TW202226225A - Apparatus and method for improved voice activity detection using zero crossing detection - Google Patents

Abstract

An apparatus and a method for improved voice activity detection (VAD) using zero crossing detection. A first VAD system outputs a pulse stream for zero crossings in an audio signal. The pulse density of the pulse stream is evaluated to identify speech. The audio signal may have noise added to it before evaluating zero crossings. A second VAD system rectifies each audio signal sample and processes each rectified sample by updating a first statistic and evaluating the rectified sample per a first threshold condition that is a function of the first statistic. Rectified samples meeting the first threshold condition may be used to update a second statistic and the rectified sample evaluated per a second threshold condition that is a function of the second statistic. Rectified samples meeting the second threshold condition may be used to update a third statistic. The audio signal sample may be selected as speech if the second statistic is less than a downscaled third statistic.

Description

Apparatus and method for improving voice activity detection with zero-crossing detection

A device and method for improving voice activity detection, particularly a device and method for improving voice activity detection by zero-crossing detection.

Smart speakers and other voice-activated devices interpret human speech as commands and perform corresponding actions. In many cases, the device can listen for keywords (such as "Alexa", "OK Google", "OK Siri"), and can listen for subsequent commands when a keyword is detected. To implement such a function, the device must always require a certain amount of power to listen for commands. One way to reduce power usage is voice activity detection (VAD), which separates noise from human speech. Using the above approach, the sound signal is evaluated to determine whether to speak a keyword only when human speech is detected.

To sum up, it can be seen that there has been a long-standing problem in the prior art that the sound signal is only evaluated to determine whether to speak a keyword when human speech is detected. Therefore, it is necessary to propose technical means to improve the implementation of voice activity detection to solve the problem. this question.

In view of the problem in the prior art that the sound signal is only evaluated to determine whether to speak a keyword when human speech is detected, the present invention discloses an apparatus and method for improving voice activity detection with zero-crossing detection, wherein:

The apparatus for improving voice activity detection by zero-crossing detection disclosed in the present invention at least includes: a processing device programmed to achieve: receiving an audio signal; crossing) generates a pulse stream; generates a pulse density stream based on the frequency of the pulses occurring over time in the pulse stream; evaluates the pulse density stream based on a threshold condition; identifies and matches the portion of the pulse density stream that meets the threshold condition The speech part of the corresponding sound signal.

The method for improving voice activity detection by zero-crossing detection disclosed in the present invention at least comprises: a processing device receives a sound signal; the processing device generates a pulse stream according to the zero-crossing detected in the sound signal; The frequency of the pulses occurring over time in the pulse stream produces the pulse density stream; the processing means evaluates the pulse density stream according to the threshold condition; the processing means identifies the speech portion of the sound signal corresponding to the portion of the pulse density stream that meets the threshold condition.

The apparatus and method disclosed in the present invention are as described above. The difference between the present invention and the prior art is that the present invention generates a pulse stream through a processing device according to the zero crossing detected in the sound signal, and generates pulses according to the pulses in the pulse stream over time. After the pulse density flow is generated at the frequency of the frequency, the pulse density flow is evaluated according to the threshold condition, and the voice part of the sound signal corresponding to the part of the pulse density flow that meets the threshold condition is identified, so as to solve the problems existing in the prior art and achieve a comparative The technical efficacy of identifying latent speech with low power and higher accuracy.

The features and implementations of the present invention will be described in detail with reference to the drawings and examples below, and the content is sufficient to enable any person skilled in the relevant art to easily and fully understand the technical means used by the present invention to solve the technical problems and the advantages of the present invention, and accordingly Implementation, thereby achieving the effect that the present invention can achieve. It is to be understood that the drawings are only used to additionally describe and explain the specificity and detail of the embodiments of the present invention, and should not be construed as limiting the present invention.

Hereinafter, the present invention will be described with reference to "FIG. 1". As shown in "FIG. 1", a speech detection system 100 (also represented as system 100 in the present invention) implements voice activity detection on an input signal. Elements of system 100 may also be represented in the form of executable code executed by a processor, different hardware elements, or other implementations. The system 100 can be used as the first device to wake up the second device in response to the speech detected in the input signal 102 . For example, the second device may be a general processor capable of performing speech-to-text, network communication, or other processing functions that can be performed by a smart speaker or other voice-controlled device.

The input signal 102 may be received by a microphone, may be a raw digital sound signal sampled by the output of the microphone, or may be obtained by preprocessing the raw digital sound signal according to one or more pre-process steps. The resulting results, such as low-pass filtering, scaling, downsampling, upsampling, or other preprocessing steps.

