JPS6039938A - Complementary speech detector - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to digital communication systems, and more particularly to the detection of speech signals in such systems.

In digital communication systems, analog speech signals are pulse code modulated (PCM), for example, at a sample rate of 8 KHz with a sample interval of 125 μs. In the United States and Canada, samples are digitized serially to 8 bits according to the LT method PCM code. C.C.I.
Similar A-law codes are followed in European countries that follow the T 'I' standard. A 4-bit code has been proposed. These codes are y-logarithmic in order to minimize encoding speed while optimizing signal-to-noise ratio over the normal range of signal levels. The first bit of each sample is a symbol pit indicating the polarity of the sample.

Successive pits in the sample are level bits that represent the level of the sample.

It is known to use so-called "speech detectors" or "silence detectors" to distinguish between speech and periods of silence.

It is possible to transmit data or other speech on the same channel during such silent periods, which allows the channel capacity to be utilized more fully.

I) CM digitizes both speech and background noise (eg, noise generated in the same room as the speaker). In addition to background noise, the communication system itself generates system noise, and this noise
It is transmitted on the same channel as the M signal. In the absence of noise, speech and silence can be distinguished by a simple P-I level detector. However, in reality, low-level speech can have the same magnitude as high-level background and system noise, requiring the use of more sophisticated means to distinguish speech from noise. .

Summary of Defense In one aspect of the invention, a flag signal is provided when a speech signal is detected in a waveform carried by a communication channel. The energy level and zero crossing speed of the waveform are calculated, and this energy level + i high: to
, J: is compared to the low energy threshold. If the energy level is equal to or greater than the high energy threshold, a flag signal is provided. On the other hand, if the energy level is below the low energy threshold, the flag signal is inhibited. If the energy level is above or above the low energy threshold, then the difference between 114'i of the waveform plus the identically separated energy levels is calculated. This level difference represents the 1.4 slope of the energy of the waveform. This energy slope is then compared to a slope threshold and a flag signal is generated if the slope is equal to or greater than the slope threshold. If the slope is below this threshold, the zero crossing speed is compared to the zero crossing threshold. A flag signal is generated if the zero-crossing velocity is greater than or equal to the zero-crossing threshold. If the zero-crossing speed is less than the zero-crossing threshold, the flag signal is inhibited.

DESCRIPTION OF EMBODIMENTS In order to fully understand the invention together with other advantages and possibilities, embodiments of the invention will now be described in detail with reference to the accompanying drawings, in which: FIG.

As a feature of the invention, three parameters are collected from the relevant channels: energy level, short-term change in energy level or short-term change in energy envelope slope, and zero-crossing velocity. , used in a complementary manner to detect speech from hacking ground and system noise.

These parameters are collected from multiple PCN4 samples within one time slot. The energy level of each sample is a function of the linear value of that sample squared. However, the size of the PCM nonlinearity sample is
For speech detection purposes, it is a satisfactory approximation of the energy level.

According to the invention, the magnitude of each sample is used to calculate the energy level and slope. This can avoid converting non-linear codes to linear codes.

The differences between time and approximate energy levels are minimal and are not distinguished in the description and claims of the invention.

The energy level E (m) is obtained by adding the magnitudes of all 64 samples within an 8 mS time slot and dividing this phase by the number of samples. E(m) is
Updated by a sliding window with a width of 2 mS, corresponding to 16 samples in one introduction. Parameter updates are performed by adding the latest 16 samples of the collected samples and removing the oldest 16 samples.

E (m) is based on odBmo, which is a standard energy level used in the four branches of communication. High and low energy thresholds THn and 1'H
L is used. If E(m) is i'llH(1'ux
(=-30d13mo) or greater than this, the signal is determined to be speech. On the other hand, if E (nl) is THL (THL = -s o
dBmo), the signal is always determined to be silent. With an energy difference of 20 dB between Tl and 1'TIL, it is difficult to distinguish between speech and noise, and the parameters of energy slope (BS) and zero crossing rate (ZCR) become effective. Speech in this range is low level, typically nasal or phricate or other low level speech phonemes at the onset or decay of the speech utterance.

