JPH0490599A - Aural operation type switch - Google Patents

Abstract

PURPOSE: To exactly generate a voice instruction signal when there is a confusing background signal by determining the existence of a voice based on the coincidence between the determination of absence of tone and the determination of presence of pitch. CONSTITUTION: A signal regulator 12 of such a device 10 receives an audio signal through a voice channel 14 and outputs a controlled attenuate signal through an analog/digital converter(ADC) 16 to a first-in first-out(FIFO) buffer 18 for applying delay. The output of the FIFO 18 is inputted to a preprocessor 20 and a variable delay circuit 22, and the output of the circuit 22 is supplied through a digital/analog converter(DAC) 24 and a channel switch 26 to one audio output signal channel 30. When such switch control is used, the audio signal to be switched in voice is generated.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application] The present invention relates to voice-triggered switching, particularly in the presence of highly confusing background signals, in response to the detection of speech information.
1indication signal). A voice-operated switch like this one (
voice operated 5w1tch) is a telephone or radio transmitter (transllllttte
rs) and speech enhancement devices that require the separation of time frames containing speech from time frames containing objectionable audio information in very noisy environments. Useful for controlling

[Prior Art and Problems to be Solved by the Invention] There are various types of conventional voice-operated switches, but they are mainly based on analog signal detection technology.

US Patent Section 4 granted to Poikela
．． 825. () No. 83 describes a two-microphone voice-operated switch (VOX) system that uses autocorrelation of analog signals ( There is a proposal for autocorrelat 1on). This technique is reminiscent of noise-canceling microphone technology, but has no particular relevance to the present invention.

U.S. Patent No. 4,484,344 to Mat et al.
A voice-operated switch based on a llabic rate filter is described. It conditions the input signal using an analog low-pass filter that limits the signal detection operation to 750 kHz or less.

U.S. Patent Section 4 granted to Luhowy
．． No. 187,396 describes an analog speech detection circuit using a syllable ratio filter. This is a hangover tile function that acts as an envelope detector.
) is used.

U.S. Pat. No. 4,052,568 to Jankowski describes a digital voice switch using a digital voice detector and noise detector that operates on a wideband spectrum of voice signals. has been done.

It also has a hangover time function and dual threshold detection.
tecoon).

U.S. Pat. A digital adaptive threshold based on

In contrast to the prior art as described above, the present invention provides a method and apparatus for voice-triggered switching, particularly for generating voice instruction signals in response to detection of voiced information in the presence of highly confusing background signals. This is what we are trying to provide.

[Means to solve the problem]

According to the present invention, in order to achieve the said object, an input audio signal is digitized, low-pass filtered and clipped to obtain a digitized, filtered and clipped signal; Autocorrelate the signal (aut
autocorrela, ting) L, , , autocorrelation (autocorrela, ting)
lation) obtain the function ACF; then: 1) the following investigation steps: determine the peak amplitude of the highest ACF;
determining the periodicity of the ACF peak whose amplitude exceeds a predetermined threshold in each of said plurality of frames;
remembering how many peaks were detected; and based on a weighted addition of a non-linear function of the amplitude of the highest ACF peak and the next highest ACF peak, and the number of detected ACF peaks with the determined periodicity. , the AC of each of the plurality of frames for the presence of a peak indicative of pitch by a step of examining the AC of each of the plurality of frames for the presence of a peak indicative of pitch.
F to obtain a pitch or no pitch determination for each of the multiple frames; and 2) analyzing the ACF of each of the multiple frames to detect tones within the frame. to determine if there is a tone or no tone for that frame (
tone/no tone decision);
Then, it is determined whether there is speech or no speech for the frame (speech/no 5peech decision).
n) and providing a determination that there is speech based on the agreement between the determination that there is no tone and the determination that there is pitch; A method is provided to indicate the presence of audio.