System 100 may include bandpass filter 104 . The band pass filter 104 may have a passband corresponding to speech, such as a 3db band. In general, the frequency band can be between 0.3 and 20,000 hertz (Hz). In other embodiments, frequency bands between 1 and 2 kilohertz may also be used. The band-pass filter 104 can achieve the functions of removing any direct circuit component (DC component) in the input signal 102 and removing noise that does not correspond to speech.

Bandpass filter 104 may output a first filtered signal that is input to adder 106 . The adder 106 may add the first filtered signal and the high frequency signal 108 to generate a sum signal. The high frequency signal 108 has a frequency and an amplitude. In some embodiments, the frequencies are selected to ensure that zero-crossing occurs between each pair of consecutive samples in the high frequency signal 108 . Therefore, the frequency of the high frequency signal 108 may be equal to half (one-half) the sampling rate of the input signal 102 .

The amplitude of the high frequency signal 108 can be calibrated to the properties of the detection microphone generating the input signal 102 and to the properties of the ambient noise that the system 100 is expected to encounter. For example, sound signals can be captured from an intended environment (eg, a sound recording in a real-world environment) without speech. When the system 100 processes the audio signal as described below, the amplitude of the high frequency signal 108 may be increased until the system 100 detects no speech. The amplitude of the high frequency signal 108 may be dynamic. For example, if the feedback from the speech-to-text element indicates that the portion of the input signal determined to contain speech does not actually contain speech, the amplitude of the high frequency signal 108 may be increased to reduce false reporting determinations (false). positive). In this specification, a "portion" of a signal refers to a series of consecutive samples in the signal.

The sum signal generated after being added by the adder 106 can be input to the zero-crossing detector 110 . The output of the zero-crossing detector 110 is a pulse stream. For example, for each zero crossing, the zero crossing detector 110 may output a first numerical value, such as 1 in binary. If there is no sign change between a sample in the sum signal and the previous sample, the zero-crossing detector 110 may output a second value, eg, 0 in binary. In some embodiments, only crossings from positive values to negative values are detected as zero crossings. In some embodiments, only crossings from negative to positive values are detected as zero crossings. In yet another part of the embodiment, a crossing from a positive value to a negative value or from a negative value to a positive value is detected as a zero crossing.

The pulse stream may be input to the pulse density detector 112 . The pulse density detector 112 generates a density stream such that the pulse density detector 112 outputs a sample in the density stream for each sample in the pulse stream, the generated density stream corresponding to each sample in the pulse stream before The number of pulses (first value) for the window of N pulses. Among them, N is greater than 1, preferably N is greater than 10, and more preferably N is greater than 100.

The density stream may be input to a low pass filter 114 which outputs a second filtered signal. A cutoff frequency, such as a 3dB cutoff frequency, may be selected to achieve a desired degree of smoothing or averaging of the second filtered signal relative to the density flow. In some embodiments, the low-pass filter 114 may function as a pulse density detector, ie, the result of the low-pass filter is generally a signal that increases with increasing pulse density and decreases with decreasing pulse density, although the corresponding Relationships may not be perfect. As such, in such embodiments, the pulse density detector 112 may be eliminated.

The second filtered signal may be used by comparator 116, which may evaluate the second filtered signal for speech thresholds and may output a speech decision 120 for each sample in the second filtered signal. The speech decision 120 may be a binary value, and the input samples in the input signal 102 are processed using the system 100 such that a corresponding speech decision 120 indicating whether the input sample is likely to correspond to speech is output. Input samples in the input signal 102 identified as speech may be passed to subsequent stages to confirm that the samples indeed contain speech, to perform speech-to-text synthesis, to store for later use, or for other purposes. Alternatively, samples of the first filtered signal that correspond in time to samples of the second filtered signal may be passed to a subsequent stage for filtering with bandpass filter 104 .

Samples with values below the speech threshold 118 can be judged to correspond to speech. In particular, the modulation of low frequencies and high amplitudes caused by speech can increase the amplitude of the sum signal above that of the high frequency signal 108, resulting in a reduction in zero crossings and a corresponding reduction in pulse density.

The speech threshold 118 may be adjusted by the statistical analyzer 122 . The statistical analyzer 122 receives the input signal 102 and/or the first filtered signal and generates an analysis value over time that characterizes one or both of the signals. These statistics can contain mean, standard deviation, maximum, minimum, root mean square, percentiles below the absolute value of the input sample (eg, the 90th percentile), or other statistical values.

For example, statistical analyzer 122 may calculate the root mean square of a segment of multiple samples in the input signal, and may scale speech threshold 118 accordingly, eg, increase speech threshold 118 as the root mean square increases, and The speech threshold is reduced 118 with a reduction in rms. In another example, the statistical analyzer 122 may calculate the root mean square of a segment of a plurality of samples in the input signal, and may use the root mean square to scale the amplitude of the high frequency signal 108, eg, increase as the root mean square increases The amplitude of the high frequency signal 108 and the amplitude of the high frequency signal 108 is reduced as the root mean square decreases. Either of the above methods can dynamically reduce erroneous judgments in response to the increase and decrease of the amplitude of the surrounding noise.