A steep energy slope (E
S) has been found to accurately represent speech. The slopes are two temporally separated (e.g. 2 ms
distant) time slots, e.g. trn and t(m-32
) is the difference between the ltMs energy levels ■ and therefore E S = E (m) − E (m −32). Since the μ law code and the A law code are y-logarithmic at low signal levels, the difference ES is also y-logarithmic. Low slopes are ambiguous with respect to speech or silence.

If a slope of less than 6 dB is measured, the channel zero-crossing rate (z crt ) is then used to distinguish speech from noise.

As already mentioned, noise can be classified into either background noise or system noise. Bank ground noise is generated at the speaker's location and is usually of low frequency. System noise occurs within the channels of a communication system. In contrast to background noise, system noise tends to be more randomly distributed across the band [1] of the channel and is usually of low level.

Low-level fricative and nasal speech tends to contain higher frequency components than background noise. A good 4Fi indication of the spectral characteristics of a signal can be obtained by the number of zero crossings within a fixed time period (ie, the zero crossing rate). A change in polarity of the zero bit of successive PCM samples represents a zero crossing. Under normal, low system noise conditions, a zero crossing rate (ZCR) above one threshold (eg, 50 zero crossings per 16 ms) is indicative of speech. Under high system noise conditions, high zero crossing rates represent either speech or system noise. To distinguish speech from normal or high level system noise, the present invention employs a technique referred to as Modified Zero Crossing Rate (MZCR). There are two aspects to the MZCR method. First, a determination is made whether high level system noise is present in the channel. If present, the detector is switched from the normal "Z CR mode" to 1MZcR mode-1.

Determination of high level system noise is done by measuring the peak ZCR, during the silent period displayed by the detector. Since background noise and speech are not enriched during silent periods, a sufficient number of zero crossings can be attributed to high level system noise. The clock generates pulses at 0.25 second intervals. If a peak ZCI reading of less than 25 zero crossings per 8 mS is obtained for 0.25 seconds, the counter is decremented by one.

If the peak ZCR does not reach 25 zero crossings per 13 m5 in 0.25 seconds, the adaptive counter increases its count by 1.

If the adaptive counter reaches -4, it is assumed that there is a high level of system noise and the speech detector is switched to its M Z CIt mode.

If the adaptive counter reaches 3, only low level system noise is present and the speech detector is switched to normal ZCR mode. The mode changes only upon completion of the 025 second count. If the counter does not reach any level, the detector mode does not change and remains in whatever mode it was in for the previous 025 seconds.

Speech or background noise is Z Cit
To avoid influencing mode selection, the counter is reset to zero at the beginning of any period in which the energy level exceeds -60 dBmo or the energy slope is 6 dB or greater, and remains in the reset state during any of these conditions. to hold.

M Z Cl(
, mode, speech detection is somewhat more difficult than in ZCrt mode. However, the frequency spectrum of the system noise has been found to be more random than the more dense 7T fricative speech.

In MZCR mode, the objective is to discard zero crossings that contribute to system noise. Therefore, a dual threshold technique is used.

It has been found that adjacent bit reversal sequences due to random system noise are shorter than bit reversal sequences due to narrowband speech. The bit reversal sequence is PCM
is the number of consecutive bit flips in the sample. During the MZCR mode period, Bjord TH+ (for example, TH1= q
(continuous sign bit reversal). Burn sequences with lengths shorter than TH+ are considered to be primarily due to high level system noise and are rejected. 'I'
Sequences as long as H+ or longer are accumulated. The number of accumulated first sequences is threshold l'
)12 (e.g. TH2-12 length sequence/16m5)
If it is equal to or greater than K, it is considered speech.

MZCR has more stringent requirements for speech than ZCR. This is necessary to reduce the possibility that high level system noise will sometimes override the speech detector. Therefore M7. CR is normal Z CR
There is a possibility that the silent period is incorrectly expressed as i/I'. For this reason, in MZ CIt7, the hangover is
This increase in hangover adds some overhead, but premature clipping of speech. It is suitable for implementation. This program accumulates data on a sample-by-sample basis and sets parameters on a window (
It must be calculated and updated on a per-window basis. This program is temporally divided into two parts, referred to in the present invention as "execution of interrupt handler" and "main program".

[The first part of the interrupt handler's execution is initiated by an interrupt pulse. The interrupt pulse is synchronized to the sampling rate.