Further according to the invention, digital low-pass filters and clipping means are coupled to filter time-invariant frames of an audio input signal: obtaining an autocorrelation function (ACF) for each of a number of said frames of said audio signal. means coupled to receive the signal processed by the filter and clipping means for processing the autocorrelation function to detect peaks indicative of the presence of pitch within each frame of the audio input signal; a first peak determination processor for determining the amplitude of the highest ACF peak; a second peak determination processor for determining the amplitude of the second highest ACF peak;
a peak determination processor; and; determining the periodicity of ACF peaks having an amplitude exceeding a predetermined threshold in each of the plurality of frames, and determining how many ACF peaks having the determined periodicity are detected. and based on a weighted addition of a non-linear function of the amplitudes of the highest and next highest ACF peaks, and the number of detected ACF peaks with the determined periodicity, pitch is/is pitched. processing means comprising a periodicity detector for providing a determination that the plurality of frames are not present; analyzing the ACF of each of the plurality of frames to detect a tone of each of the plurality of frames;
means for obtaining a tone/absence determination for said frame; autocorrelation function periodicity detection means coupled to process said autocorrelation function to detect the presence of pitch and tone in said audio input signal; ; and coupled to receive a pitched/unpitched determination and a tone/absent tone determination to indicate the presence of voice uttered at the time of the no-tone determination and the pitched determination. An apparatus is provided for indicating the presence of speech in an audio signal consisting of:

[Operating effects and modes of the invention]

In accordance with the present invention, the voice-operated switch uses digital signal processing techniques to identify the voice being uttered and determine whether the selected segment contains primarily speech or noise. ingredient (harmonic)
Detecting a frame of an audio signal having "content" is performed. The method and apparatus are constructed using conventionally manipulable DSP electronic circuitry;
A multiple-stage, del 2yed-decision adaptive digital signal processing algorithm is used. In particular, the method and apparatus provide (1) detecting an input signal at approximately 1 kHz;
(2) A low-pass filter for limiting (2) the presence of periodic components of the input signal below or above a threshold associated with peaks that recognize time-invariant frames or frames containing speech or noise; clipped (c
(3) a nonlinear filtering processor that performs nonlinear smoothing and delay of frame-level decisions and further forward or backward decision extension at the audio segment level; It consists of

The present invention will be better understood by reference to the following detailed description in conjunction with the accompanying drawings.

[Description of specific embodiments]

The present invention is implemented by hardware or software provided within a programmed digital signal processing device. For example, the voice-operated switch can be implemented as an element of other devices using digital signal processing technology. For certain applications, the present invention is manufactured by Texas Instruments, Inc.
Performance has been enhanced by supplemental digital signal processing components such as the TMS320 series devices commercially available from Motorola
It is intended to be implemented in a specialized device manufactured for the periphery of microprocessors such as the Rola 68000. It is contemplated that implementations may be made using other components without departing from the spirit and scope of the invention.

FIG. 1 shows a block diagram of a device 10 controlled by a voice-operated switch (VOX), illustrating the main functions of the voice-operated switch of the present invention. The VOX controlled device 10 has signal conditioning means 12 for receiving an audio signal via an audio channel 14 and providing a controlled attenuation signal to the next stage. This next stage is an analog-to-digital converter (ADC) 16 that converts the analog signal into digital samples. ADC16
The output of is coupled to a first-in, first-out buffer (FIFO) 18 for adding the delay needed for reliable operation in subsequent stages. The output of FIF 018 is coupled to a preprocessor 20 and variable delay circuit 22. The output of variable delay circuit 22 is coupled to a digital-to-analog converter (DAC) 24, which output is coupled to a channel switch 26. channel switch 2
6 outputs are one audio output signal channel 30
supplied to When this voice-operated switch control is used, a voice-switched audio signal is generated. Otherwise, the audio channel simply passes a conditioned audio signal containing speech and noise.

Voice-activated switching is accomplished by processing information retrieved by the briprocessor 20, the output of which is fed to a VOX processor 32. This briprocessor 20 and vOX processor 3
2 can be considered to constitute a voice-operated switch. Two control outputs are provided from the VOX processor 32. namely, the first delay control output 34 and the next audio decision output 36.

Let us now take a closer look at the signal conditioning circuit 12 shown in FIG. The signal conditioning circuit 12 preferably includes an automatic gain control (a
uto II latic gain control:
AGC) device. The AGC may also consist of a series of attenuators whose attenuation rates are mutually controlled based on the estimated energy in the signal interval. This AGC can be more tightly controlled by determining attenuation based only on intervals determined by the VOX processor to contain audio.

