Patent Grant
7610199
The present invention relates generally to the field of speech recognition and relates more particularly to methods for obtaining speech signals for speech recognition applications.
The accuracy of existing speech recognition systems is often adversely impacted by an inability to obtain a complete speech signal for processing. For example, imperfect synchronization between a user's actual speech signal and the times at which the user commands the speech recognition system to listen for the speech signal can cause an incomplete speech signal to be provided for processing. For instance, a user may begin speaking before he provides the command to process his speech (e.g., by pressing a button), or he may terminate the processing command before he is finished uttering the speech signal to be processed (e.g., by releasing or pressing a button). If the speech recognition system does not “hear” the user's entire utterance, the results that the speech recognition system subsequently produces will not be as accurate as otherwise possible. In open-microphone applications, audio gaps between two utterances (e.g., due to latency or others factors) can also produce incomplete results if an utterance is started during the audio gap.
Poor endpointing (e.g., determining the start and the end of speech in an audio signal) can also cause incomplete or inaccurate results to be produced. Good endpointing increases the accuracy of speech recognition results and reduces speech recognition system response time by eliminating background noise, silence, and other non-speech sounds (e.g., breathing, coughing, and the like) from the audio signal prior to processing. By contrast, poor endpointing may produce more flawed speech recognition results or may require the consumption of additional computational resources in order to process a speech signal containing extraneous information. Efficient and reliable endpointing is therefore extremely important in speech recognition applications.
Conventional endpointing methods typically use short-time energy or spectral energy features (possibly augmented with other features such as zero-crossing rate, pitch, or duration information) in order to determine the start and the end of speech in a given audio signal. However, such features become less reliable under conditions of actual use (e.g., noisy real-world situations), and some users elect to disable endpointing capabilities in such situations because they contribute more to recognition error than to recognition accuracy.
Thus, there is a need in the art for a method and apparatus for obtaining complete speech signals for speech recognition applications.
In one embodiment, the present invention relates to a method and apparatus for obtaining complete speech signals for speech recognition applications. In one embodiment, the method continuously records an audio stream which is converted to a sequence of frames of acoustic speech features and stored in a circular buffer. When a user command to commence or terminate speech recognition is received, the method obtains a number of frames of the audio stream occurring before or after the user command in order to identify an augmented audio signal for speech recognition processing.
In further embodiments, the method analyzes the augmented audio signal in order to locate starting and ending speech endpoints that bound at least a portion of speech to be processed for recognition. At least one of the speech endpoints is located using a Hidden Markov Model.
The teachings of the present invention can be readily understood by considering the following detailed description in conjunction with the accompanying drawings, in which:
To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are common to the figures.
The present invention relates to a method and apparatus for obtaining an improved audio signal for speech recognition processing, and to a method and apparatus for improved endpointing for speech recognition. In one embodiment, an audio stream is recorded continuously by a speech recognition system, enabling the speech recognition system to retrieve portions of a speech signal that conventional speech recognition systems might miss due to user commands that are not properly synchronized with user utterances.
In further embodiments of the invention, one or more Hidden Markov Models (HMMs) are employed to endpoint an audio signal in real time in place of a conventional signal processing endpointer. Using HMMs for this function enables speech start and end detection that is faster and more robust to noise than conventional endpointing techniques.
In step 108, the user begins speaking, at time t=S. The user command to commence speech recognition, received at time t=TS, and the actual start of the user speech, at time t=S, are only approximately synchronized; the user may begin speaking before or after the command to commence speech recognition received in step 106.
Once the user begins speaking, the method 100 proceeds to step 110 and requests a portion of the recorded audio stream from the circular buffer starting at time t=TS−N1, where N1 is an interval of time such that TS−N1<S≦TS most of the time. In one embodiment, the interval N1 is chosen by analyzing real or simulated user data and selecting the minimum value of N1 that minimizes the speech recognition error rate on that data. In some embodiments, a sufficient value for N1 is in the range of tenths of a second. In another embodiment, where the audio signal for speech recognition processing has been acquired using an open-microphone mode, N1 is approximately equal to Ts−TP, where TP is the absolute time at which the previous speech recognition process on the previous utterance ended. Thus, the current speech recognition process will start on the first audio frame that was not recognized in the previous speech recognition processing.