For a sampling rate of 8 K 1-1z the interval between the interrupt pulses is 125 m5.

The [Interrupt Handler Execution] section is approximately 86m5 long when executed on an Intel 8085A microprocessor. During this period speech data samples are accumulated and energy levels are accumulated and Z CIt- or IVI
Z Cl( is checked. "Execute interrupt handler"
After the completion of the section, the "main program" is restarted, and execution of the exemplary program is continued again until about 39 m5 after the next interrupt pulse occurs.

The main program executes functions on a 2 ms window-by-window basis. Energy level, ZC during the main program period
R and energy slope parameters are 6 new 2mS
Collected and updated for each window. Speech detection logic is executed during this main program. The system noise is measured and its detection is situationally z c
It is applicable to it mode or MZCR mode. The main program may be separated into several executable code segments that are not interrupt-based.

The initial settings of the program include printer structure settings, interrupt permissions, and input/output initial settings. The waiting loop is a window synchronization mechanism that precisely adjusts the execution timing between the main program and the interrupt handler.

As mentioned above, the invention can be implemented using a programmable microprocessor. Or a dedicated (
(digitized) logic circuits can be used.

FIG. 2 is a block diagram of logic circuitry suitable for implementing speech detection logic. The serial data stream of the communication channel is coupled to a conversion circuit 42 which converts it into a series of parallel 8-bit bytes in a known manner. The input circuits 10 have E (m) and E S (m), respectively.

Parameter update circuits 11, 12.1 for Z CIt (m) and M Z CR (m)! 1 and 14.

1) Input circuit 100 update circuit 11 for CM encoder
．． Timing markers are provided to synchronize 12, 13, and 14. E (m) circuit 11 and ES (m
) For circuit 12, the first bit of each byte is ignored;
Only the bits representing the magnitude are processed by circuit 11.12.

E (m) The purpose of the update circuit 11 is the time period of the SMS (
64 samples) every 2 ms to obtain the average energy E (m). The bit representing the magnitude of 16 consecutive samples is 2 MmW. This sum is 2
Corresponding to the energy aggregated during ms, this and h are temporarily stored. The four most recent 2 mS sums are added and the resulting sum is divided by 64. This quotient is E
(m) Output of circuit 11, representing the average energy of 64 samples updated to 2m54σ, which is the purpose of this run.

The purpose of the B S (m) update circuit 120 is to determine the slope of the energy envelope ES(m).

This circuit may use the 8 ms average energy value already obtained. The circuit 12 is pre-constructed, e.g. 64m5
The latest 8 rns average energy value obtained previously. Subtract the ms average energy value. E S (m)
The output of circuit 12 represents the absolute value of the difference and represents the energy slope updated every 2 mS, which is the purpose of this implementation.

The purpose of the Z CR(m) circuit 16 is to determine the zero crossing speed (ZCl
() is to be determined. Only the %ness bit of each byte is considered. 12 of 8 consecutive 2ms time periods
The polarity bit of 8 is temporarily set. circuit 13
The output of is the gold side number of polarity bit reversals for the eight most recent 2ms time periods, updated every 21'TIS 16m
The zero-crossing speed i1i (Z
CR).

The purpose of M Z CIt (m) update circuit 140 is to provide a modified zero-crossing speed. Circuit 14 also uses stored polarity bits, but instead of counting all polarity bit reversals, only those polarity reversals that are part of a sequence are counted. Each polarity bit in a sequence is different from the previous polarity bit. The number of polarity bit reversals in each sequence is compared to a threshold 'r H+. For example, sequences with fewer than 9 reversals within a 16m5 period should be rejected.

The number of sequences with 9 or more polarity reversals within a 16m5 period is added. The sum of this cumulative sequence represents the modified zero-crossing velocity.

Digital signals representing each oil change parameter from circuits 11, 12, 13 and 14 are fed to a comparator where they are compared to a corresponding threshold.

E(m) change The signal from circuit 11 that turns off E(m) is supplied to comparators 15 and 18. If E(m)
If (m) is equal to or greater than the high threshold 'I' I-I H, comparator 15 provides an output. This output causes OR gate 16 to provide a signal that turns on flag switch 17, which indicates the presence of speech. If E (m) is less than the lower threshold ′I″It, then comparator 1
8 supplies a signal that turns off the flag switch 17 after being delayed by the slow recovery timer 2o to inhibit the flag signal, thereby indicating the absence of speech. THH
For example, -50d13m.

