to the threshold calculation unit 6 while maintaining their stored state. The threshold calculation section 6 sets a first threshold A, which is the basis of the speech section detection process, from the 50 frames of input speech data read out from the ring buffer 3, and supplies this to the speech section detection section 7. be. Specifically, the threshold calculation unit 6 calculates the average power of the 50 frames of input audio data stored in the ring buffer 3, and calculates the power level of the ambient noise and the power level of the input audio based on this average power. A first threshold value A is set for roughly discriminating.

In accordance with the first threshold value A set in this manner, the speech section detection unit 7 detects a speech section in the input speech data according to the processing procedure shown in FIG. 2, for example, and detects the speech section detected by this processing. The input audio data (sequence of feature data) is given to the similarity calculation unit 8. The similarity calculation unit 8 calculates the similarity between the series of feature data of the speech section (input speech pattern) and the standard pattern of the recognition target category stored in the standard pattern memory 9 in advance, and calculates the similarity for each recognition target. The recognition results are obtained by comparing the similarities between the categories and evaluating the similarity calculation results.

Here, to explain the characteristic voice section detection process of this device, the voice section detection section 7 determines whether or not the start of a voice section is detected from the input voice data stored in the data buffer 4, and detects the start of the voice section. When detection is unsuccessful,
The recognition process target range is expanded to 50 frames worth of input voice data stored in the ring buffer 3, and the voice section detection process is executed.

That is, when a command to start speech recognition processing is input, first, a first threshold value A is set in the threshold calculation section B based on the input speech data stored in the ring buffer 3 (step a). Thereafter, the voice section detecting section 7 detects the start end of the voice section. To detect the start of this voice section detection, first, input voice data stored in the data buffer 4 is sequentially read out, and it is determined whether the power of the input voice data exceeds the first threshold A or not.Step b) , when the power of the input audio data changes beyond the first threshold value A, the next step is to determine whether or not a peak of the audio power is detected (step C). When the peak value of the voice power is detected in this way, a second threshold value B for start detection is set based on this detection timing and according to the peak value of the voice power, and the human power stored in the data buffer 4 is set. It is determined whether a start end is detected from the audio data (step d).

At this time, if a peak of audio power cannot be detected from the input audio data stored in the data buffer 4, the target range for peak detection is
The input audio data stored in the input audio data is extended to perform detection processing (step e). Furthermore, even after detecting the peak of voice power from the input voice data stored in the ring buffer 3 in this way, or even if the peak value of voice power is detected from the input voice data stored in the data buffer 4, the data If the start of the voice section cannot be detected from the input voice data stored in the buffer 4, then the target range for detecting the start of the voice section is extended to include the input voice data stored in the ring buffer 3. , the start end is detected from this data (step f).

After the start end of the input audio data stored in the data buffer 4 or the ring buffer 3 is detected in the above manner, a point after the peak point of the audio power of the input audio data stored in the data buffer 4 is detected. The end of the input voice section is detected by examining the data on the side (
Step g).

Note that if the input audio data stored in the data buffer 4 described above does not exceed the first threshold A,
It is assumed that no voice input has been made and error processing is activated (step h). In this way, an input speech section is detected, and the input speech data of the section is selectively extracted and subjected to the above-mentioned speech recognition processing (comparison with a standard pattern).

Such voice section detection processing will be explained in more detail with reference to FIG. FIG. 3(a) shows the process of detecting a voice section in a state where utterance is started after a command is input and data of the input voice is stored in the data buffer 4. In this case, by sequentially checking the power of the input audio data stored in the data buffer 4 after the command is input, a frame section in which the power exceeds the first threshold value A is detected, and the maximum power value among them is detected. A frame P that takes . Therefore, the input audio data is traced backwards using this maximum power frame P as a reference, and the frame S in which the power of the input audio exceeds the second threshold value B for the first time is detected. The frame S detected in this way becomes the start of the audio section. However, once the start point S of the voice section is detected in this way, the input voice data is sequentially traced using the maximum power frame P as a reference,
A frame E in which the power of the input audio first falls below a second threshold B is detected. Frame E detected in this way becomes the end of the audio section, and there is a starting point S and a ending point E.
A frame section defined by is detected as a voice section. It goes without saying that the threshold for detecting the start end and the threshold for detecting the end may be set differently.

