Patent Application
20130158996
1. Field
Embodiments of the present invention generally relate to speech recognition. More particular, embodiments of the present invention relate to the implementation of an acoustic modeling process on a dedicated processing unit.
2. Background
Real-time data pattern recognition is increasingly used to analyze data streams in electronic systems. On a vocabulary with over tens of thousands of words, speech recognition systems have achieved improved accuracy, making it an attractive feature for electronic systems. For example, speech recognition systems are increasingly common in consumer markets targeted to data pattern recognition applications such as, for example, the mobile device, server, automobile, and PC markets.
Despite the improved accuracy in speech recognition systems, significant computing resources are dedicated to the speech recognition process, in turn placing a significant load on computing systems such as, for example, multiuser/multiprogramming environments. Multiprogramming computing systems concurrently process data from various applications and, as a result, the load placed on these computing systems by the speech recognition process affects the speed at which the computing systems can process incoming voice signals as well as data from other applications. Further, for handheld devices that typically include limited memory resources (as compared to desktop computing systems), speech recognition applications not only place significant load on the handheld device's computing resources but also consume a significant portion of the handheld device's memory resources. The above speech recognition system issues of processing capability, speed, and memory resources are farther exacerbated by the need to process incoming voice signals in real-time or substantially close to real-time.
Therefore, there is a need to improve the load that speech recognition systems place on the processing capability, speed, and memory resources of computing systems.
An embodiment of the present invention includes a senone scoring unit (SSU). The SSU can include a SSU control module, a distance calculator, and an addition module. The SSU control module can be configured to receive a feature vector. The distance calculator can be configured to receive a plurality of Gaussian probability distributions via a data bus having a width of at least one Gaussian probability distribution (e.g., 768 bits) and the feature vector from the SSU control module. The distance calculator can include a plurality of arithmetic logic units (ALUs) and an accumulator. Each of the ALUs can be configured to receive a portion of the at least one Gaussian probability distribution and to calculate a dimension distance score between a dimension of the feature vector and a corresponding dimension of the at least one Gaussian probability distribution. The accumulator can be configured to sum the dimension distance scores from the plurality of ALUs to generate a Gaussian distance score. Further, the addition module can be configured to sum a plurality of Gaussian distance scores corresponding to the plurality of Gaussian probability distributions to generate a senone score. The SSU can also include a feature vector matrix module configured to store a scaling factor for the dimension of the feature vector.
Another embodiment of the present invention includes a method for acoustic modeling. The method can include the following: receiving a plurality of Gaussian probability distributions via a data bus having a width of at least one Gaussian probability distribution and a feature vector from an external computing device; calculating a plurality of dimension distance scores based on a plurality of dimensions of the feature vector and a corresponding plurality of dimensions of the at least one Gaussian probability distribution; summing the plurality of dimension distance scores to generate a Gaussian distance score for the at least one Gaussian probability distribution; and, summing a plurality of Gaussian distance scores corresponding to the plurality of Gaussian probability distributions to generate a senone score.
A further embodiment of the present invention includes a system for acoustic modeling. The system can include a memory module and a senone scoring unit (SSU). The memory module can be configured to interface with an external computing device to receive a feature vector. The SSU can include a distance calculator and an addition module, where the distance calculator includes a plurality of arithmetic logic units (ALUs) and an accumulator. Each of the ALUs can be configured to receive a portion of the at least one Gaussian probability distribution and to calculate a dimension distance score between a dimension of the feature vector and a corresponding dimension of the at least one Gaussian probability distribution. The accumulator can be configured to sum the dimension distance scores from the plurality of ALUs to generate a Gaussian distance score. Further, the addition module can be configured to sum a plurality of Gaussian distance scores corresponding to the plurality of Gaussian probability distributions to generate a senone score. The memory module and SSU can be integrated on the same chip.
Further features and advantages of embodiments of the invention, as well as the structure and operation of various embodiments of the present invention, are described in detail below with reference to the accompanying drawings. It is noted that the invention is not limited to the specific embodiments described herein. Such embodiments are presented herein for illustrative purposes only. Additional embodiments will be apparent to persons skilled in the relevant art based on the teachings contained herein.
The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate embodiments of the present invention and, together with the description, further serve to explain the principles of the invention and to enable a person skilled in the relevant art to make and use the invention.
The following detailed description refers to the accompanying drawings that illustrate exemplary embodiments consistent with this invention. Other embodiments are possible, and modifications can be made to the embodiments within the spirit and scope of the invention. Therefore, the detailed description is not meant to limit the scope of the invention. Rather, the scope of the invention is defined by the appended claims.
It would be apparent to a person skilled in the relevant art that the present invention, as described below, can be implemented in many different embodiments of software, hardware, firmware, and/or the entities illustrated in the figures. Thus, the operational behavior of embodiments of the present invention will be described with the understanding that modifications and variations of the embodiments are possible, given the level of detail presented herein.
This specification discloses one or more embodiments that incorporate the features of this invention. The disclosed embodiments merely exemplify the invention. The scope of the invention is not limited to the disclosed embodiments. The invention is defined by the claims appended hereto.
The embodiments described, and references in the specification to “one embodiment”, “an embodiment”, “an example embodiment”, etc., indicate that the embodiments described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is understood that it is within the knowledge of one skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.
In signal processing stage 110, an analog signal representation of an incoming voice signal 105 can be filtered to eliminate high frequency components of the signal that lie outside the range of frequencies that the human ear can hear. The filtered signal is then digitized using sampling and quantization techniques well known to a person skilled in the relevant art. One or more parametric digital representations (also referred to herein as “feature vectors 115”) can be extracted from the digitized waveform using techniques such as, for example, linear predictive coding and fast fourier transforms. This extraction can occur at regular time intervals, or frames, of approximately 10 ms, for example.
In acoustic modeling stage 120, feature vectors 115 from signal processing stage 110 are compared to one or more multivariate Gaussian probability distributions (also referred to herein as “Gaussian probability distributions”) stored in memory. The one or more Gaussian probability distributions stored in memory can be part of an acoustic library, in which the Gaussian probability distributions represent senones. A senone refers to a sub-phonetic unit for a language of interest, as would be understood by a person skilled in the relevant art. An individual senone can be made up of, for example, 8 components, in which each of the components can represent a 39-dimension Gaussian probability distribution.
Acoustic modeling stage 120 can process over 1000 senones, for example. As a result, the comparison of feature vectors 115 to the one or more Gaussian probability distributions can be a computationally-intensive task, as thousands of Gaussian probability distributions, for example, can be compared to feature vectors 115 every time interval or frame (e.g., 10 ms). A set of scores for each of the senones represented in the acoustic library (also referred to herein as “senone scores”) results from the comparison of each of feature vectors 115 to each of the one or more Gaussian probability distributions. Acoustic modeling stage 120 provides senone scores 125 to phoneme evaluation stage 130.
In phoneme evaluation stage 130, Hidden Markov Models (HMMs) can be used to characterize a phoneme as a set of states and an a priori set of transition probabilities between each of the states, where a state is associated with a senone. For a given observed sequence of senones, there is a most-likely sequence of states in a corresponding HMM. This corresponding HMM can be associated with an observed phoneme. A Viterbi algorithm can be used to find the likelihood of each HMM corresponding to a phoneme.
The Viterbi algorithm performs a computation that starts with a first frame and then proceeds to subsequent frames one-at-a-time in a time-synchronous manner A probability score is computed for each senone in the HMMs being considered. Therefore, a cumulative probability score can be successively computed for each of the possible senone sequences as the Viterbi algorithm analyzes sequential frames. Phoneme evaluation stage 130 provides the phoneme likelihoods or probabilities 135 (also referred to herein as a “phoneme score”) to word modeling stage 140.
In word modeling stage 140, searching techniques are used to determine a most-likely string of phonemes and subsequent words, over time. Searching techniques such as, for example, tree-based algorithms can be used to determine the most-likely string of phonemes.
Input device 210 is configured to receive an incoming voice signal (e.g., incoming voice signal 105 of
Processing unit is configured to process the digital input signal in accordance with the signal processing stage 110, acoustic modeling stage 120, phoneme evaluation stage 130, and word modeler stage 140 described above with respect to
In reference to
Phoneme evaluation module 330 receives senone scores 325 from acoustic modeling module 320. As discussed above with respect to speech recognition process 100 of
Word modeling module 340 uses searching techniques such as, for example, tree-based algorithms to determine a most-likely string of phonemes (e.g., most-likely phoneme 335), and subsequent words, over time.
An issue with conventional speech recognition system 300 of
Embodiments of the present invention address the issues discussed above with respect to conventional speech recognition systems 200 and 300 of
Although portions of the present disclosure is described in the context of a speech recognition system, a person skilled in the relevant art will recognize that the embodiments described herein are applicable to any data pattern recognition applications based on the description herein. These other data pattern recognition applications include, but are not limited to, image processing, audio processing, and handwriting recognition. These other data pattern recognition applications are within the spirit and scope of the embodiments disclosed herein.
In reference to the embodiment of
In another embodiment, acoustic modeling process 420 can compare the one or more feature vectors to all of the senones associated with an acoustic library. In this case, feedback 450 is not required, as phoneme evaluation process 430 receives an entire set of senone scores (e.g., “score all” function) from the APU for further processing.
A. System Bus Architectures for Speech Recognition Systems with an Acoustic Processing Unit
In an embodiment, the APU and CPU can be in communication with one another over a Serial Peripheral Interface (SPI) bus, a Peripheral Controller Interface (PCI) bus, an Application Programming Interface (API) bus, an Advanced Microcontroller Bus Architecture High-Performance Bus (AHB), an Advanced Peripheral Bus (APB), a memory bus, or any other type of bus. Example, non-limiting embodiments of system bus architectures for speech recognition process 400 of
As illustrated in
As illustrated in
As illustrated in
B. System-Level Architectures for Speech Recognition Systems with an Acoustic Processing Unit
Non-volatile memory device 830 can store an acoustic library to be used in a speech recognition process, in which the acoustic library can include over 1000 senones, according to an embodiment of the present invention. In an embodiment, when a senone request is received by speech recognition system 800, memory controller 820 copies the acoustic library from non-volatile memory device 830 to volatile memory device 840 via bus 825. The acoustic library transfer process between the non-volatile and volatile memory devices can be implemented using, for example, a direct memory access (DMA) operation.
In an embodiment, speech recognition system 800 can be powered on in anticipation of a senone scoring request. After power up, the acoustic library from non-volatile memory device 830 is immediately copied to volatile memory device 840. Once volatile memory device 840 has received the acoustic library, APU 810 is ready to begin processing senone scoring requests (e.g., acoustic modeling process 420 of
When the senone scoring request is received by APU 810, a selected senone from the acoustic library is copied from volatile memory device 840 to APU 810 via memory controller 820. APU 810 calculates a senone score based on the selected senone and a data stream received by APU 810 (e.g., one or more feature vectors 315 of
In an embodiment, after a predetermined time of inactivity (e.g., senone scoring inactivity by APU 810), volatile memory device 840 can be powered down. As a result, power efficiency in speech recognition system 800 can be improved, as a periodic refresh of memory cells in volatile memory device 840 will not be required. Here, the acoustic library is still stored in non-volatile memory device 830 such that the acoustic library can be retained when volatile memory device 840 is powered down. As would be understood by a person skilled in the art, when volatile memory device 840 is powered down, the contents stored therein (e.g., the acoustic library) will be lost. In an embodiment, when volatile memory device 840 is powered down, the other components of speech recognition system 800 can be powered down as well.
In step 910, a plurality of data patterns is copied from a non-volatile memory device (e.g., non-volatile memory device 830 of
In step 920, a data pattern from the volatile memory device is requested by a computational unit (e.g., APU 810 of
In step 930, after receiving the requested data pattern, the computational unit (e.g., APU 810 of
In step 940, the volatile memory device powers down. In an embodiment, the volatile memory device powers down after a predetermined time of inactivity (e.g., inactivity in the data pattern analysis by the computational unit). As a result, power efficiency can be improved, as a periodic refresh of memory cells in the volatile memory device will not be required. In an embodiment, when the volatile memory device is powered down, the other components of the system (e.g., other components of speech recognition system 800) can be powered down as well.
Non-volatile memory device 1140 can be configured to store an acoustic model that is copied to volatile memory device 1130 upon power-up of APU 1110, according to an embodiment of the present invention. In an embodiment, non-volatile memory device can be a Flash memory device and volatile memory device 1130 can be a DRAM device. Further, ASIC 1120 can be configured to perform an acoustic modeling process (e.g., acoustic modeling process 420 of
Memory device 1210 can be configured to store an acoustic library and to transfer one or more senones to ASIC 1230 via an I/O bus 1215, according to an embodiment of the present invention. In an embodiment, memory device 1210 can be a DRAM device or a Flash memory device. In another embodiment, the acoustic library can be stored in a memory device located within ASIC 1230 (not shown in
APU 1400 is in communication with a CPU (not shown in
In an embodiment, memory module 1420 and SSU 1430 can operate in two different clock domains. Memory module 1420 can operate at the clock frequency associated with the input clock signal to APU 1400 (e.g., from I/O signals 1410) and SSU 1430 can operate at a faster clock frequency based on the input clock signal, according to an embodiment of the present invention. For example, if the clock frequency associated with the input clock signal is 12 MHz, then SSU 1430 can operate at a clock-divided frequency of 60 MHz—five times faster than the clock frequency associated with the input clock signal. Techniques and methods for implementing clock dividers are known to a person skilled in the relevant art. As will be described in further detail below, the architecture of SSU 1430 can be based on the clock domain at which it operates.
In reference to
In another embodiment, memory device 1426 can be a volatile memory device such as, for example, a DRAM device. In an embodiment, when a senone request is received by APU 1400, memory controller 1424 can copy the acoustic library from a non-volatile memory device (either integrated on the same chip as APU 1400 or located external to APU 1400) to the volatile memory device. The acoustic library transfer process between the non-volatile and volatile memory devices can be implemented using, for example, a DMA operation.
Bus controller 1422 is configured to control data transfer between APU 1400 and an external CPU. In an embodiment, bus controller 1422 can control the receipt of feature vectors from the CPU and the transmission of senone scores from APU 1400 to the CPU. In an embodiment, bus controller 1422 is configured to transfer one or more feature vectors from the CPU to bridge controller 1428, which serves as an interface between memory module 1420 and SSU 1430. In turn, bridge controller 1428 transfers the one or more feature vectors to SSU 1430 for further processing. Upon calculation of a senone score, the senone score is transferred from SSU 1430 to memory module 1420 via bridge controller 1428, according to an embodiment of the present invention.
In an embodiment, bus controller 1422 can receive a control signal (via I/O signals 1410) that provides an active senone list. In an embodiment, the active senone list can be transferred to APU 1400 as a result of the phoneme evaluation process performed by the CPU (e.g., phoneme evaluation process 430 of
The active senone list indicates one or more senones stored in memory device 1426 to be used in a senone score calculation. In an embodiment, the active senone list can include a base address associated with an address space of memory device 1426 and a list of indices related to the base address at which the one or more senones are located in memory device 1426. Bus controller 1422 can send the active senone list to SSU 1430 via bridge controller 1428, in which SSU 1430 is in communication with memory device 1426 (via memory controller 1424) to access the one or more senones associated with the active senone list.
In another embodiment, bus controller 1422 can receive a control signal (via I/O signals 1410) that instructs APU 1400 to perform the senone score calculation using all of the senones contained in the acoustic library (e.g., “score all” function). Bus controller 1422 sends the “score all” instruction to SSU 1430 via bridge controller 1428, in which SSU 1430 is in communications with memory device 1426 (via memory controller 1424) to access all of the senones associated with the acoustic library.
Conventional speech recognition systems typically incorporate a feedback loop between acoustic modeling and phoneme evaluation modules (e.g., acoustic modeling module 320 and phoneme evaluation module 330 of
The “score all” function of APU 1400 not only alleviates the load on the CPU and the bandwidth of the data bus, but also provides processing of incoming voice signals in real-time or substantially close to real time. As discussed in further detail below, features of APU 1400 such as, for example, the bus width of data bus 1427 and the architecture of distance calculator 1436 of
In reference to
Further, each of the one or more senones stored in memory device 1426 is composed of one or more Gaussian probability distributions, where each of the one or more Gaussian probability distributions has the same number of dimensions as each of the one or more feature vectors (e.g., N dimensions). Each of the one or more senones stored in memory device 1426 can have, for example, 32 Gaussian probability distributions.
As discussed above, memory module 1420 and SSU 1430 can operate in two different clock domains. In an embodiment, SSU control module 1434 is configured to receive a clock signal from memory module 1420 via bridge controller 1428. The frequency of the clock signal received by SSU control module 1434 can be the same or substantially the same as the clock frequency associated with the input clock signal to APU 1400 (e.g., input clock signal from I/O signals 1410), according to an embodiment of the present invention.
In an embodiment, SSU control module 1434 can divide the frequency of its incoming clock signal and distribute that divided clock signal to other components of SSU 1430—e.g., output buffer 1432, feature vector matrix module 1435, distance calculator 1436, and addition module 1438—such that these other components operate at the clock-divided frequency. For example, if the clock frequency associated with the input clock signal (e.g., from I/O signals 1410) is 12 MHz, then SSU control module 1434 can receive the same or substantially the same clock signal from bridge controller 1428 and divide that clock frequency using known clock-dividing techniques and methods to a frequency of, for example, 60 MHz. SSU control module 1434 can distribute this clock-divided signal to the other components of SSU 1430 such that these other components operate at, for example, 60 MHz—five times faster than the clock frequency associated with the input clock signal.
For simplicity purposes, the clock signals distributed from SSU control module 1434 to the other components of SSU 1430 are not illustrated in
In reference to
For example, the active senone list can contain a base address associated with an address space of memory device 1426 and 100 indices pointing to 100 senones stored in memory device 1426. As would be understood by a person skilled in the relevant art, the indices can refer to pointers or memory address offsets in reference to the base address associated with the address space of memory device 1426. Further, as discussed above, a senone can be composed of one or more Gaussian probability distributions, where each of the one or more Gaussian probability distributions has the same number of dimensions as each of one or more feature vectors (e.g., N dimensions) received by APU 1400. For explanation purposes, this example will assume that each senone stored in memory device 1426 is composed of 32 Gaussian probability distributions. Based on the description herein, a person skilled in the relevant art will understand that each of the senones can be composed of more or less than 32 Gaussian probability distributions.
In an embodiment, for the first senone in the active senone list, control unit 1520 communicates with memory controller 1424 of
In an embodiment, the second Gaussian probability distribution can include memory address information for a third Gaussian probability distribution to be accessed by memory device 1426. Memory device 1426 can communicate this memory address information to control unit 1520 of
Control unit 1520 of
After the senones in the active senone list are used in the senone score calculations for a current feature vector, memory module 1420 can receive a control signal via I/O signals 1410 that indicates that the active senone list from the current feature vector is to be used in senone score calculations for a subsequent feature vector, according to an embodiment of the present invention. Upon receipt of the control signal from memory module 1420 via bridge controller 1428, SSU control module 1434 uses the same active senone list from the current feature vector in the senone score calculations for the subsequent feature vector. In particular, control unit 1520 of
In another embodiment, memory module 1420 can receive a control signal via I/O signals 1410 that indicates a “score all” operation. As discussed above, the “score all” function refers to an operation where a feature vector is compared to all of the senones contained in an acoustic library stored in memory device 1426. In an embodiment, control unit 1520 of
In an embodiment, similar to the active senone list example above, the second Gaussian probability distribution can include memory address information on a third Gaussian probability distribution to be accessed by memory device 1426. Memory device 1426 can communicate this memory address information to control unit 1520 of
Control unit 1520 of
In reference to
Along with mean and variance values stored for each senone in memory device 1426, an index can also be stored for each senone, in which the index points to a row in the FVTM, according to an embodiment of the present invention. The number of rows in the FVTM can vary (e.g., 10, 50, or 100 rows) and can be specific to a voice recognition system implementing APU 1400. Each row in the FVTM can have an equal number of entries as the N number of dimensions for a feature vector (e.g., 39), where each of the entries is a scaling factor that is multiplied to its corresponding feature vector dimension to produce a new feature vector, according to an embodiment of the present invention. The selected row from the FVTM (e.g., row of 39 scaling factors) is transferred to distance calculator 1436 via data bus 1439, in which distance calculator 1436 performs the multiplication operation to generate the new feature vector, as will be described in further detail below.
In an embodiment, SSU control module 1434 provides a feature vector received from the CPU and an index associated with a senone to feature vector matrix module 1435. The index indicates a particular row in the FVTM for scaling the feature vector. For example, the FVTM can have 100 rows and the index can be equal to 10. Here, for a feature vector with 39 dimensions, the 10th row of the FVTM contains 39 scaling factors, in which the row of scaling factors is transferred to distance calculator 1436 to generate the new feature vector.
In reference to
Datapath MUX 1610 is configured to receive a Gaussian probability distribution from memory device 1426 of
Further, in an embodiment, memory device 1426 and distance calculator 1436 can be integrated on the same chip, where data bus 1427 is a wide bus (of the width discussed above) integrated on the chip to provide data transfer of the Gaussian probability distribution from memory device 1426 to distance calculator 1436. In another embodiment, memory device 1426 and distance calculator 1436 can be integrated on two separate chips, where data bus 1427 is a wide bus (of the width discussed above) that is tightly coupled between the two chips such that degradation of data due to noise and interconnect parasitic effects are minimized. As will be discussed below, a benefit of a wide data bus 1427 (of the width discussed above), among others, is to increase performance of APU 1400 in the calculation of senone scores.
Datapath MUX 1610 is also configured to receive one or more control signals and a feature vector from SSU control module 1434 via data bus 1437, as well as feature vector scaling factors from feature vector buffer 1620. In an embodiment, feature vector buffer 1620 can be configured to store scaling factors (associated with a selected row of the FVTM) transferred from feature vector matrix module 1435 via data bus 1439. In another embodiment, feature vector buffer 1620 can be configured to store the FVTM. Here, one or more control signals from SSU control module 1434 via data bus 1437 can be used to select the FVTM row. Datapath MUX 1610 outputs the feature vector, selected feature vector scaling factors from the FVTM, and Gaussian probability distribution information to ALUs 16301-16308 via data bus 1612 for further processing.
In an embodiment, datapath MUX 1610 is also configured to receive a Gaussian weighting factor from the one or more controls signals from SSU control module 1434 via data bus 1437. Datapath MUX 1610 is configured to output the Gaussian weighting factor to accumulator 1640 for further processing.
In reference to
In an embodiment, datapath MUX 1610 is configured to distribute feature vector information associated with one dimension, a mean value associated with a corresponding dimension of a Gaussian probability distribution, a variance value associated with the corresponding dimension of the Gaussian probability, and feature vector scaling factors to each of ALU 16301-16308. Based on the feature vector information and the feature vector scaling factors allocated to a respective ALU, each of ALUs 16301-16308 is configured to generate a new feature vector by multiplying dimensions of the feature vector by respective scaling factors.
In an embodiment, the multiplication of the feature vector dimensions by the corresponding scaling factors is performed “on-the-fly,” meaning that the multiplication operation is performed during the calculation of the distance score. This is, in contrast, to the multiplication operation being performed for each of the rows in a FVTM and the results of the multiplication operation being stored in memory to be later accessed by each of ALUs 16301-16308. A benefit of the “on-the-fly” multiplication operation, among others, is that memory storage is not required for the results of the multiplication operation associated with non-indexed (or non-selected) rows of the FVTM. This, in turn, results in a faster generation of the new feature vector since additional clock cycles are not required to store the feature vector scaling results associated with the non-indexed rows in memory and also results in a smaller die size area for ALUs 16301-16308.
Based on the new feature vector, the mean value, and the variance value for a respective ALU, each of ALUs 16301-16308 is configured to calculate a distance score based on a feature vector dimension and a corresponding Gaussian probability distribution dimension per SSU clock cycle, according to an embodiment of the present invention. Cumulatively, in one clock cycle, ALUs 16301-16308 generate distance scores for 8 dimensions (i.e., 1 dimension calculation per ALU). The architecture and operation of the ALU is described in further detail below.
The number of ALUs in distance calculator 1436 can be dependent on the SSU clock frequency and the memory module clock frequency discussed above such that distance calculator 1436 outputs a distance score for one Gaussian probability distribution for every read access to memory device 1426, according to an embodiment of the present invention. For example, the memory module clock frequency can have an operating frequency of 12 MHz, where memory device 1426 also operates at 12 MHz (e.g., for a read access of approximately 83 ns). SSU 1430 can have an SSU clock frequency of, for example, 60 MHz to operate five times faster than the memory module cock frequency. With a feature vector of 39 dimensions and 8 ALUs, a Gaussian distance score for one Gaussian probability distribution can be calculated in 5 SSU clock cycles or 1 memory module clock cycle. Therefore, by design, the 5 SSU clock cycles is a predetermined number of clock cycles that corresponds to 1 memory module clock cycle, where as one Gaussian probability distribution is read from memory device at 1 memory module clock cycle, a Gaussian distance score for another Gaussian probability distribution is calculated by accumulator 1640.
In an embodiment, a portion of ALUs 16301-16308 can be activated on a rising edge of an SSU clock cycle, while the remaining portion of ALUs 16301-16308 can be activated on a falling edge of the SSU clock cycle. For example, ALUs 16301-16304 can be activated on the rising edge of the SSU clock cycle and ALUs 16305-16308 can be activated on the falling edge of the SSU clock cycle. As a result of staggering the activation of ALUs 16301-16308, the peak current (and peak power) generated by distance calculator 1436 can be minimized, thus decreasing the susceptibility of reliability issues in distance calculator 1436.
Based on the description herein, a person skilled in the relevant art will recognize that the architecture of distance calculator 1436 is not limited to the above example. Rather, as would be understood by a person skilled in the relevant art, distance calculator 1436 can operate at a faster or slower clock frequency of 60 MHz and that distance calculator 1436 can include more or less than 8 ALUs.
In reference to
After the Gaussian distance scores associated with all of the Gaussian probability distribution dimensions are accumulated in accumulator 1640 (e.g., 39 dimensions), accumulator 1640 multiplies the total sum by the Gaussian weighting factor to generate a weighted Gaussian distance score. In an embodiment, the Gaussian weighting factor is optional, where accumulator 1640 outputs the Gaussian distance score. In another embodiment, the Gaussian weighting factor is specific to each Gaussian and is stored in memory device 1426.
Addition module 1438 is configured to add one or more Gaussian distance scores (or weighted Gaussian distance scores) to generate a senone score. As discussed above, each senone can be composed of one or more Gaussian probability distributions, in which each Gaussian probability distribution can be associated with a Gaussian distance score. For a senone with a plurality of Gaussian probability distributions (e.g., 32 Gaussian probability distributions), addition module 1438 sums the Gaussian distance scores associated with all of the Gaussian probability distributions to generate the senone score. In an embodiment, addition module 1438 is configured to perform the summation operation in the log domain to generate the senone score.
Output buffer 1432 is configured to receive a senone score from addition module 1438 and transfer the senone score to bridge controller 1428. Bridge controller 1428, in turn, transfers the senone score to the external CPU via bus controller 1422. In an embodiment, output buffer 1432 can include a plurality of memory buffers such that, as a first senone score in a first memory buffer is being transferred to bridge controller 1428, a second senone score generated by addition module 1438 can be transferred to a second memory buffer for a subsequent transfer to bridge controller 1428.
In step 1710, a plurality of Gaussian probability distributions is received via a data bus having a width of at least one Gaussian probability distribution and a feature vector from an external computing device. The Gaussian probability distribution can be composed of, for example, 768 bits, where the width of the data bus is at least 768 bits. Further, APU 1400 of
In an embodiment, information associated with a plurality of dimensions of the feature vector, a plurality of mean values associated with the corresponding plurality of dimensions of the at least one Gaussian probability distribution, and a plurality of variance values associated with the corresponding plurality of dimensions of the at least one Gaussian probability distribution are distributed to, for example, arithmetic logic units (e.g., ALUs 16301-16308 of
In step 1720, a plurality of dimension distance scores is calculated based on a plurality of dimensions of the feature vector and a corresponding plurality of dimensions of the at least one Gaussian probability distribution. In an embodiment, the distance score calculations are based on at least one senone from an active senone list. The active senone list can include a base address associated with an address space of a memory device and one or more indices related to the base address at which the at least one senone is located in the memory device. Further, a plurality of scaling factors for the plurality of dimensions of the feature vector are stored, where the plurality of scaling factors are applied to the plurality of dimensions of the feature vector during the calculation of the plurality of dimension distance scores. Step 1720 can be performed by, for example, distance calculator 1436 of
In step 1730, the plurality of dimension distance scores are summed to generate a Gaussian distance score for the at least one Gaussian probability distribution. In an embodiment, the Gaussian distance score is generated over a predetermined number of senone scoring unit (SSU) clock cycles. The predetermined number of SSU clock cycles can equate to a read access time of the at least one Gaussian probability distribution from a memory device. Step 1730 can be performed by, for example, distance calculator 1436 of
In step 1740, a plurality of Gaussian distance scores corresponding to the plurality of Gaussian probability distributions is summed to generate a senone score. Step 1740 can be performed by, for example, distance calculator 1436 of
Embodiments of the present invention address and solve the issues discussed above with respect to conventional speech recognition system 200 of
where:
Δij=xi−μij,
varij, is the variance value of the ith dimension of the jth Gaussian probability distribution vector;
M1 and M2 are scaling factors;
C is a constant;
xi is the value of the feature vector in the ith dimension; and
μij is the mean value of the ith dimension of the jth Gaussian probability distribution vector.
Thus, in an embodiment, for a given dimension and a given Gaussian probability distribution, the one-dimensional distance score output by ALU 1800 is dependent on three variables: xi, μij, and varij. One technique for implementing this equation in software is to generate a look up table (LUT) that is indexed with these three variables. Moreover, because the score does not specifically depend on the values of xi and μij, but rather the difference between them, Δij, this LUT can be further simplified into a dimensional LUT indexed by the Δij and varij. Thus, a two-dimensional LUT could be used to implement ALUs 16301-16308.
A two-dimensional LUT, however, could have substantial drawbacks if used to implement ALUs 16301-16308 in the hardware implementation of
In an embodiment, ALU 1800 overcomes this drawback of two-dimensional LUTs by implementing a scoring function using a combination of computational logic and a one-dimensional LUT. Importantly, Equation (1) can be split into two parts: an aluij part and a LUTij part, with each specified below.
Thus, ALU 1800 computes aluij and, in parallel with the computing, retrieves LUTij. The aluij and LUTij are then combined to form the distance score. In particular, as shown in
Computational logic unit 1802 and LUT module 1804 only receive the inputs that are needed to determine their respective value. Specifically, as described above, aluij depends on three variables: xi, μij, and varij. Thus, as shown in
In an embodiment, LUT 1914 outputs a value that corresponds to
Multiplier 1916 computes a product of two terms: (1) the value retrieved from LUT 1914 and (2) the square output by squaring module 1912. Thus, the output of multiplier 1916 is
This product value is received by formatting module 1918, which formats the result so that it can be effectively combined with the output of LUT module 1804.
As shown in
The outputs from computational unit 1802 and LUT module 1804 are received at combination module 1806. Combination module 1806 includes an adder 1930, a shift module 1932, a rounding module 1934, and a saturation module 1936. Adder 1930 computes the sum of the two received values and outputs the sum. Shift module 1932 is configured to remove the fractional portion of the sum output by adder 1930. Rounding module 1934 is configured to round down the output of shift module 1934. Saturation module 1936 is configured to receive the rounded sum and saturate the value to a specific number of bits. Thus, the output of saturation module 1936 is a value having a specific number of bits that represents the one-dimensional distance score.
Transform module 2002 includes a multiplier 2020, a scale bit module 2022, and a saturation module 2024. As described above, values of feature vector can be transformed by respective entries in a feature vector transform matrix to, for example, account for learned characteristics of a speaker. In an embodiment, transform module 2002 can be configured to scale individual feature vector values xi by corresponding transform values αi. Specifically, multiplier 2020 computes a product of the feature vector value xi and the corresponding transform value αi and outputs a value to scale bit module 2022. Scale bit module 2022 shifts to the right and outputs the resulting integer to saturation module 2024. Saturation module 2024 is similar to saturation module 1936, described with reference to
Exception handling module 2012 and multiplexer 2018 are configured to address specific errors present in LUT 1914. For example, in an effort to save space, the size of LUT 1914 can be reduced. This reduction in size can cause specific values of LUT 1914 to have an error. In such an embodiment, exception handling module 2012 can recognize if the output of LUT 1914 will be one of those values, and output the correct value. Put another way, exception handling module 2012 can act as a LUT that includes an entry for each entry of LUT 1914 that may have an error due to size restrictions. Because LUT 1914 is indexed based on varij, exception handling module 2012 can recognize whether the output of LUT 1914 needs to be corrected based on the value of varij.
In a further embodiment, exception handling module 2012 can act as a two-dimensional LUT that also receives Δij. In such an embodiment, exception handling module 2012 can output specific values of aluij (e.g., as opposed to the corresponding entry from LUT 1914). Because the number of these possible errors in LUT 1914 is relatively small, exception handling module 2012 does not occupy a significant amount of space, as would other, larger two-dimensional LUTs. Furthermore, by controlling multiplexer 2018 to output the output of exception handling module 2012 instead of the output of sign bit module 1918, exception handling module 2012 can ensure that the stored value for aluij rather than the value of aluij calculated using the incorrect output of LUT 1914 is finally output to combination module 1806.
Formatting module 2014 receives the product computed by multiplier 1916. In an embodiment, formatting module 2014 is configured to reduce the number of bits in the result. While not necessary, this operation can save space and power by reducing the number of bits on the output.
Moreover, the embodiment of
A. System Overview
APU 2220 includes an acoustic model memory 2222, a first bus 2224, a memory buffer 2226, a second bus 2228, and a senone scoring unit 2230. Acoustic model memory 2222 can be configured to store a plurality of senones that together form one or more acoustic models. First bus 2224 is a wide bus that is configured to allow acoustic model memory to output an entire Gaussian probability distribution vector to memory buffer 2226. Senone scoring unit 2230 scores a senone score against a feature vector received from CPU 2210. Senone scoring unit 2230 can be implemented as described above. For example, senone scoring unit can be implemented as shown in
Memory buffer 2226 can hold a Gaussian probability distribution vector until senone scoring unit 2230 is ready to compute a Gaussian distance score for it. That is, if senone scoring unit 2230 is scoring a feature vector received from CPU 2210 against a Gaussian probability distribution vector q, memory buffer 2226 can hold the next Gaussian probability distribution vector to be scored, i.e., vector q+1.
As shown in
As noted above, the senone score is computed by calculating the scores of the J Gaussian probability distribution vectors of dimension N, and by then summing them together to get the total score. Some scoring algorithms, however, use only the most significant Gaussians in the calculation to increase the speed of the computation. When utilizing algorithms based on a partial set of Gaussians, only those bits associated with the required Gaussians need to be transferred from the acoustic model memory to senone scoring unit 2230. In other words, the largest number of contiguous bits in memory which will always be required by senone scoring unit 2230 is equal to the number of bits used to store a single Gaussian probability distribution vector. The bandwidth requirements of the memory bus as well as the number of bits that need to be read in parallel with be minimized by transferring only those bits comprising a single Gaussian probability distribution vector in each transfer. Using this number of bits per transfer, the power requirements of APU 2220 can be reduced and the transfer rate of the necessary data to senone scoring unit 2230 will be increased, resulting in an improvement of the overall system performance. Put another way, by reducing the number of bits per transfer, the power requirements of APU 2220 can be reduced and the transfer rate of the necessary data to senone scoring unit 2230 can also be increased, resulting in an improvement of the overall system performance.
As discussed above, acoustic modeling is one of the major bottlenecks in many types of speech recognition system (i.e., keyword recognition, or large vocabulary continuous speech recognition). Because of the large number of comparisons and calculations, high performance and/or parallel microprocessors are commonly used, and a high bandwidth bus between the memory storing the acoustic models and the processors is required. In the embodiment of
The number of bits per transfer can also a function of the algorithms used for acoustic modeling. When scoring algorithms based on a partial set of Gaussians are used (i.e. Gaussian Selection) then the number of bits per transfer can be equal to the size of the Gaussian used by the algorithm. Fewer number of bits per transfer requires multiple cycles to transfer the data comprising the Gaussian, while greater numbers of bits per transfer is inefficient due to data non-locality.
In an embodiment, an architecture is used for acoustic modeling hardware accelerators when scoring algorithms are used is at least partially based on a partial set of Gaussians (i.e., Gaussian Selection). This optimized architecture can result in a significant improvement in the overall system performance compared to other architectures.
In an embodiment, dedicated DRAM module 2302 is dedicated to senone scoring unit 2320 to, for example, store senones. Thus, memory interface 2316 can couple senone scoring unit 2320 to dedicated DRAM 2302.
SPI interface module 2412 can provide an interface to an SPI bus, which, in turn, can couple hardware accelerator 2400 to a CPU. Memory interface 2416 couples senone scoring unit 2414 to dedicated DRAM module 2404. In an embodiment, a voice-recognition system can be implemented in a cloud-based solution in which the senone scoring and processing necessary for voice-recognition is performed in a cloud-based voice-recognition application.
B. Software Stack
In software stack 2500, application 2502 communicates with voice recognition engine 2504, which in turn, communicates with Generic DCA 2506. In an embodiment, voice recognition engine 2504 is coupled to the Generic DCA 2506 via a DCA API. Generic DCA 2506 can be coupled to LLD 2508 via a LLD API. LLD 2508 can be coupled to HAL 2510 via an HAL API. HAL 2510 is communicatively coupled to SPI Bus Controller 2512 which is communicatively coupled to SPI bus 2514. APU 2516 is communicatively coupled to SPI bus 2514 and is communicatively coupled to the HAL 2510 via bus controller 2512 and SPI bus 2514.
In an embodiment, software stack 2500 provides a software interface between APU 2516 and application 2502 (e.g., an application that employs voice recognition). In particular, application 2502 and voice recognition engine 2504 can be “hardware agnostic.” That is, the application 2502 and voice recognition engine 2504 can complete their respective operations without detailed knowledge about how the distance, or senone, scoring is taking place.
Generic DCA 2506, LLD layer 2508, and HAL layer 2510 include hardware-specific API calls. In an embodiment, the API calls of HAL 2510 depend on the type of controller to which it is connected. In an embodiment, the bus interface for APU 2516 can be a different bus and controller combination, requiring a different HAL (with different API calls).
Generic DCA 2506 is a distance computational API. The DCA can be defined by a software developer. In an embodiment, the DCA API is specifically defined to support a voice recognition engine, such as voice recognition engine 2504. Also, Generic DCA 2506 can be implemented specifically for APU 2516. Moreover, LLD 2508 can be a functional abstraction of the senone scoring unit commands and can be a one-to-one mapping to the senone scoring unit commands. As shown in
The DCA API can include the following five functions: Create, Close, Set Feature, Compute Distance Score, and Fill Scores. In an embodiment, the Create function specifies which acoustic model is to be used. There can be one or more acoustic models stored in memory (e.g., one or more acoustic models for each language). For example, as discussed above with reference to
In an embodiment, the Set Feature function is used to set the senone scoring requests into their respective frames by passing a specific framelD, a passID, and the feature vector. As noted above, the input audio signal can be broken up into frames (e.g., by voice recognition engine 2504). An exemplary frame comprises spectral characteristics of a portion of the audio input signal. In an embodiment, a frame can be 12 milliseconds (ms) long. The Set Feature function can convert each frame into 39 dimensions (e.g., 39 8-bit values). The Set Feature function can specify a particular frame's ID and the associated feature vector.
In an embodiment, the Distance Compute Score function calculates the senone score (e.g., Gaussian probability), which, as noted above, can be implemented as a distance calculation. This function can be used to begin and prepare the senone scoring. For example, the feature vector can be input into APU 2516 and APU 2516 will score against all the senones stored in the acoustic model, or at least a selected portion of the senones. This score will then be given back to the upper layer. In an embodiment, the Distance Compute Score function can specify that a portion or the complete acoustic model will be used for the senone scoring.
In an embodiment, the Fill Scores function takes the senone scoring result and returns it to the upper software layers, including application 2502 and voice recognition engine 2504.
In an embodiment, voice recognition engine 2504 can be used for any form of pattern recognition, e.g., pattern recognition forms that use a Hidden Markov model for pattern recognition. In another embodiment, another form of pattern recognition also uses Gaussian calculations. Examples of pattern recognition can include, but are not limited to the above described senone scoring for speech recognition, image processing and handwritten recognition.
As noted above, application 2502 and voice recognition engine 2504 are agnostic to any hardware used to determine the senone score. In an embodiment, a particular APU can be swapped out for different hardware without application 2502 and voice recognition engine 2504 knowing or being effected. When application 2502 and voice recognition engine 2504 are agnostic to any type of hardware used for the senone scoring, a first hardware accelerator can be replaced with a second hardware accelerator of a different design without requiring any redesign of application 2502 and voice recognition engine 2504. In other words, as discussed herein, while the APU Library of calls are specific to the type and design of hardware accelerator used, the Generic DCA Library calls are not hardware specific.
In an embodiment, a software architecture, as illustrated in
In an embodiment, APU 2516 computes senone scores using the Gaussian Mixture Model. APU 2516 can compute these scores much faster (e.g., by an order of magnitude) than an embedded processor (e.g., a cortex A8 embedded processor) making speech recognition more practical in on-board speech recognition systems with APU 2516. Offloading the senone scoring (or distance computation) to APU 2516 not only improves the user experience (by reducing the computational latency) but also allows CPU 2210 to attend to other tasks in the system. The software architecture plays an important role in reducing the CPU load and the latency.
In an embodiment, voice recognition engine 2504 is not directly aware of APU 2516. For example, voice recognition engine 2504 can use Generic DCA API 2506 to compute the distances (also referred to as senone scores). The specific implementation of the Generic DCA library discussed here has been designed specifically to use APU 2516, with a plurality of function calls to the APU discussed below. This differs from a fully software implementation of the Generic DCA library. This specific implementation translates the Generic DCA library calls to a sequence of APU library calls. The details of the implementation are described below. The definition and implementation of the APU library is specific to the current implementation of the APU and is also described Below.
In an embodiment, Generic DCA 2506 operates as an interface layer between the voice recognition engine 2504 and APU 2516. For example, voice recognition engine 2504 can utilize generic API calls to the Generic DCA to request senone scoring. Generic DCA 2506 then utilizes an APU-specific library of API calls, described further below, to direct the APU hardware accelerator to perform the requested senone scoring. Because voice recognition engine 2504 is not aware of APU 2516, voice recognition engine 2504 can take advantage of the following benefits. For example, voice recognition engine 2504 may only need to know the message passing formats of APU 2516. Voice recognition engine 2504 also does not need to know the tasks to be performed by APU 2516. Moreover, there is a swap-out benefit. That is, APU 2516 can be replaced or redesigned without requiring any redesign of voice recognition engine 2504. Only the interface, in this embodiment Generic DCA 2506, needs to have the hardware specific API calls to ensure the required interoperability between voice recognition engine 2504 and APU 2516.
In one exemplary embodiment, a Generic DCA Library comprises the following list of functions:
Function name: distance_computation_create
description: stores these parameters as part of the state of distance computation.
Function name: distance_computation_setfeature
Description: store the feature vector corresponding to the frame Id.
Function name: distance_computation_computescores
Input parameters:
Description: specifies the senones to be scored for a given frame.
Function name: distance_computation_fillscores
Description: store the senone scores in the buffer.
Function name: distance_computation_setfeaturematrix
Description: stores the feature vector transformation matrix given by “pMatrix” in APU.
The distance_computation_setfeaturematrix function is called between utterances to adapt the recognition to the specific speaker. The APU uses this matrix when computing the senone scores for the next utterance.
In an embodiment, “distance_computation_computescores” and “distance_computation_fillscores” can be implemented such that the computational latency and the CPU load are minimized. For example, these functions can be implemented so as to achieve the concurrent operation embodied in
In one exemplary embodiment, an APU Library supports the following functions:
Function name: apu_set_acoustic_model
Function name: apu_load_feature_vector
Function name: apu_score_senone_chunk
Function name: apu_score_range
Function name: apu_read_senone_scores
Function name: apu_check_score_ready_status
Function name: apu_read_score_length
Function name: apu_read_status
Function name: apu_read_configuration
Function name: apu_write_configuration
In an embodiment, the APU can be used for scoring the senones for each frame of a given utterance. The acoustic model of choice is communicated to the APU at the beginning as part of the function distance_computation_create. The feature vector for a given frame is passed to the APU via the function distance_computation_setfeature. The senones to be scored for a given frame are passed to the APU via the function distance_computation_computescores. The actual scores computed by the APU can be passed back to the Voice Recognition Engine via the function distance_computation_fillscores.
The control flows from top to bottom of stack 2500 illustrated in
C. Concurrent Search and Distance Score Computation
In one embodiment, a thread (e.g. an executable process) separate from a thread that is being executed by application 2502 or voice recognition engine 2504 can be created for APU 2516. For there to be separate threads, there must be no dependency (that a further action of a first actor is dependent upon the actions of a second actor). Breaking any dependency between application 2502 and voice recognition engine 2504 and APU 2516 allows application 2502 and voice recognition engine 2504 to operate in parallel with APU 2516. In one exemplary embodiment, a dependency between application 2502 and voice recognition engine 2504 on one hand and APU 2516 on the other can be avoided through the use of frames, e.g., lasting approximately 10-12 ms (although the invention is not limited to this embodiment). For example, while the application 2502 is using the senone score for frame n, APU 2516 can be performing a senone score for frame n+1.
More specifically, a voice recognition operation requires two discrete operations: scoring and searching. As described above, the scoring operation involves a comparison between Gaussian probability distribution vectors of a senone with the feature vector corresponding to a specific frame. In an embodiment, software stack 2500 can be configured such that these two operations occur in parallel. In particular, as shown in
For any given frame, the search can follow the distance computation as illustrated in
In step 2702, the received audio signal is divided into frames. For example, in
In step 2704, a search thread and a distance computation thread are created. For example, in
In step 2706, a distance score is computed using an APU. For example, in
In step 2708, a search operation is performed using the computed score for the frame. For example, in
In step 2710, it is determined whether the frame was the last frame of the audio signal. If so, method 2700 ends. If not, method 2700 proceeds to step 2712.
In step 2712, concurrently with the search operation of step 2708 a distance score for the next frame is computing using the APU. For example, in
Various aspects of the present invention may be implemented in software, firmware, hardware, or a combination thereof,
It should be noted that the simulation, synthesis and/or manufacture of various embodiments of this invention may be accomplished, in part, through the use of computer readable code, including general programming languages (such as C or C++), hardware description languages (HDL) such as, for example, Verilog HDL, VHDL, Altera HDL (AHDL), or other available programming and/or schematic capture tools (such as circuit capture tools). This computer readable code can be disposed in any known computer-usable medium including a semiconductor, magnetic disk, optical disk (such as CD-ROM, DVD-ROM). As such, the code can be transmitted over communication networks including the Internet. It is understood that the functions accomplished and/or structure provided by the systems and techniques described above can be represented in a core (e.g., an APU core) that is embodied in program code and can be transformed to hardware as part of the production of integrated circuits.
Computer system 2800 includes one or more processors, such as processor 2804, Processor 2804 may be a special purpose or a general-purpose processor such as, for example, the APU and CPU of
Computer system 2800 also includes a main memory 2808, preferably random access memory (RAM), and may also include a secondary memory 2810. Secondary memory 2810 can include, for example, a hard disk drive 2812, a removable storage drive 2814, and/or a memory stick. Removable storage drive 2814 can include a floppy disk drive, a magnetic tape drive, an optical disk drive, a flash memory, or the like. The removable storage drive 2814 reads from and/or writes to a removable storage unit 2818 in a well-known manner. Removable storage unit 2818 can comprise a floppy disk, magnetic tape, optical disk, etc. which is read by and written to by removable storage drive 2814. As will be appreciated by persons skilled in the relevant art, removable storage unit 2818 includes a computer-usable storage medium having stored therein computer software and/or data.
Computer system 2800 (optionally) includes a display interface 2802 (which can include input and output devices such as keyboards, mice, etc.) that forwards graphics, text, and other data from communication infrastructure 2806 (or from a frame buffer not shown) for display on display unit 2830.
In alternative implementations, secondary memory 2810 can include other similar devices for allowing computer programs or other instructions to be loaded into computer system 2800. Such devices can include, for example, a removable storage unit 2822 and an interface 2820. Examples of such devices can include a program cartridge and cartridge interface (such as those found in video game devices), a removable memory chip (e.g., EPROM or PROM) and associated socket, and other removable storage units 2822 and interfaces 2820 which allow software and data to be transferred from the removable storage unit 2822 to computer system 2800.
Computer system 2800 can also include a communications interface 2824. Communications interface 2824 allows software and data to be transferred between computer system 2800 and external devices. Communications interface 2824 can include a modem, a network interface (such as an Ethernet card), a communications port, a PCMCIA slot and card, or the like. Software and data transferred via communications interface 2824 are in the form of signals which may be electronic, electromagnetic, optical, or other signals capable of being received by communications interface 2824. These signals are provided to communications interface 2824 via a communications path 2826. Communications path 2826 carries signals and can be implemented using wire or cable, fiber optics, a phone line, a cellular phone link, a RF link or other communications channels.
In this document, the terms “computer program medium” and “computer-usable medium” are used to generally refer to media such as removable storage unit 2818, removable storage unit 2822, and a hard disk installed in hard disk drive 2812. Computer program medium and computer-usable medium can also refer to memories, such as main memory 2808 and secondary memory 2810, which can be memory semiconductors (e.g., DRAMs, etc.). These computer program products provide software to computer system 2800.
Computer programs (also called computer control logic) are stored in main memory 2808 and/or secondary memory 2810. Computer programs may also be received via communications interface 2824. Such computer programs, when executed, enable computer system 2800 to implement embodiments of the present invention as discussed herein. In particular, the computer programs, when executed, enable processor 2804 to implement processes of embodiments of the present invention, such as the steps in the method illustrated by flowchart 900 of
Embodiments of the present invention are also directed to computer program products including software stored on any computer-usable medium. Such software, when executed in one or more data processing device, causes a data processing device(s) to operate as described herein. Embodiments of the present invention employ any computer-usable or -readable medium, known now or in the future. Examples of computer-usable mediums include, but are not limited to, primary storage devices (e.g., any type of random access memory), secondary storage devices (e.g., hard drives, floppy disks, CD ROMS, ZIP disks, tapes, magnetic storage devices, optical storage devices, MEMS, nanotechnological storage devices, etc.), and communication mediums (e.g., wired and wireless communications networks, local area networks, wide area networks, intranets, etc.).
It is to be appreciated that the Detailed Description section, and not the Summary and Abstract sections, is intended to be used to interpret the claims. The Summary and Abstract sections may set forth one or more but not all exemplary embodiments of the present invention as contemplated by the inventors, and thus, are not intended to limit the present invention and the appended claims in any way.
Embodiments of the present invention have been described above with the aid of functional building blocks illustrating the implementation of specified functions and relationships thereof. The boundaries of these functional building blocks have been arbitrarily defined herein for the convenience of the description. Alternate boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed.
The foregoing description of the specific embodiments will so fully reveal the general nature of the invention that others can, by applying knowledge within the skill of the relevant art, readily modify and/or adapt for various applications such specific embodiments, without undue experimentation, without departing from the general concept of the present invention. Therefore, such adaptations and modifications are intended to be within the meaning and range of equivalents of the disclosed embodiments, based on the teaching and guidance presented herein. It is to be understood that the phraseology or terminology herein is for the purpose of description and not of limitation, such that the terminology or phraseology of the present specification is to be interpreted by the skilled artisan in light of the teachings and guidance.
The breadth and scope of the present invention should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following claims and their equivalents.
This application claims the benefit of U.S. Provisional Patent Application No. 61/577,595, filed Dec. 19, 2011, titled “Senone Scoring Unit” and U.S. Provisional Patent Application No. 61/589,113, filed Jan. 20, 2012, titled “HW/SW Architecture for Speech Recognition,” both of which are incorporated herein by reference in their entireties.
| Number | Date | Country | |
|---|---|---|---|
| 61577595 | Dec 2011 | US | |
| 61589113 | Jan 2012 | US |