'I'HL may be -50 dHmo. The flag switch 17 may be a flip-flop circuit. The slow recovery timer may be an up/down counter.

If E (m) is the threshold 'l''Hri and 'l'
If it is between Hh, the comparator 18 is A N
I) Give a signal to gate 21. AND gate 21 generates an output signal only when it receives signals from both comparators 18 and 01 (gate 220).
2 is a predetermined ES (m) as described later.

A signal is supplied in response to the Z CH, (+n) or MZCR (m) level. The output of AND gate 21 is applied to the input of OR gate 16 which turns on flag switch 17 in response to a signal.

The output of the ES (m) update circuit 12 is given to a comparator 23 . If E S (m) is E S (m) threshold (E S (m) l' H= 6 dB
), the comparator 25 provides a signal to the OR gate 22 that turns on the flag switch 17.

If E S (m) is lower than its corresponding threshold, comparator 23 provides a signal to one of the AND gate 240 inputs. AND gate 24 is turned ON only when the other input receives a signal from OR gate 25.
A signal is supplied to the R gate 19. This (JR Gate 25
provides a signal only if the M2', CR or ZCR requirements as dictated by the signal from comparator 68 or 69 are met. 0 if there is no signal from either AND gate 24 or comparator 18.
11L gate 19 does not provide a signal and flag switch 1
7 is not treated as off. If the signal from gate 24 is not present, gate 19 turns off flag switch 17 in response to the signal from comparator 18.

E (m) lies between the two thresholds, and B
If S (m) is lower than that threshold, then
The logic circuit is selected to be in either ZCR mode or MZCR mode. If E S (m) is lower than the threshold, comparator 2 lowers line 2
An incremental timer 26 is used which generates a series of pulses approximately 0.25 seconds apart on the clock. These pulses act as a signal to the comparator 28. During this time, Z CI (, (m) '9i new circuit 13
The output of is coupled to peak ZCR, circuit 39. This beak ZCR circuit 29 has a plurality of 7. From CR(m), a signal representative of the peak ZCR is generated which is supplied to the comparator 28.

Comparator 28 has two outputs coupled to counter 30. If the peak ZCR is equal to or exceeds the threshold during the time increment period, one output of comparator 28 decrements counter 60 by one. On the other hand, if the peak ZCR is smaller than the threshold, the other output of the comparator 2B is
Increase by one. Counter 60 is connected to comparator 61, filter 2, each having an output coupled to ZCR mode selector switch 36.

When the count value of counter 60 reaches a low threshold (e.g. -4), comparator 31 operates mode switch 33 to provide a signal indicative of the MZCR mode, and also activates mode switch 33 to provide a signal representing the MZCR mode.
The timer 20 is set for one hour over time. On the other hand, when the count value reaches a high threshold (for example, +6), the comparator 2 operates the mode switch 33 to supply a signal instructing the ZCR mode, and the hangover one hour selector 35 sets the hangover one hour of the timer 20. Set. The output of each comparator 61.32 is coupled to a separate input of an OR gate 66. Comparator 31 or 32
The OR gate 36 supplies a reset signal to the counter 3o via the delay circuit 37. The mode selector switch 63 may be a flip-flop circuit.

The output of mode selector switch ro 3 is comparator 38
and 69 available terminals. Only one of comparators 68 and 39 is enabled at a time. When the mode selector switch 33 supplies the ZCR mode signal, the comparator 68 becomes usable nP.
become. The signal representing Z CI'(, (m) is Z CI
t (m) is supplied from the update circuit 13 to the controller L/-38. Z CR(m) is the threshold (e.g. 16
(50 zero crossings between m5), the comparator 68 is enabled by the mode selector switch 66 and the gates 22, 21 .
16, a signal is supplied to turn on the flag switch 17. If Z CR(m) is less than this threshold, comparator 38
and supplies a signal to turn off the flag switch 17 via the hangover timer 20. During normal ZCR mode, the no-hangover timer 20 is set to 32m5 by the hangover timer 35.

