TIME-BASED FREQUENCY TUNING OF ANALOG-TO-INFORMATION FEATURE EXTRACTION
1. A method comprising:
- receiving an analog signal;
extracting a first feature from the analog signal, wherein the first feature is associated with a first frequency band;
extracting a second feature from the analog signal, wherein the second feature is associated with a second frequency band that is different from the first frequency band;
multiplexing the first feature and the second feature to form a feature sequence; and
performing analog-to-digital conversion of the feature sequence to form a digitized feature sequence.
A sound recognition system including time-dependent analog filtered feature extraction and sequencing. An analog front end (AFE) in the system receives input analog signals, such as signals representing an audio input to a microphone. Features in the input signal are extracted, by measuring such attributes as zero crossing events and total energy in filtered versions of the signal with different frequency characteristics at different times during the audio event. In one embodiment, a tunable analog filter is controlled to change its frequency characteristics at different times during the event. In another embodiment, multiple analog filters with different filter characteristics filter the input signal in parallel, and signal features are extracted from each filtered signal; a multiplexer selects the desired features at different times during the event.
- 1. A method comprising:
receiving an analog signal; extracting a first feature from the analog signal, wherein the first feature is associated with a first frequency band; extracting a second feature from the analog signal, wherein the second feature is associated with a second frequency band that is different from the first frequency band; multiplexing the first feature and the second feature to form a feature sequence; and performing analog-to-digital conversion of the feature sequence to form a digitized feature sequence.
- View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9)
- 10. An integrated circuit comprising:
first feature extraction circuitry coupled to receive an analog signal and to provide a first signal feature from the analog signal, wherein the first signal feature is associated with a first frequency characteristic; second feature extraction circuitry coupled to receive the analog signal and to provide a second signal feature from the analog signal, wherein the second signal feature is associated with a second frequency characteristic that is different from the first frequency characteristic; a multiplexer coupled to the first feature extraction circuitry and the second feature extraction circuitry to provide a feature sequence based on the first and second signal features; and an analog-to-digital converter coupled to the multiplexer to provide a digitized feature sequence in response to the feature sequence.
- View Dependent Claims (11, 12, 13, 14, 15)
- 16. An integrated circuit comprising:
first feature extraction circuitry coupled to receive an analog signal and to extract a first signal feature based on a first property of the analog signal from a first group consisting of;
zero-crossing and total energy of a first frequency band;
second feature extraction circuitry coupled to receive the analog signal and to extract a second signal feature based on a second property of the analog signal from a second group consisting of;
zero-crossing and total energy of a second frequency band that is different from the first frequency band;
a multiplexer coupled to the first feature extraction circuitry and the second feature extraction circuitry to provide a feature sequence comprising the first and second signal features; and an event trigger coupled to the multiplexer to receive the feature sequence and to provide a trigger in response to a match of the feature sequence.
- View Dependent Claims (17, 18, 19, 20)
This application is a continuation of prior U.S. application Ser. No. 14/920,210, filed Oct. 22, 2015, to Zhenyong Zhang, et al., titled “Time-Based Frequency Tuning of Analog-to-Information Feature Extraction,” which is herein incorporated by reference in its entirety.
This invention is in the field of active sensing of audio inputs. Embodiments are directed to the detection of particular features in sensed audio.
Recent advancements in semiconductor manufacturing and sensor technologies have enabled new capabilities in the use of low power networks of sensors and controllers to monitor environments and control processes. These networks are being envisioned for deployment in a wide range of applications, including transportation, manufacturing, biomedical, environmental management, safety, and security. Many of these low power networks involve machine-to-machine (“M2M”) communications over a wide-area network, such a network now often referred to as the “Internet of Things” (“IoT”).
The particular environmental attributes or events that are contemplated to serve as input to sensors in these networks are also wide-ranging, including conditions such as temperature, humidity, seismic activity, pressures, mechanical strain or vibrations, and so on. Audio attributes or events are also contemplated to be sensed in these networked systems. For example, in the security context, sensors may be deployed to detect particular sounds such as gunshots, glass breaking, human voices, footsteps, automobiles in the vicinity, animals gnawing power cables, weather conditions, and the like.
The sensing of audio signals or inputs is also carried out by such user devices as mobile telephones, personal computers, tablet computers, automobile audio systems, home entertainment or lighting systems, and the like. For example, voice activation of a software “app” is commonly available in modern mobile telephone handsets. Conventional voice activation typically operates by detecting particular features or “signatures” in sensed audio, and invoking corresponding applications or actions in response. Other types of audio inputs that can be sensed by these user devices include background sound, such as whether the user is an office environment, restaurant, in a moving automobile or other conveyance, in response to which the device modifies its response or operation.
Low power operation is critical in low-power network devices and in battery-powered mobile devices, to allow for maximum flexibility and battery life, and minimum form factor. For example, it has been observed that some types of sensors, such as wireless environmental sensors deployed in the IoT context, can use a large fraction of their available power on environmental or channel monitoring while waiting for an anticipated event to occur. This is particularly true for acoustic sensors, considering the significant amount of power typically required in voice and sound recognition. Conventional sensors of this type typically operate according to a low power, or “sleep,” operating mode in which the back end of the sensor assembly (e.g., the signal transmitter circuitry) is effectively powered down pending receipt of a signal indicating the occurrence of the anticipated event. While this approach can significantly reduce power consumption of the sensor assembly, many low duty cycle systems in which each sensor assembly spends a very small amount of time performing data transmission still consume significant power during idle periods, so much so as to constitute a major portion of the overall power budget.
Digital logic 330 of system 300 converts digital samples 324 to sound information (D2I) in this conventional system 300. Digital logic 330 is typically realized by a general purpose microcontroller units (MCU), a specialty digital signal processor (DSP), an application specific integrated circuit (ASIC), or another type of programmable logic, and in this arrangement partitions the samples into frames 340 and then transforms 342 the framed samples into information features using a defined transform function 344. These information features are then mapped to sound signatures (I2S) by pattern recognition and tracking logic 350.
Recognition logic 350 is typically implemented by one or more types of known pattern recognition techniques, such as a Neural Network, a Classification Tree, Hidden Markov models, Conditional Random Fields, Support Vector Machine, etc., and operates in a periodic manner as represented by time points t0 360, t1 361, t2 362, etc. For example, each information feature (e.g., feature 346) generated by transformation 342 is compared to a database 370 of pre-identified features. At each time step, recognition logic 350 attempts to find a match between a sequence of information features produced by transformation logic 342 and a sequence of sound signatures stored in data base 370. Each candidate signatures 352 that is identified is assigned a score value indicating the degree of match between it and features in database 370. Those signatures 352 having a score for exceeding a threshold value, are identified by recognizer 300 as a match with a known signature.
Because the complex signal segmentation, signal transformation and final pattern recognition operations are performed in the digital domain in recognition system 300, high-performance and high-precision realizations of ADC 322 and the rest of analog-front-end (AFE) 320 are required to provide an adequate digital signal for the following complex digital processing. For example, audio recognition of a sound signal with an 8 kHz bandwidth by a typical conventional sound recognition system will require an ADC with 16-bit accuracy operating at a sample rate of 16 KSps (samples per second) or higher. In addition, because the raw input signal 310 is essentially recorded by system 300, that signal could potentially be reconstructed from stored data, raising privacy and security issues.
Furthermore, to mitigate the problem of high power consumption in battery powered applications, system 300 may be toggle between normal detection and standby operational modes at some duty cycle. For example, from time to time the whole system may be turned on and run in full-power mode for detection, followed by intervals in low-power standby mode. However, such duty cycled operation increases the possibility of missing an event during the standby mode.
By way of further background, U.S. Patent Application Publication No. US 2015/0066498, published Mar. 5, 2015, commonly assigned herewith and incorporated herein by this reference, describes a low power sound recognition sensor configured to receive an analog signal that may contain a signature sound. In this sensor, the received analog signal is evaluated using a detection portion of the analog section to determine when background noise on the analog signal is exceeded. A feature extraction portion of the analog section is triggered to extract sparse sound parameter information from the analog signal when the background noise is exceeded. An initial truncated portion of the sound parameter information is compared to a truncated sound parameter database stored locally with the sound recognition sensor to detect when there is a likelihood that the expected sound is being received in the analog signal. A trigger signal is generated to trigger classification logic when the likelihood that the expected sound is being received exceeds a threshold value.
By way of further background, U.S. Patent Application Publication No. US 2015/0066495, published Mar. 5, 2015, commonly assigned herewith and incorporated herein by this reference, describes a low power sound recognition sensor configured to receive an analog signal that may contain a signature sound. In this sensor, sparse sound parameter information is extracted from the analog signal and compared to a sound parameter reference stored locally with the sound recognition sensor to detect when the signature sound is received in the analog signal. A portion of the sparse sound parameter information is differential zero crossing (ZC) counts. Differential ZC rate may be determined by measuring a number of times the analog signal crosses a threshold value during each of a sequence of time frames to form a sequence of ZC counts and taking a difference between selected pairs of ZC counts to form a sequence of differential ZC counts.
Disclosed embodiments provide an audio recognition system and method that efficiently identifies particular audio events with reduced power consumption.
Disclosed embodiments provide such a system and method that identifies particular audio events with improved accuracy.
Disclosed embodiments provide such a system and method that enables increased hardware efficiency, particularly in connection with analog circuitry and functions.
Disclosed embodiments provide such a system and method that can perform such audio recognition with higher frequency band resolution without increasing detection channel complexity.
Disclosed embodiments provide such a system and method that reduces analog filter mismatch in the audio recognition system.
Other objects and advantages of the disclosed embodiments will be apparent to those of ordinary skill in the art having reference to the following specification together with its drawings.
According to certain embodiments, analog audio detection is performed on a received audio signal by dividing the signal duration into multiple intervals, for example into frames. Analog signal features are identified from signals filtered with different frequency characteristics at different times in the signal, thus identifying signal features at particular frequencies at particular points in time in the input signal. An output feature sequence is constructed from the identified analog signal features, and compared with pre-defined feature sequences for the detected events.
The one or more embodiments described in this specification are implemented into a voice recognition function, for example in a mobile telephone handset, as it is contemplated that such implementation is particularly advantageous in that context. However, it is also contemplated that concepts of this invention may be beneficially applied and implemented in other applications, for example in sound detection as may be carried out by remote sensors, security and other environmental sensors, and the like. Accordingly, it is to be understood that the following description is provided by way of example only, and is not intended to limit the true scope of this invention as claimed.
As will be described in further detail below in connection with these embodiments, AFE 10 also performs analog domain processing to extract particular features in the received input signal. These typically “sparse” extracted analog features are classified, for example by comparison with signature features stored in signature/imposter database 17, and then digitized and forwarded to digital microcontroller unit (MCU) 20, which may be realized by way of a general purpose microcontroller unit, specialty digital signal processor (DSP), application specific integrated circuit (ASIC), or the like. MCU 20 applies one or more type of known pattern recognition techniques, such as a Neural Network, a Classification Tree, Hidden Markov models, Conditional Random Fields, Support Vector Machine, and the like to carry out digital domain pattern recognition on the digitized features extracted by AFE 10 in this arrangement. Upon MCU 20 detecting a sound signature from those features, the corresponding information is forwarded from sound recognition system 5 to the appropriate destination function in the system in which system 5 is implemented, in the conventional manner. According to this arrangement, sound recognition system 5 only digitizes the extracted features, i.e. those features that contain useful and recognizable information, rather than the entire input signal, and performs digital pattern recognition based on those features, rather than a digitized version of the entire input signal. According to this arrangement, because the input sound is processed and framed in the analog domain, much of the noise and interference that may be present on a sound signal is removed prior to digitization, which in turn reduces the precision needed within AFE 10, particularly the speed and performance requirements for analog-to-digital conversion (ADC) functions within AFE 10. The resulting relaxation of performance requirements for AFE 10 enables sound recognition system 5 to operate at extremely low power levels, as is critical in modern battery-powered systems.
As shown in
The above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498 describe approaches to analog feature extraction in which multiple analog channels operate on the analog signal to extract different analog features. As described in those publications, one or more channels may extract such attributes as zero-crossing information and total energy from respective filtered versions of the analog input signal, using a selected band pass, low pass, high pass or other type of filter. The extracted features may be based on differential zero-crossing (ZC) counts, for example differences in ZC rate between adjacent sound frames (i.e., in the time-domain), determining ZC rate differences by using different threshold voltages instead of only one reference threshold (i.e., in the amplitude-domain); determining ZC rate difference by using different sampling clock frequencies (i.e., in the frequency-domain), with these and other differential ZC measures used individually or combined to recognize particular features. The total energy values extracted from the analog signal and various filtered versions of that signal can be analyzed to detect energy values in particular bands of frequencies, which can also indicate particular features.
According to the approaches in the above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498, the analog feature extraction channels are applied over the duration of the received signal.
It has been discovered, in connection with this invention, that signal features in a particular frequency band at a particular time interval within the signal can be more important to signature recognition than features in other frequency bands during that interval, and more important than features in that same particular frequency band at other times in the signal. According to these embodiments, time-dependent analog filtered feature extraction and sequencing function 35 (
It is contemplated that the particular sequence of filter frequency characteristics to be applied over the duration of the input signal will typically be determined by on-line training function 18 in its development of signature/imposter database 17. In general, this training will operate to identify the most unique features of the sound event to be detected, such as described in the above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498, with the addition of the necessary training to identify the particular frequency bands and frame intervals at which those features occur within the signal. According to these embodiments, this training results in the determination of a sequence of filter frequency bands and corresponding signal features to be applied or detected, as the case may be, over the duration of the signal.
An example of the operation of time-dependent analog filtered feature extraction and sequencing function 35 according to these embodiments is illustrated in
As noted above, the sequence of filter characteristics selected by time base controller 42 over the sequence of m frames can be pre-defined based on the result of on-line training function 18, or otherwise corresponding to the pre-known feature sequence in signature/imposter database 17 for the sound signature to be detected.
According to this embodiment, therefore, a sequence of framed filtered analog signals F(n), each filtered according to a filter characteristic that may vary among the frames of the sequence of m frames, is provided by tunable filter 40 to feature extraction function 45. Feature extraction function 45 is constructed to extract one or more features from the filtered signal in each frame. For example, as described in the above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498, feature extraction function 45 may be constructed to extract features such as ZC counts, ZC differentials, total energy, and the like. It is contemplated that those skilled in the art having reference to this specification along with the above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498 will be readily able to realize the zero-crossing circuitry, integrator circuitry, and the like for extracting the desired features from the signal F(n) produced by tunable filter 40 according to this embodiment, without undue experimentation. Feature extraction function 45 thus produces a frame by frame sequence E(F(n))/ZC(F(n)) of the extracted features, where those features are extracted from particular frequencies of the input signal at various times within the duration of the signal.
This sequence E(F(n))/ZC(F(n)) of extracted features is then provided to event trigger 36 in analog feature extraction function 28, as shown in
The filtered signals produced by analog filters 50a through 50k are then applied to corresponding feature extraction functions 55a, 55b, . . . , 55k, which are constructed to extract one or more features from the corresponding filtered signal. It is contemplated that feature extraction functions 55a through 55k may be constructed similarly as feature extraction function 45 described above and in the above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498, with each instance extracting features such as ZC counts, ZC differentials, total energy, and the like. It is contemplated that those skilled in the art having reference to this specification along with the above-incorporated U.S. Patent Application Publications No. US 2015/0066495 and No. US 2015/0066498 will be readily able to realize feature extraction functions 55a through 55k, in the form of zero-crossing circuitry, integrator circuitry, and the like, as appropriate for extracting the desired features from the filtered signals from corresponding analog filters 50a through 50k, without undue experimentation. It is contemplated that the filtered output from one or more of analog filters 50a through 50k may be presented to more than one corresponding feature extraction function 55a through 55k. For example, as shown in
According to this embodiment, in which the multiple analog filters 50a through 50k may each be enabled to filter input signal i(t) over its entire duration, the outputs of each of feature extraction functions 55a through 55k are applied to corresponding inputs of multiplexer 60. The output of multiplexer 60 presents the feature sequence E(F(n))/ZC(F(n)) to trigger logic 36 and ADC 29 (
As in the embodiment of
In this implementation, SP unit 1004 includes an A2I sound extraction module in the form of sound recognition system 5 described above, which allows mobile phone 1000 to operate in an ultralow power consumption mode while continuously monitoring for a spoken word command or other sounds that may be configured to wake up mobile phone 1000. Robust sound features may be extracted and provided to digital baseband module 1002 for use in classification and recognition of a vocabulary of command words that then invoke various operating features of mobile phone 1000. For example, voice dialing to contacts in an address book may be performed. Robust sound features may be sent to a cloud based training server via RF transceiver 1006, as described in more detail above.
RF transceiver 1006 is a digital radio processor and includes a receiver for receiving a stream of coded data frames from a cellular base station via antenna 1007 and a transmitter for transmitting a stream of coded data frames to the cellular base station via antenna 1007. RF transceiver 1006 is coupled to DBB 1002 which provides processing of the frames of encoded data being received and transmitted by cell phone 1000.
DBB unit 1002 may send or receive data to various devices connected to universal serial bus (USB) port 1026. DBB 1002 can be connected to subscriber identity module (SIM) card 1010 and stores and retrieves information used for making calls via the cellular system. DBB 1002 can also connected to memory 1012 that augments the onboard memory and is used for various processing needs. DBB 1002 can be connected to Bluetooth baseband unit 1030 for wireless connection to a microphone 1032a and headset 1032b for sending and receiving voice data. DBB 1002 can also be connected to display 1020 and can send information to it for interaction with a user of the mobile UE 1000 during a call process. Touch screen 1021 may be connected to DBB 1002 for haptic feedback. Display 1020 may also display pictures received from the network, from a local camera 1028, or from other sources such as USB 1026. DBB 1002 may also send a video stream to display 1020 that is received from various sources such as the cellular network via RF transceiver 1006 or camera 1028. DBB 1002 may also send a video stream to an external video display unit via encoder 1022 over composite output terminal 1024. Encoder unit 1022 can provide encoding according to PAL/SECAM/NTSC video standards. In some embodiments, audio codec 1009 receives an audio stream from FM Radio tuner 1008 and sends an audio stream to stereo headset 1016 and/or stereo speakers 1018. In other embodiments, there may be other sources of an audio stream, such a compact disc (CD) player, a solid state memory module, etc.
The analog filtered feature extraction and sequencing function according to this embodiment provides important benefits in the recognition of audio events, commands, and the like. One such benefit resulting from the analog feature extraction according to these embodiments is reduction in the complexity of the downstream digital sound recognition process. Rather than receiving and processing multiple analog feature sequences processed by multiple analog channels, these embodiments can present a single sequence of extracted features, which allows the digital classifier to be significantly less complex. These embodiments also improve the potential frequency band resolution of the sound recognition process over fixed frequency band implementations, in which the frequency band resolution is proportional to the channel count. In these embodiments, different frequency bands can be assigned to certain time intervals of the input signal, allowing a single channel to attain good resolution over multiple frequencies. This attribute of these embodiments also improves the overall accuracy and efficiency of the sound recognition process, by allowing the training process to extract the most unique features of the audio event to be detected, isolated in both time and frequency, which reduces the computational work for recognizing a signature while improving the accuracy of the recognition.
Some of the embodiments described above provide hardware efficiency and improved hardware performance. More specifically, the use of a tunable analog filter that applies different frequency characteristics at different times during the signal duration reduces the number of analog filters and also the number of feature extraction functions in the analog front end from the multi-channel approach. In addition, embodiments that use the tunable analog filter eliminate the potential for filter mismatch among multiple filters operating in parallel; rather, many of the same circuit elements are used to apply the multiple filter characteristics at different times.
It is contemplated that those skilled in the art having reference to this specification will recognize variations and alternatives to the described embodiments, and it is to be understood that such variations and alternatives are intended to fall within the scope of the claims. For example, while these embodiments perform the analog filtering and feature extraction after framing of the input analog signal, it is contemplated that framing could alternatively be performed after feature extraction and recognition. In addition, other embodiments may include other types of analog signal processing circuits that may be tailored to extraction of sound information that may be useful for detecting a particular type of sound, such as motor or engine operation, electric arc, car crashing, breaking sound, animal chewing power cables, rain, wind, etc. It is contemplated that those skilled in the art having reference to this specification can readily implement and realize such alternatives, without undue experimentation.
While one or more embodiments have been described in this specification, it is of course contemplated that modifications of, and alternatives to, these embodiments, such modifications and alternatives capable of obtaining one or more the advantages and benefits of this invention, will be apparent to those of ordinary skill in the art having reference to this specification and its drawings. It is contemplated that such modifications and alternatives are within the scope of this invention as subsequently claimed herein.