"FIG. 2" illustrates a schematic diagram of the signals used and generated in accordance with the system 100. FIG. The upper graph 200a includes a speech signal 202 of the amplitude of the speech signal over a series of samples, such as the original speech signal or the filtered signal output by the speech bandpass filter 104 . Curve 204 illustrates speech decisions 120 in relation to the samples of graph 200a above, with higher numerical values representing samples recognized as speech, and lower numerical values representing non-speech. Clearly, the higher-amplitude portion with the envelope corresponding to the speech is correctly identified as speech, while the low-amplitude noise is not identified as speech.

Note that some parts identified as non-speech may correspond to specific parts of speech, e.g. unvoiced friction such as /s/, /sh/, and /f/, which are difficult to distinguish from noise . However, these parts are ephemeral and can be obtained by extending the part of speech to include a certain period (eg, less than 200 milliseconds) between the part and the part recognized as speech or at the beginning or end of the part recognized as speech.

Graph 200b illustrates a plot 206 of the amplitude of a low-pass filtered signal, eg, the output of low-pass filter 114, over a series of samples. Curve 208 represents the threshold value used by comparator 116 . In this embodiment, samples of the input signal 102 corresponding to samples of the low-pass filtered signal below the threshold will be recognized as speech.

The present invention will be described with reference to "FIG. 3". As shown in "FIG. 3", the voice activity detection system 300 (also represented as system 300 in the present invention) implements another approach to achieve voice activity detection. System 300 is more complex than system 100 at the expense of more complex calculations. However, the system 300 is still very computationally efficient and can implement the state of the voice activity detection algorithm stored between the processed samples using only low storage space requirement addition, multiplication and subtraction operations. As such, system 300 can be used to implement voice activity detection to trigger wake-up of a processing device capable of performing more complex operations than system 300, such as a general purpose processor.

In some embodiments, the schemes of "Fig. 3" and "Fig. 4" are implemented after the voice activity detection is implemented using the system 100, even though the portion of the signal identified as speech using the system 100 can be processed using the system 300 to Confirm that the signal does contain speech. System 300 may be represented by executable code executed by a processor, various hardware elements, or other implementations. System 100 and system 300 may be implemented by different executable codes on the same hardware device or on different hardware devices.

The system 300 can receive the input signal 302, which is the original audio signal or a filtered version of the audio signal. The input signal 302 may be the portion of the input signal 102 that is recognized by the system 100 as speech. Alternatively, a portion of the signal produced by bandpass filtering (the output of bandpass filter 104 ) may be used as input signal 302 .

When the input signal 302 is not band-pass filtered, the input signal 302 can be processed by the speech band-pass filter 304 to obtain the first filtered signal. The speech bandpass filter 304 may be configured as the bandpass filter 104 described above.

The system 300 may also include a Teager energy calculator 306 . The Tiger energy calculator 306 outputs the Tiger energy signal (T) of the signal input to the Tiger energy calculator 306 (the input signal 302 or the first filtered signal). For example, for a given input signal (s) the Tiger energy (T[n]) of an individual sample (s[n]) of the input signal (s) can be calculated according to equation (1). In formula (1), k is a time offset, such as a value from 1 to 5. The value of k can be a function of the sampling rate, and can be higher as the sampling rate increases. T[n] = (s[n]*s[n]) – (s[n-k]*s[n+k])…Formula (1)

System 300 may include rectifier 308 . The rectifier 308 outputs the absolute value of the input signal (the input signal 302, the first filtered signal, or the Tiger energy signal).

System 300 may also include a first low pass filter 310 . For example, for the input signal designated as x (the output of the rectifier 308), low pass filtering may be performed to obtain a first low pass signal. The first low-pass signal may be input to a selective sampling stage 312, which selects samples of x with reference to the first low-pass signal. The selective sampling stage 312 may select those samples whose magnitudes are statistically significant outliers relative to the first low-pass signal. Since the selection is based on the properties of x, it may be non-uniform, i.e. at non-uniform intervals according to the change in the magnitude of x.

Those samples of x that are selectively sampled in the selective sampling stage 312 may be input to a second low pass filter 314 to obtain a second low pass signal. The samples of x sampled in selective sampling stage 312 may then be selectively sampled again in selective sampling stage 316, resulting in a further reduction of the samples. The selective sampling in the selective sampling stage 316 may be selected from the samples selected in the selective sampling stage 312 having magnitudes of statistically significant outliers relative to the second low-pass signal.