When the mode selector switch 33 supplies the MZCR, mode signal, the comparator 39 is enabled. M
The signal representing ZCR(m) (the number of bit-reversal sequences equal to or exceeding TI-h) is MZ
It is supplied from the CR(m) update circuit 14 to a comparator 69. M Z CR (nl) is threshold TH2
If it is equal to or exceeds , then the comparator 39 is
When enabled by the mode selector switch 33, the flag switch 17 is
Generates a signal to turn on. If M Z CR(m) is less than this threshold, comparator 69
provides a signal via gate 25.24.19 and hangover timer 20 that turns off flag switch 17 and inhibits the flag signal. During the I'vl z CR mode, the hangover timer 2o is set to 64m5 by the hangover hour selector 34.

During the ZCR(m) mode, the hangover hour is set to a short delay of, for example, 32m5, but the
During It (m) mode, the hangover timer is set to a long delay, for example 64m5.

Means are provided for resetting timer 26 and peak ZCR circuit 29. Compa Lake 31.32.1
The outputs of 5 and 23 are combined into an OR gate 400A. The presence of a signal from one of these comparators indicating high E (m), high E S (rn) or completed peak Z CIt count and mode selection resets timer 26 and peak Z CR circuit 29. The output of the increment counter 26 is output via a delay circuit 41 to O
R gate coupled to 400 Å force. Timer 26 and peak Z CR circuit 29 are reset after the expiration of the incremental timer 26 time period.

The specific threshold values and time periods used in the foregoing discussion were found through numerous objective and subjective tests for optimization. Other values and time periods can be substituted with some degradation in performance. Furthermore, although the invention has been described using an 8-bit PCM system, it is readily applicable with other systems, including 4-bit PCM systems.

While the presently preferred embodiment of the invention has been described and illustrated, it will be apparent to those skilled in the art that various changes and modifications can be made without departing from the scope of the invention as defined by the claims. For Nagisa, this is obvious.

[Brief explanation of the drawing]

FIG. 1 is a flowchart of a program suitable for implementing the present invention, and FIG. 2 is a logic circuit diagram suitable for implementing the present invention. 10: Input circuits 11 to 14: Parameter update circuits 15.18.23: Comparators 16.19, 22.25: Ort gate 17: Flag switch 20: Late recovery timer 21.24: AND gate 26: Incremental timer 28: Comparator 29: Peak ZCR circuit 30: Counter 51.32: Comparator 33: Mode selector switch 34: Hangover one hour selector 36.40: (JR gate 38.39: Comparator Hiroshi Kazama 1)

Claims (10)