ADC 12 may be a conventional linear 12-bit converter with anti-aliasing, or an ADC converter used in digital telephony.
law (A-lay) or MU-law (MU-1aw)
It may also be a codec. sampling·
The rate is 8000 samples per second, which is adequate for audio processing. The DAC 24 is for reconstructing analog signals for subsequent use and is of a complementary format to the ADC 16 format.

PIF018 is a digital delay fine that introduces a delay of approximately 1/4 second (250 milliseconds). Briprocessor 20 conditions the sample signals and sorts them into a sequence of overlapping frames for use by VOX processor 32, as described below. The vOX processor 32, as described later,
Speech/no-5
peech).

Variable delay circuit 22 is provided to compensate for changes in parameters that affect the delay introduced by vOX processor 32.

The channel switch is closed by the VOX processor 32 when passing audio segments and opened when blocking non-audio segments.

The device of FIG. 1 is intended to be illustrative of the present invention, not limiting as to the particular features of the invention, and is illustrative of one embodiment of a device that may be considered a voice-operated switch. It is something. The actual switching decisions are made in an element shown as OX processor 32.

FIG. 2 shows a block diagram of the preprocessor 20 of the present invention. Briprocessor 20 produces digitized input signals that are processed in VOX processor 32. According to the invention, vOX processor 32
When speech is present in the audio signal, it makes a preliminary decision based on pitch information in a constant vocalized speech segment with a duration of 16 ms, and continuously over a period extended forward and backward. The inadequacy of this determination is compensated for by giving a gender, and the unpronounced sounds before and after the sound are compensated for.

The preprocessor 20 consists of a low pass filter 38, a down sampler 40, a center clipper 42, and a frame segmenter 44. Low pass filter 38 receives the digital signals from selected stages of FIFO 18 and passes the filtered digital signals to down sampler 40 . Down sampler 40 is coupled to frame segmenter 44. The output of frame segmenter 44 is coupled to the input of center clipper 42. The output of the center clipper 42 is as explained later.
It is coupled to an input of VOX processor 32.

The low-pass filter 38 is a digital filter with a cut-off frequency (cutof'f' frequency) of 1000 Hz or less, preferably 800 Hz.

This is to improve the useful pitch S/N ratio characteristics in the spectrum from 50Hz to 500Hz,
It is known that most of the pitch frequencies of phonemes pronounced in real-time normal conversation are included in this range of 50 to 500 Hz.

Down sampler 40 is a mechanism for reducing the filtered signal. A resolution of 8000 samples per second is no longer necessary. I mean,
This is because the effective bandwidth is only 800H2. The down sampler 40 thus functions to discard, say, three out of four samples each time, while retaining sufficient information to make the necessary decisions about the remaining bandwidth of the signal. This also reduces the complexity of signal processing. (However, signals that have been filtered but not discarded can be processed by a selected adaptive process such as autocorrelation.)
pre-processing). ) Frame segmenter 44 performs a segmentation process to divide the stream of digital audio samples into useful processing frames. In particular, the digital audio samples are grouped in frame segmenter 44 into frames with preferably 509o overlap between successive intervals. In the presently preferred embodiment, the frame length is selected to be 256 samples or 32 milliseconds. Deciding on the frame level is
Generated every 16 milliseconds. Because of the overlap, transitions between spoken audio segments are handled more smoothly, and second-level decisions are twice as effective as frame-level decisions.

The center clipper 42 is for flattening the spectrum (spectrum + flattener).
), the vocal tract transfer function (vocal tract tra
nsfer function) and convert each harmonic to the fundamental wave.
It is suppressed so that the amplitude is approximately the same as (al). In the specific procedure, we looked at the peak amplitude in the first third and last third of the segment (i.e., a 32 ms audio segment) and set the clipping level to these two measured maxima. Set to a predetermined percentage of the smaller of the values. The clipping level input 43, which is a parameter supplied by the VOX processor 32, is preferably about 0.65 times the smaller maximum value. For a detailed explanation of this center-clipping technique, please refer to Rabiner (L, R, Rab
iner) and Sheikh 7- (17J, 5chafer
``Digital processing of audio signals (Digital
Processing of 5peech Sig
nals) J (076B 2New Jersey Ingredients Cliffs Prentice Hall Co. Pren.
tice-Hall, Inc., Englewood C.
l1ff's, NJO7632) 15 of 1978
Given on pages 0-154.

To understand the need for a center clipper, it is helpful to review the classical model of speech production. Speech production can be thought of as the excitation of the vocal cords, which produce both vocal vibrations and unspoken "white noise-like" sounds.When the vocal cords vibrate at the pitch frequency, frequency selection occurs. generates a train of impulses at the pitch frequency represented by the vocal tract transfer function that produces an attenuation of The peaks and valleys of this spectrum are shown.The peaks of this spectrum
frequencies) J, which corresponds to the resonant frequency of the vocal tract.

In accordance with the present invention, VOX processor 32 utilizes the presence of pitch in the voiced speech to determine whether there is speech in the audio signal. but,
If this excitation or pitch is to be enhanced to improve its detectability, it is believed that it is desirable and necessary to remove formant spectral structure from the speech spectrum prior to detection. The particular type of V used here.

In the The correlation peak is higher than the autocorrelation peak due to the periodicity of the vocal excitation, resulting in false readings. This is especially true if this reading is based on the selection of the highest peak within the segment. In order to minimize this problem, it is desirable to process the audio signal so that its periodicity is more noticeable by suppressing peaks caused by other factors. Therefore, in the present invention, a centered chestnut sonomer spectral flattening technique is used as explained above.

FIG. 3 shows a block diagram of the VOX processor 32 of the present invention. vOX processor 32 is best described by the corresponding software algorithms used in the present invention.

The vOX algorithm uses a first level determining means 50, a second level determining means 52, and a third level determining means 54. The first rubel determining means 50 operates on overlapping frames of the signal and determines whether the frame is of the first category of vocalized speech or of the second category of unspoken speech, i.e. noise or silence. , evaluate whether it is. The algorithm of the first rubel determines whether the input frame is (1) uttered voice V or tone T, or (2) unvoiced 1X voice U or noise N or silence S. Using the pitch as an index of the first element 56 of the second level determining means 52
as a binary decision. No. 1 rubel determining means 50
also extracts the pitch information P and the extracted tone T
is applied to a delayed tone detector element 58 of the second level determining means 52. The first element 56 that receives the VT/UNS decision is an intermediate smoother 56, a nonlinear filter used to smooth the decision and pass decisions that exhibit sharp, consistent transitions. Delayed decision tone detector 58 is a detector that detects the presence of constant frequency tones having a duration of several frames or more in the range from 50 Hz to 500 Hz. Intermediate smoother 5
6 and the output of the tone detector 58 of the delayed decision is coupled to a decision combiner 60 where the tone output decision T of the tone detector 58 coincides with the voice/tone output decision VT of the intermediate smoother 56. A decision is made to block the audio decision.

The third level determining means 54 operates over several frames. All second level decisions are therefore stored in a decision storage means.
62 to provide the necessary delay for the third level decision. This decision storage means includes a decision extender/modifier that provides a final speech/non-speech decision for each different frame.
der/a+odifier) 64. This decision extender/modifier 64 removes extremely short audio segments that indicate false audio detections and does this when an unspoken audio segment is adjacent to a vocal audio segment. It is intended to extend second-level decisions to be included in this decision, to fill short gaps of silence, and to provide a hang-time delay. ing. Synchronizer 66 is used to ensure that equal delays are provided between FIFO 18 and VOX processor 32. This synchronizer 66 controls the variable delay circuit 22.

FIG. 4 shows a detailed block diagram of the rubel determining means 50 of the present invention. This rubel determining means 50 is
Autocorrelator (autocorrelator') (ACF) 68,
ACF normalizer 70 , positive peak detector 72 , audio signal detector 74
, a first peak determination processor 76 , a second peak determination processor 78 , a periodicity detector 80 , a periodicity function processor 81 , selected weighting functions 82 , 84 , and 86
, multipliers 88, 90 and 92, an adder for integrating the first peak determination processor 76, the second peak determination processor 78, and the periodicity function processor 81.
94, a comparator 96, and a deterministic combiner 98.

Autocorrelator-6 shown in this preferred embodiment
8 is 32 samples reduced from 256 samples to 64 samples from the frame segmenter of preprocessor 20.
Taking overlapping frames of millisecond length, we calculate the unnormalized autocorrelation function between the maximum and minimum lags, and the resulting autocorrelation function ACF (k), k
- the minimum value, and the maximum value are given to the ACF standardizer 70 and the audio signal detector 74; Here, the preferred minimum lag corresponding to a high pitch of 500Hz is 4, and 50
The preferred maximum lag corresponding to a low pitch of Hz is 40. Lag, ACF at zero (ACF (0))
is known as frame energy.

The audio signal detector 74 uses the minimum energy level (4-5 bits of a 12-bit signal) as a parameter input to determine the frame energy (ACF (0)).
Detect situations where there is no audio signal. A signal indicating the presence or absence of an audio signal is provided to a decision combiner 98. This is the only stage in the decision process where the signal level is the basis for the decision.

ACF normalizer 70 receives the output signal of autocorrelator 68 and normalizes the energy and envelope. Normalization of the energy is performed by dividing the output of the normalization function from k = min lag to -max lag by the frame energy ACF (0).

The normalization of the envelope is done by multiplying the ACF by the inverted triangle factor, thus giving the roll-off (r) of the triangle envelope of the ACF.

for r) characteristic is replaced by a rectangular envelope for the ACF.

The positive peak detector 72 detects a preselected number of peaks that exceed a standardized threshold, and more accurately calculates the value of the ACF and the lag of each peak. Preferred values for the normalized threshold range from 0.1 to 0.2. The output, in the form of a list of peaks with ACF values and lags, is provided to a first peak determination processor 76, a second peak determination processor 78, and a periodicity detector 80.

A first peak decision processor 76 receives as input the maximum value of the ACF peak and outputs a positive decision if this value exceeds a threshold PIMAX-T indicating the presence of a pitch in the preselected signal. . A non-linear function is applied to reflect the possible presence of pitch at different levels of PIMAX. Typical values for PIMAX-T are:
It ranges from 0.4 to 0.6, and as this value decreases, the probability of detecting audio and false alarms increases.

A second peak determination processor 78 receives as its input the second highest ACF peak and has a value between 0.35 and 0.5.
5, i.e. the threshold for the second ACF peak, except that P2MAXT is used as the threshold.
It is a non-linear function equal to the first peak determination processor 76.

Periodicity detector 80 confirms the periodicity of the ACF peak. For a pronounced frame, the lag of the ACF peaks forms an arithmetic sequence with the first element, zero, and the difference between successive elements corresponding to the pitch period.

The allowed lag compensates for the difference between the ideal sequence and the detected sequence. The periodicity detector 80 has as an output:
Give the following values = (1) the theoretical number of peaks (TNPKS) calculated by dividing the maximum lag by the delay of the first peak; (2) the number of actual peaks forming an approximate arithmetic sequence; number (below the peak when there is no lag) (ANPKS) (3) Pitch period evaluation or sequence difference.

This evaluation of the pitch period is performed by a pitch consistency detector (pitch consistency detector) of the second level determining means 52.
ydetector) ()-n detector), and other values are provided to the periodicity determination processor 81.

A periodicity determination processor 81 receives the output parameters and assigns a value to each combination from a look-up table indicating the probability that the received signal is periodic. Since these values are primarily empirically corrected for periodicity detector 80, no specific algorithm is provided in this preferred embodiment.

The output of each decision processor 76.78.81 is a non-deterministic decision indicating the probability that an uttered segment or tone (pitch) has been detected.
on). The fitness of the resulting decision (f 1
In order to increase exjbi I 1ty, a weighting factor (weighting
coefricient) 82, 84, 86 are provided, which are multipliers 88, 90, 92, respectively.
Weight the values of non-deterministic decisions by multiplying by . The respective outputs are summed by an adder 94 and provided to a comparator 96 which is set to a threshold value, preferably zero.

A positive result thus indicates the presence of pitch in the signal.

The final step in determining the second level is the decision combiner 98. This determines the pitch and determines whether there is an audio signal in the signal detector 74 (audio/no au old □
If there is no audio signal, then even if there is a full output of the adder 94, the output of the third level determining means 54 will be DN
S (no voice or tone). However, if the pitch evaluation is not successful, the VT/UNS determination is also sent to the second level determination processor 52.

Referring again to FIG. 3, the major elements of second level determining means 52 are shown. Intermediate smoother 56 looks at the given odd number of previous rubel decisions and determines which of the two states is more common. It shows as its output the pre-given odd or multiple state of the second level decision. Therefore, it operates to eliminate short duration transitions induced by noise. Intermediate smoothers of this type include Rabiner (L, R, Rabiner) and Schafer (R).
, W., 5chafer), ``Digital Processing of Audio Signals J, 1978 (cited above), pp. 158-161.

The pitch estimate is fed to a tone detector 58, more precisely to a pitch-dispersion detector 58, which has the consistency tolerance and the window width as parameter inputs. If this pitch estimate is within a consistency tolerance for a duration greater than a predetermined minimum tone duration, a decision T that a tone is present is issued to the decision combiner.

A decision combiner 60 of the second level decision means 52 combines the smoothed output of the intermediate smoother 56 and the tone determination T of the tone detector to determine whether the signal is voiced or unvoiced. , a signal indicating noise or silence (DNS) is generated by suppressing frames containing tones. This determination of ■/UNS is provided to the decision storage means 62 of the third level determination means 54, where the determination of the audio segment level is made.

FIG. 5 shows a decision storage means 62, a decision expansion/modifier 64
A portion of the third level determining means 54 having the following is shown.

As explained earlier, the decisions of all frames are caught and stored in the decision storage means 62 for a period of -periods.

Some audio segment level decision processes are:
This is done on accumulated data.

The first short speech segment tester 100 is provided to delete or modify all V-segments whose duration is less than a preselected minimum kV for UNS determination.

initial backward dilator
extensi.

nHO2 and final back dilator
ardextension) 104 is provided to test the backward extension in time of every voice decision V. The purpose is for a segment of audio to include a segment without audio to which it should be passed as a determination of its associated audio. A typical extension is 5 to 10 frames. (The sum of the initial backward expansion time and the final backward expansion time has a direct impact on the time delay, so care must be taken not to increase the time if you wish to shorten VoX hangs.) An initial forward dilator 106 and a final forward dilator 108 are provided to test the temporal forward expansion of all voice segments V. The purpose is to include a segment of speech with a following segment of speech that should be passed as a determination of its associated speech, and to provide a limited amount of hang between words and sentences. It is. A typical value for the initial forward expansion parameter is 5 frames. (Forward extension has no direct effect on VOX time delay.) Short Silence Interval.
The tester 110 is provided to convert silence intervals shorter than a preselected length kS into a determination V that there is speech.

The final rear dilator 104 is typically set from zero to 15 frames. Parameters are selected based on the total time delay allowed.

The final forward dilator 108 is set to a minimum of 10 frames to include the non-speech frames that follow the detected speech. This maximum value is limited only by available memory.

Values from 500 milliseconds to 3 seconds are considered sufficient for anticipated applications.

To aid in understanding the invention, flowcharts of specific embodiments are shown in FIGS. 6A-6D. A step-by-step explanation of this process is provided below.

201.202, 203, 204: This is the sampling and low-pass filtering block.

205, 206, 207: When another 128 samples are accumulated, a new 256 50% overlapping frame is formed.

208 209. 210, 211 + central clipping 212: Compute the autocorrelation function (DACF) for 64 reduced samples per frame. Rug (la)
g) varies from [minimum lag] to [maximum lag].

213: DACF with lag or 0 is frame energy.

214: Compare the frame energy with a threshold to make an audio/no audio decision.

215+DACF is normalized by frame energy.

216.217: As the lag increases, fewer terms are included in the sum, so A and CF have triangular envelopes. Therefore, the DACF is divided by this envelope. The resulting function 1; i NDA CF color specified. NDACF (0)-1゜218:l:"- is a point that is greater than two adjacent points. Only peaks that exceed the threshold are considered, and only the first n peaks that exceed the threshold be considered.

219-227: This is a loop. This purpose is 2
18 to find the exact lag and value of all detected peaks. Fit the vertical parabola to the three highest peaks. Gives an approximation to the lag of the axis of symmetry or the lag of the peak of the parabola.

and using all 256 samples of the frame
CF is calculated again to obtain a more accurate value. This ACF
is designated UDACF and is calculated around the approximated position of the peak. Therefore, UDACF is standardized in the same way that DACF was standardized, and becomes NUDACF.

Finally, the exact location and value of each peak is memorized.

228: PIMAX is the highest NUDACF peak value and P2MAX is the next highest NUDACF peak value.

229: The purpose of this block is to measure how close the positions of some peaks are to an arithmetic sequence with zero being the first element of the sequence.

The procedure is as follows.

The first peak, i.e. the one with the minimum lag, is selected.

This lag can now be considered as the difference in the arithmetic sequence, D
is specified.

Is there a peak in lag 2*D± [periodicity tolerance]? If No, the longest sequence corresponding to the first peak is 1.

If YES, is there another peak at lag 3*D (within tolerance)? If NO, the longest sequence corresponding to the first peak is 2.

If YES, are there other peaks in 4*D?
This process is continued until the length of the longest sequence corresponding to the first peak is determined.

Select a second peak and repeat the above procedure to determine the length of the longest sequence corresponding to the second peak.

Similarly, determine the length of the longest sequence corresponding to the following peak.

Remember the length of the longest sequence.

230. Pitch is defined as the difference between the longest of all longest sequences.

[Periodicity Index] is the length of the longest sequence among all the longest sequences corresponding to any peak.

232: The theoretical number of peaks [T N P O) is
This is the number of peaks that should exist in the case of a completely periodic waveform with a fundamental frequency of frame [pitch], and ACF is calculated by the maximum lag of [maximum lag].

233.234,235: PIMAX, P2MAX, T
The table is examined using the NPO and the periodicity index. The table of periodic soft decision functions is two-dimensional.

236: Weights are parameters of the algorithm. These values are optimized for a particular application.

237/239: From now on, each frame is represented by a bit, where z; tv, T, o means u, N or 5t-i. K
There is an infinite string of such bits called VOX.

This is provided as a cyclic buffer in the software.

240.241.242: This block performs an intermediate smoothing function. The output of the intermediate smoother is the new bit string LVO
Stored in X. The reason for storing processed decisions in a new bit string rather than in the same column is simply to save the unprocessed decisions for the next intermediate smoothing window.

This block is a tone detector.

244.245, 246.247: This block removes speech segments that are too short. Transition from 1 to 0 or L
When a VOX (i.e. intermediate smoothed decision) is detected (244), the previous (KV) bits or 0 to 1
is scanned for transitions to (245) and if there is one, the utterance segment is too short and is therefore deleted by modifying the 1's to 0's during the transition (247).

If there is no such transition, the utterance segment is long enough so that it is extended forward [IFE)+CFFEI frames after being copied into C246) bit stream M V OX.

248 249: This block is exactly (KV) “
Detect the case where 1" adjacent bits are accumulated and extend "1" backward by the initial backward extension [IBE].

250.251: This block removes silent segments that are too short.

252: This block performs final backward extension (FBE). MVOX is 70R for N V OX l:
Note that VOX is computed by blocks of deleted audio or deleted silence. This procedure ensures that the final expansion is not erased.

261, the final decision is output based on the bit string NVOX.

The present invention has been described above with reference to specific examples. Other embodiments, such as those implemented by hardware or in other programmed or software forms, will be apparent to those skilled in the art. Accordingly, it is not intended that the invention be limited, except as in the scope of the appended claims.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a device using voice-operated switching means of the present invention, FIG. 2 is a block diagram of a preprocessor of the present invention, FIG. 3 is a block diagram of a vOX processor of the present invention, and FIG. The figure is a detailed block diagram of the third level determining means of the present invention, FIG. 5 is a block diagram of the third level determining means of the present invention, and FIGS. 6(A) to (D) are part of the process of the present invention. It is a flow chart showing an example.