On the other hand, if the voice starts slightly before the command is input, the first part of the voice is stored in the ring buffer 3, as shown in FIG. 3(b), and the data buffer 4 is The input audio data will be stored with the leading portion missing. In such a case, even if the human voice data is traced backwards based on the maximum power frame P detected as described above, the input voice data stored in the data buffer 4 will still have a voice power that is below the second threshold. cannot be detected. Therefore, the search target is expanded to include the input voice data stored in the ring buffer 3 described above, and a frame S in which the power of the human voice exceeds the second threshold value B is detected from this data. As a result, the start point S of the voice section is detected from the input voice data that is stored in the ring buffer 3 and that is continuous before the input voice data stored in the data buffer 4. Thereafter, by detecting the end E of the voice section in the same manner as in the example described above, it becomes possible to detect all of the input voice data without omitting the beginning of the input voice.

By the way, if the utterance starts much earlier than the input timing of the command, as shown in Fig. 3 (C).
As shown in FIG. 3, there are cases where a peak frame of audio power cannot be detected from the input audio data stored in the data buffer 4. In other words, since the input audio data stored in the data buffer 4 exceeds the first threshold value A from the beginning of the input audio data, the input audio data stored in the data buffer 4 exceeds the first threshold A. , it may not be possible to determine whether the peak is the true peak frame of audio power or not.

In such a case, the audio power peak value detection range itself is expanded to include the input audio data stored in the ring buffer 3, and the peak frame is detected from among these continuous input audio data.

Then, using the detected peak frame as a reference, the leading edge of the input audio is extracted by detecting its peak S and terminal E from a series of input audio data stored in the ring buffer 3 and data buffer 4. It becomes possible to accurately detect the voice section and extract the input voice data of the detected voice section without any loss.

In addition, if the command is manually input much later than the timing of the utterance, the power of the input voice data stored in the data buffer 4 will exceed the first threshold value A as shown in FIG. 3(d). , and since it is almost impossible to obtain all of the input audio data even if you trace the input audio data back through Recken Buffer 3, in this case, start error processing. Take measures such as prompting the user to re-enter the voice.

According to the present device configured in this way, even if the timing at which the command to start the recognition processing operation is given and the timing at which the voice utterance starts is different, and the input of the command is slightly delayed from the timing at which the voice is input, Even if there is, the beginning part of the voice uttered before the command input timing is stored in the ring buffer 3, so by using this data appropriately, the voice section can be detected without dropping the beginning part of the input voice. This allows the input audio data to be extracted accurately. In addition, data extraction can be performed without missing the beginning of the audio by simply storing input audio data for a certain number of frames up to just before a command is given. As a result, it becomes possible to easily and reliably detect input voice data and submit it to voice recognition processing without making complicated timing adjustments, making it possible to suppress the occurrence of erroneous recognition and recognition rejects. .

Note that the present invention is not limited to the embodiments described above. For example, the number of frames of audio data to be sequentially stored in the ring buffer 3 can be determined by taking into consideration the frame period and timing deviation. Further, as the speech recognition processing method itself, various methods that have been proposed in the past can be appropriately adopted. Further, the threshold value setting algorithm is not particularly limited, and can be implemented with various modifications without departing from the gist thereof.

[Effects of the Invention] As explained above, according to the present invention, the input voice data up to just before a recognition processing operation start command is given is sequentially stored in a ring buffer, and the input voice data stored in this ring buffer is Since the data is used appropriately to detect voice sections and extract the voice data, it is possible to perform voice recognition processing using voice data without missing the beginning part very easily and effectively. This produces a certain effect.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram of the main parts of a speech recognition device according to an embodiment of the present invention, FIG. 2 is a diagram showing the flow of a speech section detection process in the embodiment device, and FIG. 3 is an operation of the embodiment device. FIG. 1...Feature extraction unit, 2...Input switch, 3...
Ring buffer, 4... Data buffer, 5... Control unit, 6... Threshold calculation unit, 7... Voice section detection unit,
8... Similarity calculation unit, 9... Standard pattern memory. Applicant's Representative Patent Attorney Takehiko Suzue

Claims (1)

[Claims] The data buffer includes a ring buffer that sequentially stores a certain number of frames of input audio data up to just before a command to start recognition processing is input, and a data buffer that stores input audio data after the command is input. When the start of the input voice is not detected from the input voice data stored in the ring buffer, the input voice data stored in the ring buffer and the voice data stored in the data buffer are consecutively connected to detect a voice section. A speech recognition device characterized by: