Patent Application
20060161430
(1) Field of the Invention
The present invention generally relates to speech detection and/or recognition and more particularly to a system, a circuit and a concomitant method thereof for detecting the presence of a desired signal component within an acoustical signal, especially recognizing a component characterizing human speech. Even more particularly, the present invention is providing a human speaker recognition by means of a detection system with automatically generated activation trigger impulses at the moment a voice activity is detected.
(2) Description of the Prior Art
Sound or acoustical signals are besides others, such as video signals e.g., one main category of analog and—most often also noise polluted—signals modern telecommunications are dealing with; where all signals together—generally after transformation into digital form—are termed as communication data signals. Analyzing and processing such sound signals is an important task in many technical fields, such as speech transmitting and voice recording and becoming even more relevant nowadays, speech pattern or voice recognition e.g. for a command identification to control modern electronic appliances such as mobile phones, navigation systems or personal data assistants by spoken commands, for example to dial the phone number with phones or entering a destination address with navigation systems. In real world environments, many observed acoustical signals to be processed are typically composites of a plurality of signal components. Looking at an audio signal picked up by a microphone within a moving vehicle, the enregistered audio signal may comprise a plurality of signal components, such as audio signals attributed to the engine and the gearbox of the car, the tires rolling on the surface of the road, the sound of wind, noise from other vehicles passing by, speech signals of people chatting within the vehicle and the like. Furthermore, most audio signals are non-stationary, since the signal components vary in time as the situation is changing. In such real world environments, it is often necessary to detect the presence of a desired signal component, e.g., a speech component in an audio signal. Speech detection has many practical applications, including but not limited to, voice or speech recognition applications for spoken commands. Such applications of speech detection as in Voice Operated Trans-(X)-mission (VOX) systems or in baby phones need to have a voice activation module included in order to decide when a voice signal starts so as to generate an activation trigger impulse. Normally the sound level is used for this purpose. The disadvantage thereby is, that—without additional precautions—some kind of noise can also lead to an activation signal. The invention reduces such a misclassification by detecting voice only appearance in a more reliable manner. For speech recognition as known in the art this is an advantageous feature. In general, speech audio input is digitized and then processed to facilitate identification of specific spoken words contained in the speech input. Pursuant to one approach called pattern-matching, so-called features are extracted from the digitized speech and then compared against previously stored patterns to enable such recognition of the speech content. It is easily understandable that, in general, pattern matching can be more successfully accomplished when the input can be accurately characterized as being either speech or non-speech audio input. For example, when information is available to identify a given segment of audio input as being non-speech, that information can be used to beneficially influence the functionality of the pattern matching activity by, for example, simplifying or even eliminating said pattern matching for that particular non-speech segment. Unfortunately, the benefits of voice activity detection are not ordinarily available in speech recognition systems, as the identification of speech is very complex, time-consuming and costly and also considered being not reliable enough. This is where this invention might also come in.
The main problems in performing a reliable human speech detection and voice activation lie in the fact, that the speech detection procedures have to be adapted to all the possible environmental and operational situations in such a way, that always the most apt procedures i.e. algorithms and their optimum parameters are chosen, as no unique procedure on its own is capable of fulfilling all the desired requirements under all conditions. In order to substantiate said situations a bit more, showing all the diversity of environmental and operational situations, a rather casual catalog of questions to be considered is given in the following, whereby no claim for completeness is made. This list of questions is given in order to decide which algorithm is best suited for the specific application and thus illustrates the vast range of possible considerations to be made.
Such questions may be, for example, questions about the audio signal itself, about the environment, about technical and manufacturing aspects, such as:
And so on. Depending on the outcome of the answers to these questions it will then be decided, which algorithm is best suited for the specific application. Some relevant answers to these questions will be given later.
Preferred prior art realizations are implementing speech detection and voice activation procedures via single chip or multiple chip solutions as integrated circuits. These solutions are therefore on one hand, only usable with optimum results for certain well defined cases, thus exhibiting however a somewhat limited complexity or are on the other hand very complex and use extremely demanding algorithms requiring great processing power, thus offering however greater flexibility with respect to their adaptability. The limitation in applicability of such a low-cost circuit on one hand and the complexity and the power demands of such a higher quality circuit on the other hand are the main disadvantages of these prior art solutions. These disadvantages pose major problems for the propagation of that sort of circuits. It is therefore a challenge for the designer of such devices and circuits to achieve a high-quality and also low-cost solution.
Several prior art inventions referring to such solutions describe related methods, devices and circuits, and there are also several such solutions available with various patents referring to comparable approaches, out of which some are listed in the following:
U.S. Pat. No. 6,691,087 (to Parra et al.) shows a method and an apparatus for adaptive speech detection by applying a probabilistic description to the classification and tracking of signal components, wherein a signal processing system for detecting the presence of a desired signal component by applying a probabilistic description to the classification and tracking of various signal components (e.g., desired versus non-desired signal components) in an input signal is disclosed.
U.S. Pat. No. 6,691,089 (to Su et al.) discloses user configurable levels of security for a speaker verification system, whereby a text-prompted speaker verification system that can be configured by users based on a desired level of security is employed. A user is prompted for a multiple-digit (or multiple-word) password. The number of digits or words used for each password is defined by the system in accordance with a user set preferred level of security. The level of training required by the system is defined by the user in accordance with a preferred level of security. The set of words used to generate passwords can also be user configurable based upon the desired level of security. The level of security associated with the frequency of false accept errors verses false reject errors is user configurable for each particular application.
U. S. Patent Application 20020116186 (to Strauss et al.) describes an integrated voice activity detector for integrated telecommunications processing for detecting whether voice is present. In one embodiment, the integrated voice activation detector includes a semiconductor integrated circuit having at least one signal processing unit to perform voice detection and a storage device to store signal processing instructions for execution by the at least one signal processing unit to: detect whether noise is present to determine whether a noise flag should be set, detect a predetermined number of zero crossings to determine whether a zero crossing flag should be set, detect whether a threshold amount of energy is present to determine whether an energy flag should be set, and detect whether instantaneous energy is present to determine whether an instantaneous energy flag should be set. Utilizing a combination of the noise, zero crossing, energy, and instantaneous energy flags the integrated voice activation detector determines whether voice is present.
U. S. Patent Application 20030120487 (to Wang) describes the dynamic adjustment of noise separation in data handling, particularly voice activation wherein data handling dynamically responds to changing noise power conditions to separate valid data from noise. A reference power level acts as a threshold between dynamically assumed noise and valid data, and dynamically refers to the reference power level changing adaptively with the background noise. The introduction of dynamic noise control in VOX (Voice Activated Transmission) improves a VOX device operation in a noisy environment, even when the background noise profiles are changing. Processing is on a frame by frame basis for successive frames. The threshold is adaptively changed when a comparison of frame signal power to the threshold indicates speech or the absence of speech in the compared frame repeatedly and continuously for a period of time involving plural successive frames having no valid speech or noise above the threshold to correspondingly reduce or increase the threshold by changing the threshold to a value that is a function of the input signal power.
U. S. Patent Application 20040030544 (to Ramabadran) describes a distributed speech recognition with back-end voice activity detection apparatus and method, where a back-end pattern matching unit can be informed of voice activity detection information as developed through use of a back-end voice activity detector. Although no specific voice activity detection information is developed or forwarded by the front-end of the system, precursor information as developed at the back-end can be used by the voice activity detector to nevertheless ascertain with relative accuracy the presence or absence of voice in a given set of corresponding voice recognition features as developed by the front-end of the system.
Although these papers describe circuits and/or methods close to the field of the invention they differ in essential features from the method, the system and especially the circuit introduced here.
A principal object of the present invention is to realize a very flexible and adaptable voice activation circuits module in form of very manufacturable integrated circuits at low cost.
Another principal object of the present invention is to provide an adaptable and flexible method for operating said voice activation circuits module implementable with the help of integrated circuits.
Also another principal object of the present invention is to include determinations of “Noise estimation and “Speech estimation” values, done effectively without use of Fast Fourier Transform (FFT) methods or zero crossing algorithms only by analyzing the modulation properties of human voice.
Also an object of the present invention is to include tailorable operating features into a modular device for implementing multiple voice activation circuits and at the same time to reach for a low-cost realization with modern integrated circuit technologies.
Further an object of the present invention is to always operate the voice activation device with its optimum voice activation algorithm.
Also further an object of the present invention is the inclusion of multiple diverse voice activation algorithms into the voice activation device.
Another further object of the present invention is to combine the function of multiple diverse voice activation algorithms within the voice activation device operating.
Also an object of the present invention is to establish a building block system for a voice activation device, capable of being tailored to function effectively under different acoustical conditions.
Also another object of the present invention is to facilitate by said building block approach for said voice activation device solving operating problems necessitating future expansions of the circuit.
Further another object of the present invention is to streamline the production by implementing the voice activation device with a limited gate count, i.e. to limit its complexity counted by number of transistor functions needed.
A further object of the present invention is to make the voice activation circuit as flexible as possible by previsioning modules and interconnections necessary to implement algorithms of future developments.
A still further object of the present invention is to reduce the power consumption of the circuit by realizing inherent appropriate design features.
Another further object of the present invention is to reduce the cost of manufacturing by implementing the circuit as a monolithic integrated circuit in low cost CMOS technology.
Another still further object of the present invention is to reduce cost by effectively minimizing the number of expensive components.
In accordance with the objects of this invention, a new system for a tailorable and adaptable implementation of a voice activation function is described, capable of a practical application of multiple voice activation algorithms, receiving an audio input signal and furnishing a trigger impulse as output signal, comprising an analog audio signal pick-up sensor; an analog/digital converting means digitizing said audio signal and thus transforming said audio signal into a digital signal, then named ‘Digital Audio Input Signal’; a modular assembly of multiple voice activation algorithm specific circuits made up of building block modules containing processing means for amplitude and energy values of said ‘Digital Audio Input Signal’ as well as and especially for Noise and Speech estimation calculations, intermediate storing means, comparing means, connecting means and means for selecting and operating said voice activation algorithms; and a means for generating said trigger impulse.
Also in accordance with the objects of this invention, a new method for a general tailorable and adaptable voice activation circuits system is described, capable of implementing multiple diverse voice activation algorithms with an input terminal for an audio input signal and an output terminal for a generated voice activation trigger signal and being composed of four levels of building block modules together with two levels of connection layers, altogether being dynamically set-up, configured and operated within the framework of a flexible timing schedule, comprising at first providing as processing means—four first level modules named “Amplitude Processing” block, “Energy Processing” block, “Noise Processing” block and “Speech Processing” block, which act on its input signal named ‘Digital Audio Input Signal’ either directly or indirectly, i.e. either on its amplitude value as input variable or on processed derivatives thereof, i.e. on energy, noise and speech values as processing variables; providing as storing means four pairs of second level modules designated as value and threshold storing blocks or units respectively, namely for intermediate storage of pairs of amplitude, signal energy, noise energy and speech energy values in each case, named “Amplitude Threshold” and “Amplitude Value”, “Energy Threshold” and “Energy Value”, “Noise Threshold” and “Noise Energy Value”, as well as “Speech Threshold” and “Speech Energy Value”; providing as comparing means within a third level of modules four comparator blocks, named “Amplitude Comparator”, “Energy Comparator”, “Noise Comparator”, and “Speech Comparator”; providing as triggering means and fourth module level an “IRQ Logic” block together with its “IRQ Status/Config” block, delivering an IRQ output signal for voice activation; providing also a “First Interconnection Layer” within and between said first level modules for processing said ‘Digital Audio Input Signal’ values from its amplitude, energy, noise and speech variables and said second level modules, whereby said amplitude value of said ‘Digital Audio Input Signal’ may be fed into said “Amplitude Processing” block, and/or into said “Energy Processing” block, and/or into said “Noise Processing” block and/or into said “Speech Processing” block, thus receiving from each other already processed values as possible input and/or control signals separately or in parallel and whereby finally from all said processing the resulting variables with their calculated and/or estimated values are fed into said respective second level storing units, named “Amplitude Value”, “Signal Energy Value”, “Noise Energy Value”, and “Speech Energy Value”; providing further a “Second Interconnection Layer” between said second and third level of modules for storing and comparing said processed values of said amplitude, energy, noise (SNR) and speech variables, whereby always the corresponding values of threshold and variable result pairs are fed into their respective comparator blocks located within said third level of modules and whereby said comparator blocks may also receive via an extra input additional control signals from others of said second level modules; providing an extra “Config” block for setting-up and configuring all necessary threshold values and operating states for said blocks within all four levels of modules according to said voice activation algorithm to be actually implemented; connecting the output of each of said comparators in module level three to said fourth level “IRQ Logic” block as inputs; establishing a recursively adapting and iteratively looping and timing schedule as operating scheme for said tailorable voice activation