The samples of x selected in selective sampling stage 316 may be low pass filtered again at low pass filter 318 to obtain a third low pass signal. The third low pass signal may be further processed, eg by a scale down stage 320 to obtain a scaled down signal. In some embodiments, this may include multiplying the third low pass signal by a downscaling factor less than one. The function of the downscaling stage 320 may be to at least partially compensate for the fact that the third low-pass signal is taken from the samples of the higher amplitude x remaining after the selective sampling stages 312 and 316 . The reduction factor can be chosen experimentally for a given situation, for example, by gradually reducing the reduction factor from 1 until the number of false positives reaches a desired value, such as 0.1% of the sample.

Differential stage 322 may calculate the difference between the downscaled signal and the second low-pass signal to obtain a difference signal. For example, for a sample in the downscaled signal, a sample in the first lowpass signal with the same index or in the same position in the series of samples of the second lowpass signal can be identified and the downscaled signal subtracted from the sample.

The samples in the difference signal may be interpreted to obtain speech decisions 324 . In some embodiments, the downscaling factor may be chosen such that those disparity values greater than zero are likely to be speech samples with acceptable confidence. Obviously, this will happen when the second low pass signal is smaller than the downscaled signal. When the difference value is determined to correspond to speech, samples of the input signal 302 with the same index may be determined to correspond to speech and may be passed to another device or another processing stage to perform speech-to-text analysis or other processing.

Various modifications may be made to system 300. For example, a single low pass filter 310 and selective sampling stage 312 may be used, followed by low pass filter 318, and low pass filter 314 and selective sampling stage 316 may be omitted. Alternatively, one or more combinations, each comprising a low pass filter followed by a selective sampling stage, may be inserted between the selective sampling stage 316 and the low pass filter 318 .

Differential stage 322 may take as input any signal from any preceding stage, such as the signal output from some or all of low pass filters 310 , 314 , 318 . The difference function can thus be a function that can include scaling, addition or subtraction of these signals to achieve the required accuracy for speech recognition.

System 300 may implement Algorithm 1 below. Algorithm 1 may be performed sequentially on each sample s[n] of the input signal 302 (n is an index from 0 to N-1, where N is the total number of samples). Although "s[n]" is used, it should be understood that signal samples derived from s[n] may be used, such as a bandpass filtered version of the input signal 302 or the calculated Tiger energy T[n], as above for the input Signal 302 or a bandpass filtered version of the input signal 302 is depicted. Algorithm 2 can be used as an alternative to Algorithm 1, where attenuation is performed to account for periods of zero amplitude or non-speech. Algorithm 1: Selective sampling of voice activity detection with reference to a low-pass filtered signal x = Abs(s[n]); //absolute value of s[n] f1 = alpha * f1 + (1-alpha) * x; if (x > m * f1) { f3 = alpha * f3 + (1-alpha) * x; } if (x > m * f3) { f5 = alpha * f5 + (1-alpha) * x; } d = (f5 * mult) - f3; if (d > 0) { speech = 1; } Algorithm 2: Selectively sampled speech activity detection with reference to a low-pass filtered signal with filter value attenuation x = Abs(s[n]); //absolute value of s[n] f1 = alpha * f1 + (1-alpha) * x; if (x + offset3 > m * f1) { f3 = alpha * f3 + (1-alpha) * x; } else { f3 = beta*f3 f5 = beta*f5 } if (x + offset5 > m * f3) { f5 = alpha * f5 + (1-alpha) * x; } else { f5 = beta*f5 } d = (f5 * mult) - f3; if (d > 0) { speech = 1; }

The computations of f1, f3 and f5 in Algorithms 1 and 2 implement low-pass filtering (low-pass filters 310, 314, 318, respectively). alpha is a low pass filter coefficient and can be a value between 0.98 and 0.9999. For example, a value of 0.99 has been found to be valid. The alpha values used to compute f1, f3, f5 or any other low pass filtering steps can be the same or different alpha values and can be used for different low pass filtering steps.

The "if" statements in Algorithms 1, 2 may correspond to selective sampling stages 312, 316. The value of m can be chosen according to the adjustment process. In some embodiments, m may be a value between 1.3 and 1.7. For example, a value of 1.5 has been found to be acceptable. The m values used in the "if" statements can be the same or different m values, and can be used to evaluate the low pass filtered signals f1 and f3, or to calculate many other low pass filtered signals however.

The reduction factor of the reduction stage 320 can be achieved by multiplying f5 by "mult" in Algorithms 1 and 2. Thus, mult may be a value smaller than the one selected above with respect to the downscaling factor in order to achieve an acceptable number of false positives.

The differencing stage 322 corresponds to the calculation of d. In the case where d is greater than zero, the samples s[n] from which x is calculated can be considered to correspond to speech according to Algorithms 1 and 2. Note that in embodiments where only one instance of filtering and selective sampling is performed, the equation "d = (f5*mult) - f3" may be replaced by "d = (f3 *mult) - f1". In a similar fashion, where more than two filters and selective sampling are performed, d can be computed as "d = (fx * mult) - fy", where fx is the filter result in the last filter instance and fy is the result of filtering in the previous filtering instance, such as the second-to-last instance.

The beta value in Algorithm 2 can be a decay factor less than 1, for example between 0.999 and 0.9999. The decay produced by multiplying beta can be very slow, indicating that for many samples, say hundreds or thousands, speech may not be detected. In the absence of a decay factor, f1, f3, and/or f5 may change suddenly, resulting in unnecessary false positives. In Algorithm 1, decay according to beta can be omitted, and the possibility of false positives is handled by subsequent stages or simply accepted.

In some embodiments, the occurrence of a series of zero-valued input samples can be handled by using offset3 and offset5 as shown in Algorithm 2. The values of offset3 and offset5 can be the same or different. The values of offset3 and offset5 may be on the order of the smallest value representable using the number of bits and format of x. For example, assuming x is a 12-bit unsigned integer (x is an absolute value, so always positive), offset3 and offset5 might equal 2^(-11) (2 to the power of -11). Alternatively, offset3 and offset5 may be equal to some multiple of the smallest representable value (eg, 2 to 10). As can be seen from Algorithm 2, when there is a series of zero-valued samples, the low-pass filter value f1 will eventually also reach zero. By adding offset3 or offset5 to a zero-valued x will still satisfy the condition of the "if" statement, thus avoiding discontinuities and ensuring that f3 and f5 will also decay to zero in response to a series of zero-valued samples. The use of offset3 and offset5 shown in Algorithm 2 can be used instead of, or in combination with, the decay using beta. Likewise, decay using beta can be used in "if

” is used when offset3 and offset5 are not used in the statement. Obviously, Algorithms 1 and 2 only require multiplication, addition and subtraction operations. The values used in multiple iterations include only alpha, m, mult, f1, f3, and f5 (and beta to implement decay). Therefore, the computational and memory requirements required to implement Algorithm 1 are very low. Therefore, Algorithm 1 provides a low-power and high-accuracy method for identifying latent speech.

"FIG. 4" shows a diagram of various signals that may exist during the implementation of the system 300. FIG. Graph 400a contains a graph 402 of the signal for speech and periodic noise periods. Curve 404 represents speech decisions for the signal sample represented by curve 402 (high values represent speech, low values represent non-speech).

Diagram 400b presents a graph of the internal signals of system 300 . Includes curve 406 for f1, curve 408 for f3, and curve 410 for f5. Obviously, each signal is smoothed relative to the previously computed signal (f3 is smoother than f1, f5 is smoother than f3). It is also evident that the periods of noise in the original signal that are not recognized as speech (curve 402) are lower than f1, f3, and f5, which are amplified by an additional m before comparison.

"FIG. 5" is a block diagram of the system 500, which can be combined with the voice activity detection system 100 and the voice activity detection system 300 as described above. System 500 may include microphone 502, which may be a single microphone or an array of microphones. The output of the microphone 502 may be pre-processed by low pass filtering, band pass filtering, or other types of processing in order to condition the output for subsequent processing.

The output of microphone 502 may be input to system 100 for voice activity detection. The system 100 identifies a first portion of the microphone output that may correspond to speech using the manner described above with respect to "FIG. 1" and "FIG. 2." Refer to "Picture 1" and "Picture 2". The first portion may be input to the system 300 for voice activity detection. For example, system 100 may wake up system 300 to process the first portion when the first portion of the region identified by system 300 is powered off or powered off. Less power is used in sleep mode than when system 300 is awake. System 300 may process the first portion and identify a second portion that may correspond to speech using the methods described above with respect to "Fig. 1" and "Fig. 2." Refer to "Picture 3" and "Picture 4". It is contemplated that some portions that are recognized as speech by system 100 will not be recognized as speech by system 300 .

The second portion recognized by the voice activity detection system 300 may be input to another voice processing system 504. Speech processing system 504 may perform any speech processing function known in the art, such as speech-to-text, speech authentication, or the like.

The elements in "FIG. 5" (system 100, system 300, speech processing system 504) may all be separate hardware devices, such as separate semiconductor chips, separate circuit boards, or separate stand-alone computing devices. Alternatively, any two or more of the elements (system 100, system 300, speech processing system 504) may be different executable modules executing on the same hardware device.

FIG. 6 shows a block diagram of computing device 600 . Computing device 600 may be used to perform various processes, such as those discussed herein.

Computing device 600 includes one or more processors 602, one or more storage devices 604, one or more interfaces 606, one or more mass storage devices 608, one or more input/output (I/O) devices 611 and display device 630 , the aforementioned processors, interfaces, and various devices are coupled to bus 612 . Processor 602 includes one or more processors or controllers that may execute instructions stored in storage device 604 and/or mass storage device 608 . The processor 602 may also include various types of computer-readable media, such as cache memory.

The storage device 604 includes various computer-readable media, such as volatile memory and/or non-volatile memory, volatile memory such as random access memory (RAM) 614, non-volatile memory such as read-only memory ( ROM) 616. The storage device 604 may also include a rewritable memory, such as a flash memory (Flash Memory).

Mass storage device 608 includes various computer readable media such as magnetic tape, platter/disk, optical disk, solid state memory (eg, flash memory), and the like. As shown in "FIG. 6", a specific mass storage device is a hard disk drive 624. Mass storage device 608 may also include various drives to enable reading from and/or writing to various computer-readable media. Mass storage device 608 includes removable media 626 and/or non-removable media.

Input/output devices 610 include various devices that allow data and/or other information to be input to or retrieved from computing device 600 . Examples of input/output devices 610 include cursor control devices, keyboards, keypads, microphones, monitors or other display devices, speakers, printers, network interface cards, modems, lenses, cameras/charge-coupled devices (Charge-Coupled Devices) Device, CCD) or other devices, etc.

Display device 630 includes any type of device capable of displaying information to one or more users of computing device 600 . Examples of the display device 630 include monitors, display terminals, video projection devices, and the like.

Interface 606 includes various interfaces that allow computing device 600 to interact with other systems, devices, or computing environments. Examples of interface 606 include any number of different network interfaces 620, such as interfaces for local area network (LAN), wide area network (WAN), wireless network, and the Internet. Other interfaces include user interface 618 and peripheral device interface 622. Interface 606 may also include one or more peripheral interfaces, such as for printers, pointing devices (mouse, trackpad, etc.), keyboards, and the like.

Bus 612 allows processor 602, storage device 604, interface 606, mass storage device 608, input/output device 610, and display device 630 to connect with other components connected to bus 612, which represents various types of bus One or more of the bus architectures, such as system bus, PCI bus, IEEE 1394, USB, etc.

For illustrative purposes, programs and other executable program elements are represented in this disclosure as discrete blocks, although it should be understood that such programs and elements may reside in different storage elements of computing device 600 at different times, and executed by the processor. Alternatively, the systems and processes described herein may be implemented in hardware or a combination of hardware, software and/or firmware. For example, one or more application specific integrated circuits (ASICs) may be programmed to perform one or more of the systems and procedures described in this disclosure.

In the foregoing disclosure, reference is made to the drawings, which form a part hereof, and which illustrate, by way of illustration, specific embodiments in which the invention may be practiced. It is to be understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the present invention. The embodiments described in the specification may include a particular feature, structure or characteristic, but each embodiment does not necessarily include the particular feature, structure or characteristic. Moreover, these terms are not necessarily referring to the same embodiment. Furthermore, when a particular feature, structure or characteristic is described in conjunction with one embodiment, it is believed to be within the purview of those skilled in the art to affect such feature, structure or characteristic in combination with other embodiments, whether or not expressly described.

Implementations of the systems, devices, and methods disclosed herein may include or utilize a special purpose or general purpose computer containing computer hardware, such as one or more processors and system memory as discussed herein. Embodiments within the scope of the present disclosure may also include physical and other computer-readable media for carrying or storing computer-executable instructions and/or data structures. Such computer-readable media can be any available media that can be accessed by a general purpose or special purpose computer system. Computer-readable media storing computer-executable instructions are computer storage media or devices or devices. Computer-readable media that carry computer-executable instructions are transmission media. Thus, by way of example and not limitation, implementations of the invention may include at least two distinct computer-readable media: computer storage media (or apparatus or devices) and transmission media.

Computer storage media (or devices or devices) include RAM, ROM, EEPROM, CD-ROM, (RAM-based) SSD, flash memory, phase change memory (PCM), other types of memory, other optical disk storage devices , disk storage device or other magnetic storage device, or any other medium general purpose or special purpose computer that can be used to store the desired code means in the form of computer executable instructions or data structures and which can be accessed by a computer.

Implementations of the devices, systems and methods disclosed in the present invention may communicate through a computer network. "Network" is defined as one or more data links capable of transmitting electronic data between computer systems and/or modules and/or other electronic devices. When information is transmitted or provided to a computer over a network or other communication connection (wired, wireless, or a combination of wired or wireless), the computer correctly considers the connection to be a transmission medium. Transmission media can include networks and/or data links, which can be used to carry the desired code means in the form of computer-executable instructions or data structures, and which can be accessed by a general purpose or special purpose computer. Combinations of the above should also be included within the scope of computer-readable media.

Computer-executable instructions include, for example, instructions and materials which, when executed at a processor, cause a general purpose computer, special purpose computer or special purpose processing device to perform a specified function or combination of functions. Computer-executable instructions may be intermediate format instructions such as binary, assembly language, etc., or even source code. Although the present invention has been described in language specific to structural features and/or methodological acts, it is to be understood that the inventive title defined in the claims is not necessarily limited to the above-described features or acts. Rather, the described features and acts are disclosed as example forms of implementing the claims.

Those skilled in the art will understand that the present disclosure may be practiced in networked computing environments with many types of computer system configurations, including dashboard, vehicle-mounted computers, personal computers, desktop computers, notebook computers, message processors, handheld devices , multiprocessor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, mobile phones, PDAs, tablet computers, pagers, routers, switches, various storage devices, etc. The invention may also be practiced in a distributed system environment where local and remote computer systems are linked by a network (via a data link directly connected by a cable, a wireless data link, or through a combination of a cable and a wireless data link ) to perform the task. In a distributed system environment, program modules may be located in both local and remote memory storage devices.

Furthermore, where appropriate, the functions described herein may be performed in one or more of the following: hardware, software, firmware, digital components, or analog components. For example, one or more application-specific integrated circuits may be programmed to perform one or more of the systems and procedures described in this disclosure. Certain terms are used throughout the description and claims to refer to specific system elements. As will be understood by those skilled in the art, elements may be replaced by different names. The present invention does not intend to distinguish between elements with different names but different functions.

Note that the sensor embodiments discussed above may include computer hardware, software, firmware, or any combination thereof to perform at least a portion of their functions. For example, a sensor may include code configured to execute in one or more processors, and may contain hardware logic/circuitry controlled by the code. These exemplary devices are provided herein for purposes of illustration and not limitation. As known to those skilled in the relevant art, the disclosed embodiments may be implemented in other types of devices.

At least some embodiments of the present disclosure have been directed to computer program products including logic (eg, in the form of software) stored on any computer-usable medium. Such software, when executed in one or more data processing devices, causes the devices to operate as described in the present invention.

While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example only, and not limitation. It will be apparent to those skilled in the relevant art that various changes in form and details may be made therein without departing from the spirit and scope of the present disclosure. Thus, the breadth and scope of the present disclosure should not be limited by any of the above-described embodiments, but should be defined only in accordance with the claims and their equivalents. The foregoing description has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the disclosure to the precise form disclosed. Many modifications and variations are possible in light of the above teachings. Furthermore, it should be noted that any or all of the foregoing alternative implementations may be used in any desired combination to form further hybrid implementations of the present invention.

100: Voice Activity Detection System 102: Input signal 104: Bandpass filter 106: Adder 108: high frequency signal 110: Zero Crossing Detector 112: Pulse Density Detector 114: Low pass filter 116: Comparator 118: Voice threshold 120: Voice Judgment 122: Statistical Analyzer 200a: Diagram 200b: Diagram 202: Voice Signal 204: Curves 206: Curves 208: Curves 300: Voice Activity Detection System 302: input signal 304: Speech Bandpass Filter 306: Tiger Energy Calculator 308: Rectifier 310: Low Pass Filter 312: Selective sampling stage 314: Low Pass Filter 316: Selective sampling stage 318: Low Pass Filter 320: Shrinking Stage 322: Differential Stage 324: Voice Judgment 400a: Diagram 400b: Diagram 402: Curve 404: Curve 406~410: Curve 500: System 502: Microphone 504: Speech Processing System 600: Computing Equipment 602: Processor 604: Storage Device 606: Interface 608: Mass Storage Device 610: Input/Output Devices 612: Busbar 614: Random Access Memory 616: Read only memory 618: User Interface 620: Web Interface 622: Peripheral Device Interface 630: Display device

FIG. 1 is a block diagram of an element for realizing voice activity detection based on detection of zero-crossing detection according to an embodiment of the present invention. FIG. 2 is a schematic diagram of a sound signal and a signal derived from the sound signal based on detection of zero-crossing detection to realize voice activity detection included in an embodiment of the present invention. FIG. 3 is a block diagram of an element for realizing voice activity detection based on the statistical characteristics of a sample of a sound signal according to an embodiment of the present invention. FIG. 4 is a schematic diagram of a sound signal and a signal derived from the sound signal for realizing voice activity detection based on the statistical characteristics of a sample of the sound signal included in the embodiment of the present invention. FIG. 5 is a block diagram of a speech processing system according to an embodiment of the present invention. FIG. 6 is a block diagram of a computing device according to an embodiment of the present invention.

100: Voice Activity Detection System

102: Input signal

104: Bandpass filter

106: Adder

108: high frequency signal

110: Zero Crossing Detector

112: Pulse Density Detector

114: Low pass filter

116: Comparator

118: Voice threshold

120: Voice Judgment

122: Statistical Analyzer

Claims (20)

A device for improving voice activity detection with zero-crossing detection, the device comprising at least: a processing device programmed to achieve: receive an audio signal; generating a pulse stream based on the zero crossing detected in the sound signal; generating a pulse density stream according to the frequency of the pulses occurring over time in the pulse stream; evaluating the pulse density flow according to a threshold condition; Identifying the speech portion of the sound signal corresponding to the portion of the pulse density flow that meets the threshold condition. The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 1, wherein the processing means is programmed to identify the voice portion of the sound signal based on identifying the portion of the pulse density stream below the threshold condition. The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 1, wherein the processing device is further programmed to receive an original signal and add a noise signal to the original signal to obtain the sound signal . The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 1, wherein the processing means is further programmed to receive an original signal, and to perform band-pass filtering on the original signal to obtain a filtering signal, and adding a noise signal to the filtering signal to obtain the sound signal. The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 4, wherein the frequency of the noise signal is one-half the sampling frequency of the original signal. The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 4, wherein the amplitude of the noise signal is such that the zero-crossing rate of the audio signal is one-half the sampling rate of the silent portion of the original signal. The apparatus for improving voice activity detection with zero-crossing detection as described in claim 4, wherein the passband of the bandpass filtering is between 0.3 and 20,000 hertz (Hz). The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 1, wherein the processing means is programmed to generate the pulse density stream and the threshold condition as a function of the frequency of the pulses occurring in the pulse stream over time Evaluating the pulse density flow is low-pass filtering the pulse flow to obtain a filtered flow and identifying portions of the filtered flow that have amplitudes below the threshold condition. The apparatus for improving voice activity detection with zero-crossing detection as claimed in claim 1, wherein the processing device is further programmed to adjust the threshold condition according to the statistical value of the voice signal. The apparatus for improving voice activity detection with zero-crossing detection as recited in claim 1, wherein the apparatus further comprises a microphone coupled to the processing device, the processing device being further programmed to receive signal from one of the microphones The output gets the sound signal. A method of improving voice activity detection with zero-crossing detection, the method comprising at least the following steps: a processing device receives an audio signal; the processing device generates a pulse stream according to the zero crossings detected in the sound signal; The processing device generates a pulse density flow according to the frequency of the pulses occurring over time in the pulse flow; the processing device evaluates the pulse density flow according to a threshold condition; and The processing device identifies the speech portion of the sound signal corresponding to the portion of the pulse density stream that meets the threshold condition. The method for improving voice activity detection with zero-crossing detection as claimed in claim 11, wherein the method further comprises the processing device identifying a portion of the speech portion of the audio signal based on identifying the portion of the pulse density stream below the threshold condition step. The method for improving voice activity detection by zero-crossing detection as claimed in claim 11, wherein the method further comprises the steps of receiving an original signal, and adding a noise signal to the original signal to obtain the voice signal. The method for improving voice activity detection with zero-crossing detection as recited in claim 11, wherein the method further comprises receiving an original signal, bandpass filtering the original signal to obtain a filtered signal, and adding a noise signal to the original signal The step of filtering the signal to obtain the sound signal. The method for improving voice activity detection with zero-crossing detection as recited in claim 14, wherein the frequency of the noise signal is one-half the sampling frequency of the original signal. The method of improving voice activity detection with zero-crossing detection as recited in claim 14, wherein the amplitude of the noise signal is such that the zero-crossing rate of the audio signal is one-half the sampling rate of the silent portion of the original signal. The method for improving voice activity detection with zero-crossing detection as described in claim 14, wherein the passband of the bandpass filtering is between 0.3 and 20,000 Hz. The method for improving voice activity detection with zero-crossing detection as recited in claim 11, wherein the steps of generating the pulse density stream and evaluating the pulse density stream for the threshold condition further comprise low-pass filtering the pulse stream to obtain a filtered stream and the step of identifying the portion of the filtered stream that has an amplitude below the threshold condition. The method for improving voice activity detection by zero-crossing detection as claimed in claim 11, wherein the method further comprises the step of adjusting the threshold condition according to the statistical value of the voice signal. The method for improving voice activity detection with zero-crossing detection as claimed in claim 11, wherein the method further comprises the step of acquiring the sound signal from an output of a microphone coupled to the processing device.

Images