Claims (13)

[Claims] (1) Digitize, low-pass filter, and clip the input audio signal to obtain a digitized, filtered, and clipped signal; then autocorrelate the clipped signal to produce a large number of frames. Obtain the autocorrelation function ACF for each of the multiple frames; then: 1) determine the amplitude of the highest ACF peak; determine the amplitude of the next highest ACF peak; and then perform the following investigation steps: determine the periodicity of the ACF peaks whose amplitude exceeds a predetermined threshold, memorize how many ACF peaks with the determined periodicity are detected; and identify the highest ACF peak and the next highest
by a probing step consisting of a weighted addition of a non-linear function of the amplitudes of the F peaks and a step of providing a pitch or no pitch determination based on the number of detected ACF peaks with the determined periodicity; , the AC of each of the plurality of frames for the presence of a peak indicative of pitch.
F to obtain a pitch or no pitch determination for each of the multiple frames; and 2) analyzing the ACF of each of the multiple frames to detect tones within the frame. and obtain a tone or no tone determination for the frame; and make a voice or no voice determination for the frame, and determine a toneless determination and a pitch determination. A method for indicating the presence of speech in an audio signal in each of a number of time-invariant frames, comprising the steps of: making a determination of the presence of speech based on a match of the signals; 2. The method of claim 1, further comprising the step of: (2) after said digitizing step, dividing said frames into overlapping parts. (3) said clipping step is adapted in relation to the lower of the two highest peaks of the filtered signal occurring in the first third of frames and the second third of frames; 2. The method of claim 1, including setting a clipping level to . (4) The method of claim 1, wherein the autocorrelation step includes normalizing the autocorrelation function. (5) The investigation step includes the highest ACF peak and 2
a preliminary quantitative value corresponding to a first possibility of pitch determination, comprising an act of comparing said ACF peaks including a second highest ACF peak, said highest ACF peak to a first threshold; and said second highest AC
A second step in determining pitch by comparing the F peak with a second threshold.
5. The method according to claim 4, wherein preliminary quantitative values are obtained corresponding to the probability of . (6) The analysis step includes AC
6. The method of claim 5, comprising checking the consistency of lags between F-peaks and deriving an estimate of the pitch of the ACF. 7. The method of claim 6, wherein the step of analyzing further comprises detecting matched tones across multiple frames for application to the step of making a determination. (8) Prior to the step of making said decision, there is a preliminary pitch/no pitch over a number of frames to suppress excessive transitions between the decision that there is a pitch and the decision that there is no pitch. 2. The method of claim 1, further comprising the step of smoothing the determination. (9) storing a large number of voice/no voice determinations to perform a sufficient number of accumulations to generate voice segment level decisions, and utterances before and after the voiced voice; 2. The method of claim 1, further comprising the step of generating audio segment level determinations of sufficient duration to include unused audio. (10) digital low-pass filter and clipping means coupled to filter time-invariant frames of the audio input signal; means coupled to receive the signal processed by the filter and clipping means; means coupled to process the autocorrelation function to detect peaks indicative of the presence of pitch within each frame of the audio input signal; a first peak determination processor for determining the amplitude of the highest ACF peak; a second peak determination processor for determining the amplitude of the second highest ACF peak;
a peak determination processor; and; determining the periodicity of an ACF peak having an amplitude exceeding a predetermined threshold in each of the plurality of frames, and determining how many ACF peaks having the determined periodicity are detected. and based on a weighted addition of a non-linear function of the amplitudes of the highest and next highest ACF peaks, and the number of detected ACF peaks with the determined periodicity, pitch is/is pitched. processing means comprising a periodicity detector for providing a determination that there is no tone; analyzing the ACF of each of the plurality of frames to detect the tone of each of the plurality of frames; means for obtaining a determination; autocorrelation function periodicity detection means coupled to process the autocorrelation function to detect the presence of pitch and tone in the audio input signal; and a determination that there is no tone and pitch. decision combining means coupled to receive a pitch/no pitch determination and a tone/no tone determination to indicate the presence of uttered speech when the decision is made; A device that indicates the presence of audio. (11) further comprising audio segment level determining means responsive to an output of said determining and combining means indicating the presence of uttered audio in a given frame, said audio segment level determining means being configured to select a sufficient number of frames; means for capturing and processing to produce speech segment level determinations, initial backward extension means, initial forward extension means, final backward extension means, final forward extension means, short speech segment testing means;
and short silence interval testing means, said extending means and testing means extending the time base of said speech segment level determining means to include unspoken speech and gaps between words. (12) Claim 1 further comprising means for synchronizing said audio segment level determining means with a corresponding audio segment.
1. The device according to 1. (13) The apparatus according to claim 10, further comprising means for dividing the frame into frames that temporally overlap with each other.