circuits system capable of implementing multiple diverse voice activation algorithms and thus being able to being continuously adapted for its optimum operation; initializing with pre-set operating states and pre-set threshold values a start-up operating cycle of said operating scheme for said voice activation circuit; starting said operating scheme for said adaptable voice activation circuits system by feeding said ‘Digital Audio Input Signal’ as sampled digital amplitude values into the circuit, namely said “First Interconnection Layer”, for further processing e.g. by calculating said signal energy, and/or by estimating said noise energy and/or said speech energy; deciding upon said voice activation algorithm to be chosen for actual implementation with the help of crucial variable values such as said amplitude value from said audio signal input variable and also said already calculated and estimated signal energy, noise energy and speech energy values as processing variables critical and crucial for said voice activation algorithm and in conjunction with some sort of a decision table, leading to optimum choices for said voice activation algorithms; setting-up the operating function of said “First Interconnection Layer” element appropriately with the help of said “Status/Config” block considering the requirements of said voice activation algorithm to be actually implemented for the connections within and between said first and second level modules; setting-up the operating function of said “Second Interconnection Layer” element appropriately with the help of said “Status/Config” block considering the requirements of said voice activation algorithm to be actually implemented for the connections within and between said second and third level modules; configuring said necessary operating states e.g. in internal modules each with specific registers by algorithm defining values corresponding to said actually chosen voice activation algorithm for future operations; setting-up the operating function of said “IRQ Logic” block appropriately with the help of said “IRQ Status/Config” block considering said voice activation algorithm to be actually implemented; processing continuously within said “Energy Processing” block e.g. said “Signal Energy Value” calculation, acting on said input signal named ‘Digital Audio Input Signal’; processing continuously within said “Noise Processing” block e.g. said “Noise Energy Value” estimation, which depends on its input signal, e.g. said already formerly calculated “Signal Energy Value”; processing continuously within said “Speech Estimation” block e.g. said “Speech Energy Value”, which depends on its input signal, e.g. said already formerly calculated “Signal Energy Value”; storing within its corresponding storing units located within module level two the results of said preceding “Amplitude Processing”, “Energy Processing”, “Noise Processing” and “Speech Processing” operations, namely said “Amplitude Value”, “Signal Energy Value”, “Noise Energy Value”, and “Speech Energy Value” all taken directly or indirectly from said ‘Digital Audio Input Signal’; setting-up within said storing units said respective threshold values named “Amplitude Threshold”, “Energy Threshold”, “Noise Threshold” and “Speech Threshold” corresponding to said actually chosen voice activation algorithm for future comparing operations; comparing with the help of said “Amplitude Comparator”, “Energy Comparator”, “Noise Comparator”, and “Speech Comparator” said “Amplitude Threshold” and “Amplitude Value”, said “Energy Threshold” and “Signal Energy Value”, said “Noise Threshold” and “Noise Energy Value”, as well as said “Speech Threshold” and “Speech Energy Value”; evaluating the outcome of the former comparing operations within said “IRQ Logic” block with respect to said earlier set-up operating function; generating, depending on said “IRQ Logic” evaluation in the case where applicable a trigger event as IRQ impulse signalling a recognized speech element for said voice activation; and re-starting again said once established operating scheme for said voice activation circuits system from said starting point above and continue its looping schedule.
Further in accordance with the objects of this invention, a circuit, implementing said new method is achieved, realizing a voice activation system capable of implementing multiple voice activation algorithms and being composed of four levels of building block modules as well as connection means, receiving an audio input signal and furnishing a trigger impulse as output signal, comprising an input terminal as entry for said audio input signal into a first level of modules; a first level of modules consisting of a set of processing modules including modules for signal amplitude preparation, energy calculation and especially noise and speech estimation; a second level of modules consisting of a set of intermediate storage modules for threshold and signal values; a multipurpose connection means in order to transfer said audio input signal to said first level modules and to appropriately connect said first level modules to each other and to said second level of modules; a third level of modules consisting of comparator modules; a fourth level of modules as trigger generating means including additional configuration, setup and logic modules; and an output terminal for said IRQ signal as said output signal in form of said trigger impulse.
In the accompanying drawings forming a material part of this description, the details of the invention are shown:
The preferred embodiments disclose a novel optimized circuit with a modules conception for a speech detection and voice activation system using modern integrated circuits and an exemplary implementation thereto.
As already stated above speech detection procedures do have to be adapted to all the possible environmental and operational situations in such a way, that always the most apt procedures i.e. algorithms and their optimum parameters are chosen, as no unique procedure on its own is capable of fulfilling all the desired requirements under all conditions. Therefore it is suitable to answer certain relevant questions about the audio signal itself, about the environment, about technical and manufacturing aspects as re-listed in the following:
This question is the base for the algorithm. If the signal, which has to be detected, is loud in comparison to the background noises, the used algorithm can be very simple. Unfortunately, in most cases the background noises can be often very loud and the application has to cope with it. If the background noise is low however, only the signal amplitude or the signal energy value can be used.
If the algorithm has to handle loud background noises, it would be good to know more about the sound signal. If it is a speech signal, special characteristics of the speech can be used to differentiate between the activation signal and the background noise. In the case of baby sounds, the voice activation can use characteristics from baby sounds. If the activation sound is artificial, the algorithm can be adapted to this special sound. This is however only really useful for speech or baby sounds. For other especially artificial sounds only the amplitude or energy values should be used.
After carefully considering and evaluating the answers given to such questions—whereby that list above may be expanded and amended in many ways—a choice has to be made out of a pool of voice activation algorithms for the specific application. One question to be answered hereby is who, where and when makes this choice? It can either be made during design i.e. statically before operation. Or it can be done by the user dynamically responding during a configuration phase. Both ways of working are possible.
Then this choice of algorithms has to be adequately implemented in the most efficient and economical way. Therefore, now referring to
What shall be especially emphasized here is the introduction of two separate and specific blocks for “Noise Estimation” and “Speech Estimation” in combination with “Amplitude Preparation” and “Energy Calculation” and their corresponding threshold triggering operations as main parts of this invention.
Said building blocks are adaptively tailored to handle certain relevant and well known case specific operational characteristics describing the acoustical differing cases analyzed by such a list of questions as collocated above and leading to said choice of algorithms. Said algorithms are then realized and activated by tailoring said building blocks within said actual voice activation circuit device according to the method of this invention, explained and described with the help of a flow diagram given later in
In the following the introduction of two so called “interconnection layers” is explained, expanding the structure of
The block diagram in
Studying FIGS. 1H-IL the generalized method according to this more general module structure of
Delving now into
Contemplating now
Before describing the particular methods according to
In the following some more detailed remarks to the interaction between said levels of main modules when implementing said voice activation algorithms are made, thus clarifying some of the underlying operating principles, which inversely in other words, could also be transformed into or deduced from every single description of the corresponding voice activation algorithms, whatever is appropriate.
Module 320, denominated as “Amplitude Comparator”, which compares the actual “Amplitude Value” 220—directly derived from said Digital Audio Input Signal 110—with the previously stored “Amplitude Threshold” 225 is the primary module for implementing a “Threshold Detection on Signal Amplitude” algorithm ALGO1, to be more explicitly described later. Whenever the “Amplitude Value” 220 exceeds the “Amplitude Threshold” 225 the “Amplitude Comparator” 320 signs this to the IRQ Logic 400. For the implementation of a “Threshold Detection on Signal Energy” algorithm ALGO3 said module 140 provides an “Energy Calculation” function, which is realized as e.g. an ordinary low pass filter on the absolute signal value or squared signal calculating the actual “Energy Value” 240. Said “Energy Comparator” 340 compares the actual “Energy Value” 240 with an “Energy Threshold” 245. If the “Energy Value” 240 exceeds the “Energy Threshold” 245 the “Energy Comparator” 340 signs this to the “IRQ Logic” 400. An “Automatic Threshold Adaptation on Background Noise” algorithm ALGO2 is implemented starting with module 160, which includes the “Noise Estimation” operation, which is realized by a minimum detection unit detecting the minimum of the energy in a moving window. This minimum is the estimation for the noise in the signal, termed as “Noise Energy Value” 260. The “SNR Comparator” 360 calculates from the actual “Noise Energy Value” 260 and the actual “Speech Energy Value” 280 the actual SNR and compares it with an “SNR Threshold” 265. If the SNR exceeds the “SNR Threshold” 265 the “SNR Comparator” 360 signs this to the “IRQ Logic” 400. The implementation of a “Threshold Detection on Speech Energy” algorithm ALGO4 includes module 180, which is described as the “Speech Estimation” unit which performs a subtraction of the “Noise Energy Value” 260 from the energy value stored in “Speech Energy Value” 280. The “Speech Comparator” 380 compares the “Speech Energy Value” 280 with a “Speech Threshold” 285 and signs the result to the IRQ Logic 400. A description for “SNR” algorithm ALGO5 has to include mentioning the use of all available modules in order to calculate the Signal-to-Noise Ratio (SNR), which is defined as the ratio of Speech energy to Noise energy, wherein the energy ‘E’ accumulated within a certain number ‘n’ of samples of digital amplitude values is generally calculated in digital signal processing systems from sampled digital Signal amplitude values ‘s(n)’ as ‘E’=[Sum of all ‘n’ values (‘s(n)’ times ‘s(n)’)] or using the much easier to implement procedure of Low-Pass (LP) filtering said (‘s(n)’ times ‘s(n)’) product, as it is done in said “Energy Calculation” block. The determination of the Signal-to-Noise Ratio (SNR) needs using the resulting energy values from said “Noise Estimation” block and said “Speech Estimation” block, whereby Speech is defined as the difference of Signal energy minus Noise energy.
Completing now, the IRQ Logic 400 can be configured in such a way, that one can select which type of voice activation should be used, whereby said voice activation algorithms as directly implemented or even boolean combinations of these algorithms can be set-up. As the circuit is already capable to evaluate all the described signal parameters it could be advantageous also to use said parameters to perform other auxiliary functions, e.g. using the feature noise estimation for the control of a speaker volume.
A first method, belonging to the block diagram of
The method continues with step 520 starting said operating scheme for said adaptable voice activation circuits system by feeding said ‘Digital Audio Input Signal’ as sampled digital amplitude values into the circuit, and by calculating said signal energy within said “Energy Calculation” block, and estimating said noise energy (also used for SNR determination) and said speech energy within said “Noise Estimation” block and said “Speech Estimation” block; then step 530 decides upon said voice activation algorithm to be chosen for actual implementation with the help of crucial variable values such as said amplitude value from said audio signal input variable and also said already calculated and estimated signal energy, noise energy and speech energy values as processing variables critical and crucial for said voice activation algorithm and in conjunction with some sort of decision table, leading to optimum choices for said voice activation algorithms. It shall be emphasized here, that within steps 520 & 530 said “Noise estimation and “Speech estimation” can be done effectively without use of Fast Fourier Transform (FFT) methods or zero crossing algorithms only by analyzing the modulation properties of human voice.
Two more steps, 532 and 534, are needed to configure said necessary operating states e.g. in internal modules each with specific registers by algorithm defining values corresponding to said actually chosen voice activation algorithm for future operations and to set-up the operating function of said “IRQ Logic” block appropriately with the help of said “IRQ Status/Config” block considering said voice activation algorithm to be actually implemented. The method now calculates continuously within said “Energy Calculation” block said “Energy Value”, acting on said input signal named ‘Digital Audio Input Signal’ in step 540, in steps 542 and 544 estimates continuously within said “Noise Estimation” block said “Noise Energy Value”, and within said “Speech Estimation” block said “Speech Energy Value”, which both depend on that input signal, namely said already formerly in step 540 calculated “Energy Value”. Step 550 then stores within its corresponding storing units located within module level two the results of said preceding “Energy Calculation”, “Noise Estimation” and “Speech Estimation” operations, namely said “Energy Value”, “Noise Energy Value”, and “Speech Energy Value” as well as said “Amplitude Value” taken directly from said ‘Digital Audio Input Signal’. It is now in step 552, the method sets-up within said storing units said respective threshold values named “Amplitude Threshold”, “Energy Threshold”, “SNR Threshold” and “Speech Threshold” corresponding to said actually chosen voice activation algorithm for future comparing operations before step 560 compares with the help of said “Amplitude Comparator”, “Energy Comparator”, “Noise (SNR) Comparator”, and “Speech Comparator” said “Amplitude Threshold” and “Amplitude Value”, said “Energy Threshold” and “Energy Value”, said “SNR Threshold” and “Noise Energy Value”, as well as said “Speech Threshold” and “Speech Energy Value”. Coming near the end of the method, step 570 evaluates the outcome of the former comparing operations within said “IRQ Logic” block with respect to said earlier set-up operating function and generates in step 580, depending on said “IRQ Logic” evaluation in the case where applicable a trigger event as IRQ impulse signalling a recognized speech element for said voice activation. Finally step 590 serves to re-start again said once established operating scheme for said voice activation circuits system from said starting point above and continue its looping schedule.
This concludes the description of the first method belonging to the structure of
The method continues with step 620 starting said operating scheme for said adaptable voice activation circuits system by feeding said ‘Digital Audio Input Signal’ as sampled digital amplitude values into the circuit, namely said “First Interconnection Layer”, for further processing e.g. by calculating said signal energy, and/or by estimating said noise energy and/or said speech energy; then step 630 decides upon said voice activation algorithm to be chosen for actual implementation with the help of crucial variable values such as said amplitude value from said audio signal input variable and also said already calculated and estimated signal energy, noise energy and speech energy values as processing variables critical and crucial for said voice activation algorithm and in conjunction with some sort of a decision table, leading to optimum choices for said voice activation algorithms. Two steps, 632 and 634, are needed to set-up the operating function of said “First Interconnection Layer” element appropriately with the help of said “Status/Config” block considering the requirements of said voice activation algorithm to be actually implemented for the connections within and between said first and second level modules and to set-up the operating function of said “Second Interconnection Layer” element appropriately with the help of said “Status/Config” block considering the requirements of said voice activation algorithm to be actually implemented for the connections within and between said second and third level modules. Two more steps, 636 and 638, are needed to further configure said necessary operating states e.g. in internal modules each with specific registers by algorithm defining values corresponding to said actually chosen voice activation algorithm for future operations and to set-up the operating function of said “IRQ Logic” block appropriately with the help of said “IRQ Status/Config” block considering said voice activation algorithm to be actually implemented. The method now processes continuously in the following steps—640, 642 and 644—within said “Energy Processing” block e.g. said “Signal Energy Value” calculation, acting on said input signal named ‘Digital Audio Input Signal’ and within said “Noise Processing” block e.g. said “Noise Energy Value” estimation, and within said “Speech Estimation” block e.g. said “Speech Energy Value”, which both depend on that input signal, e.g. said already formerly in step 640 calculated “Signal Energy Value”. Step 650 then stores within its corresponding storing units located within module level two the results of said preceding “Amplitude Processing”, “Energy Processing”, “Noise Processing” and “Speech Processing” operations, namely said “Amplitude Value”, “Signal Energy Value”, “Noise Energy Value”, and “Speech Energy Value” all taken directly or indirectly from said ‘Digital Audio Input Signal’. It is now in step 652, the method sets-up within said storing units said respective threshold values named “Amplitude Threshold”, “Energy Threshold”, “Noise Threshold” and “Speech Threshold” corresponding to said actually chosen voice activation algorithm for future comparing operations before step 660 compares with the help of said “Amplitude Comparator”, “Energy Comparator”, “Noise Comparator”, “SNR Comparator”, and “Speech Comparator” said “Amplitude Threshold” and “Amplitude Value”, said “Energy Threshold” and “Signal Energy Value”, said “Noise Threshold” and “Noise Energy Value”, as well as said “Speech Threshold” and “Speech Energy Value”. Coming now near the end of the method, step 670 evaluates the outcome of the former comparing operations within said “IRQ Logic” block with respect to said earlier set-up operating function and generates in step 680, depending on said “IRQ Logic” evaluation in the case where applicable a trigger event as IRQ impulse signalling a recognized speech element for said voice activation. Finally step 690 serves to re-start again said once established operating scheme for said voice activation circuits system from said starting point above and continue its looping schedule.