[Claims] (1) A method for providing a flag signal upon detection of a speech signal in a waveform carried by a communication channel, comprising the steps of: coupling to the waveform carried by the channel; and determining the energy level and zero-crossing velocity of the waveform. calculating; comparing the energy level to high and low energy thresholds; providing a flag signal if the energy level is equal to or greater than the high energy threshold; inhibiting a flag signal if the energy level is less than the lower energy threshold; continuing if the energy level is equal to or greater than the lower energy threshold; calculating a level dip difference or energy slope; comparing said energy slope to a slope threshold; and providing a flag signal when said slope is equal to or exceeds said slope threshold. , continuing if the slope is below the threshold; comparing the zero-crossing rate to a zero-crossing threshold; and setting a flag if the zero-crossing rate is equal to or exceeds the zero-crossing threshold. providing a signal, the zero crossing speed being 7! A method for detecting a conbrimentalis beach, comprising the steps of: prohibiting flag transmission when the lfi is below a zero-crossing threshold. 2. The method of claim 1, wherein inhibiting the flag signal is delayed until the end of a time period that elapses when the zero-crossing rate is less than the zero-crossing threshold. (3) A method for providing a flag signal upon detection of a digitized speech signal in a waveform of serial bits carried by a communication channel, comprising the steps of: coupling to the waveform carried by the channel; converting bits into parallel bytes; calculating the energy level during a time slot of said waveform by binary addition of a plurality of bytes; and comparing the average energy level of the plurality of time slots with high and low energy thresholds. providing a flag signal if the energy level is equal to or greater than the higher energy threshold; and inhibiting the flag signal if the energy level is more similar than the lower energy threshold. , continuing if said energy level is equal to or greater than said high energy threshold; and calculating a difference or energy slope between the time-segmented energy levels of two time slots; comparing the slope to a slope threshold; providing a flag signal when the slope equals or exceeds the slope; continuing if the slope is less than the threshold; calculating the zero-crossing rate of the waveform by counting the reversals of polarity bits of the bytes; pt comparing the zero-crossing rate to a zero-crossing threshold; and whether the zero-crossing rate is equal to the zero-crossing threshold. A complementary speech detection method comprising the steps of: providing a flag signal when the zero-crossing rate exceeds the zero-crossing threshold; and disabling the flag signal when the zero-crossing rate is lower than the zero-crossing threshold. 4. The method of claim 3, wherein inhibiting the flag signal is delayed until the end of a time period that elapses when the zero-crossing rate is less than the zero-crossing threshold. 5. The method of claim 6, wherein the average energy level is calculated after each time slot. (6) A method for providing a club signal upon detecting a digitized speech signal in a digitized waveform of serial bits carried by a communication channel, comprising the steps of: converting the serial bits into parallel bytes; and from the change in the polarity bits of the bytes, a plurality of time periods during which energy levels are lower than a high energy threshold and the energy slope is less than the energy slope threshold. calculating a peak number of zero crossings during each of the following steps: adding to a count the number of time periods during which the peak zero crossing is less than a peak zero crossing threshold; subtracting a number of time periods from said count that is greater than said count; indicating a high level system noise condition when said count is less than or equal to a predetermined negative number; indicating a low level system noise condition when equal to or greater than a positive number; and determining the average energy level of said waveform aggregated over one time slot by binary addition of the polarities of said bytes. calculating, and comparing an average energy level of a time slot of a count with high and low energy thresholds, and providing a flag signal when the energy level is equal to or greater than the high energy threshold. disabling a flag signal if the energy level is below the lower energy threshold; and calculating a difference between the temporally spaced energy levels of two time slots, i.e., an energy slope. comparing the energy slope to a slope threshold; providing a flag subdivision if the slope is equal to or exceeds the slope threshold; calculating and counting the number of long bit-flip sequences having successive bit-flips equal to or greater than a successive bit-flip threshold of the second bit-flip sequence; comparing the number of long bit-flip sequences to a threshold; and determining whether the number of long bit-flip sequences is equal to or less than the second bit-flip sequence threshold
retaining a flag signal when the number of long bit-flip sequences is less than the second bit-flip sequence threshold, and inhibiting the flag signal if the number of long bit-flip sequences is less than the second bit-flip sequence threshold; calculate the zero-crossing speed of the waveform if
comparing to a zero-crossing threshold; maintaining a flag signal if the zero-crossing rate is equal to or exceeds the zero-crossing threshold; and if the zero-crossing rate is less than the zero-crossing threshold; and inhibiting a flag signal. 7. The method of claim 6, wherein the prohibition of the flag is deferred until the end of a period of time that elapses when the zero-crossing rate is less than the zero-crossing threshold. 8. The method of claim 7, wherein the average energy level is recalculated after each time slot. (9) An apparatus for providing a flag signal upon detection of a speech signal in a waveform carried by a communication channel, comprising: means for coupling to the waveform carried by the channel; and an energy level and zero-crossing rate of the waveform. means for comparing said energy level with a high and a low energy threshold; and means for providing a flag signal if said energy level is equal to or greater than said high energy threshold. means for inhibiting a flag signal if said energy level is less than said lower energy threshold; and said energy level is equal to or greater than said lower energy threshold; . means for calculating a difference between time-divided energy levels of the waveform, ie, an energy slope; means for comparing the energy slope with a slope threshold; means for providing a flag signal when the slope is equal to or exceeds the threshold; and means for continuing when the slope is less than the threshold; means for comparing to a zero-crossing threshold; means for providing a flag signal if the zero-crossing speed rq is equal to or exceeds the zero-crossing threshold; and means for inhibiting the flag signal when the signal is lower than the flag signal. 10. The apparatus of claim 9, wherein said flag signal 70 is delayed until the end of a time period that elapses when said zero crossing speed is less than said zero crossing threshold. 11. The apparatus of claim 10, wherein: (l) the average energy level is recalculated after each time slot.