In the next paragraph said different voice activation algorithms are described in more detail, the implementations of which are covered by the voice activation module circuit shown with the help of the block diagram in
As introduction however, explaining a method to discriminate between the different voice activation algorithms, some words about experimental testing and theoretical analysis procedures and the presentation of its results have to be premised. When looking at
The algorithms considered here for voice activation purposes are basically the already known five algorithms ALGO1 to ALGO5, namely said “Threshold Detection on Signal Amplitude” algorithm—ALGO1; said “Automatic Threshold Adaptation on Background Noise” algorithm—ALGO2; said “Threshold Detection on Signal Energy” algorithm—ALGO3; said “Threshold Detection on Speech Energy” algorithm—ALGO4; and said “Signal to Noise Ratio (SNR)” algorithm—ALGO5 and now thoroughly explained:
“Threshold Detection on Signal Amplitude” Algorithm ALGOL:
The signal amplitude includes all sound information coming from the microphone limited only by the frequency characteristic of the microphone and the amplifiers. A threshold is used to determine, if a sound is loud enough to sign activation. Although “is loud enough” normally means the energy is high enough, the amplitude gives a more or less good substitution to the normally used energy. But there are some exceptions: there might be a high amplitude value although there is only very limited energy in the signal; the worst case would be a delta peak. On the other hand there might be a high energy, but overall a very small amplitude. In these special cases the amplitude would not show the loudness of the signal. The ‘Modulated White Noise’ diagram in
“Automatic Threshold Adaptation on Background Noise” Algorithm ALGO2:
The “Threshold Detection on Signal Amplitude” algorithm ALGO1 can be enhanced by measuring the background noise level and its subtraction from the actual amplitude level, or by increasing the detection threshold accordingly. The white noise modulation diagram in
“Threshold Detection on Signal Energy” Algorithm ALGO3:
As mentioned before, the amplitude is not a measure for the loudness. The signal energy is the low pass filtered square of the signal amplitude. In many cases, the calculation of the square is too complicated and the absolute amplitude values are low pass filtered instead. This is a good substitute for the signal energy. For the diagram shown in
“Threshold Detection on Speech Energy” Algorithm ALGO4:
If the audio signal for activation consists of speech (or baby sounds) the signal energy algorithm ALGO3 can be enhanced in a similar way as done for the amplitude detection enhancement in algorithm ALGO2. If the slow changing noises are estimated and subtracted from the signal energy, the speech energy would be the result. As one can see from the diagram in
The crucial points of both algorithms ALGO3 and ALGO4 are thus the application of Noise estimation and espec. Speech estimation methods, implemented by special hardware processing blocks. These signal processing hardware is also needed to realize the following algorithm.
“Signal to Noise Ratio (SNR)” Algorithm ALGO5:
This SNR algorithm (ALGO5) takes into account that a person unconsciously speaks louder if there are high background noises. In this noisy environment the activation should be detected at higher speech energy levels than in environments with low background noises. Said SNR algorithm (ALGO5) sets its activation threshold (of the speech energy) on a defined percentage of the noise energy, which for example can be set to a value of 25% to 400%. In the case of 100% the activation is detected, when the speech energy level is as high (or higher) as the noise energy level. This algorithm should be combined with the previous algorithms, because in silent environments the threshold is so small that calculation errors can lead to a misclassification. A good combination would be to use the speech energy algorithm ALGO4 until the noise energy rises to similar values as the speech energy threshold and then to switch to this SNR algorithm (ALGO5) for higher noises. This algorithm can be used in very low SNR environments; it is (nearly) independent from environmental conditions. Similar to the speech energy algorithm ALGO4 the SNR algorithm is fast (<15 ms), the power consumption is low and the area of needed silicon is minimal. Because of the attenuation of the different noises, there are only few misclassifications and it has a good reliability.
Each algorithm has advantages and disadvantages. To choose the right algorithm it is important to answer the list of questions from above. Helpful therefore would be a summary of all relevant features and characteristical data for each algorithm given at its best in form of a decision table. In the following TABLE 1 an overview for all the pertaining voice activation algorithms and their pertinent characteristics is given, whereby the column headers (on the top of the table) are characterizing the pertinent voice activation algorithm, and the row headers (on the left side of the table) are listing relevant aspects and variables for each algorithm.
Remarks:
“Hardware” Yes signifies, that a practical hardware implementation is provided.
“Delay” is the time for producing an activation signal after an activation event.
“Size” should give a hint for the silicon area needed. Precise values cannot be given, because these values depend directly on the technology used.
“SNR” gives a hint about the environmental conditions, where the algorithm should be used.
Recurring back to
Summarizing the essential features of the invention we find, that a circuit and a method are given, to realize a very flexible voice activation system using a modular building block approach, that is adaptively tailored to handle certain relevant and case specific operational characteristics describing most of the possible acoustical differing environmental cases to be found in the field of speech recognition. Included are determinations of “Noise estimation and “Speech estimation” values, done effectively without use of Fast Fourier Transform (FFT) methods or zero crossing algorithms only by analyzing the modulation properties of human voice.
As shown in the preferred embodiments and evaluated by circuit analysis, the novel system, circuits and methods provide an effective and manufacturable alternative to the prior art.
While the invention has been particularly shown and described with reference to the preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made without departing from the spirit and scope of the invention.
| Number | Date | Country | Kind |
|---|---|---|---|
| 05368003.9 | Jan 2005 | EP | regional |