In step 112, the method 100 receives a user command (e.g., via a button press or other means) to terminate speech recognition, at time t=TE. In step 114, the user stops speaking, at time t=E. The user command to terminate speech recognition, received at time t=TE, and the actual end of the user speech, at time t=E, are only approximately synchronized; the user may stop speaking before or after the command to terminate speech recognition received in step 112.
In step 116, the method 100 requests a portion of the audio stream from the circular buffer up to time t=TE+N2, where N2 is an interval of time such that TE≦E<TE+N2 most of the time. In one embodiment, N2 is chosen by analyzing real or simulated user data and selecting the minimum value of N2 that minimizes the speech recognition error rate on that data. Thus, an augmented audio signal starting at time TS−N1 and ending at time TE+N2 is identified.
In step 118 (illustrated in phantom), the method 100 optionally performs an endpoint search on at least a portion of the augmented audio signal. In one embodiment, an endpointing search in accordance with step 118 is performed using a conventional endpointing technique. In another embodiment, an endpointing search in accordance with step 118 is performed using one or more Hidden Markov Models (HMMs), as described in further detail below in connection with
In step 120, the method 100 applies speech recognition processing to the endpointed audio signal. Speech recognition processing may be applied in accordance with any known speech recognition technique.
The method 100 then returns to step 104 and continues to record the audio stream to the circular buffer. Recording of the audio stream to the circular buffer is performed in parallel with the speech recognition processes, e.g., steps 106-120 of the method 100.
The method 100 affords greater flexibility in choosing speech signals for recognition processing than conventional speech recognition techniques. Importantly, the method 100 improves the likelihood that a user's entire utterance is provided for recognition processing, even when user operation of the speech recognition system would normally provide an incomplete speech signal. Because the method 100 continuously records the audio stream containing the speech signals, the method 100 can “back up” or “go forward” to retrieve portions of a speech signal that conventional speech recognition systems might miss due to user commands that are not properly synchronized with user utterances. Thus, more complete and more accurate speech recognition results are produced.
Moreover, because the audio stream is continuously recorded even when speech is not being actively processed, the method 100 enables new interaction strategies. For example, speech recognition processing can be applied to an audio stream immediately upon command, from a specified point in time (e.g., in the future or recent past), or from a last detected speech endpoint (e.g., a speech starting or speech ending point), among other times. Thus, speech recognition can be performed, on the user's command, from a frame that is not necessarily the most recently recorded frame (e.g., occurring some time before or after the most recently recorded frame).
In step 206, the method 200 performs a speech endpointing search using an endpointing HMM to detect the start of the speech in the received audio signal. In one embodiment, the endpointing HMM recognizes speech and silence in parallel, enabling the method 200 to hypothesize the start of speech when speech is more likely than silence. Many topologies can be used for the speech HMM, and a standard silence HMM may also be used. In one embodiment, the topology of the speech HMM is defined as a sequence of one or more reject “phones”, where a reject phone is an HMM model trained on all types of speech. In another embodiment, the topology of the speech HMM is defined as a sequence (or sequence of loops) of context-independent (CI) or other phones. In further embodiments, the endpointing HMM has a pre-determined but configurable minimum duration, which may be a function of the number of reject or other phones in sequence in the speech HMM, and which enables the endpointer to more easily reject short noises as speech.
In one embodiment, the method 200 identifies the speech starting frame when it detects a predefined sufficient number of frames of speech in the audio signal. The number of frames of speech that are required to indicate a speech endpoint may be adjusted as appropriate for different speech recognition applications. Embodiments of methods for implementing an endpointing HMM in accordance with step 206 are described in further detail below with reference to
In step 208, once the speech starting frame, FSD, is detected, the method 200 backs up a pre-defined number B of frames to a frame FS preceding the speech starting frame FSD, such that FS=FSD−B becomes the new “start frame” for the speech for the purposes of the speech recognition process. In one embodiment, the number B of frames by which the method 200 backs up is relatively small (e.g., approximately 10 frames), but is large enough to ensure that the speech recognition process begins on a frame of silence.
In step 210, the method 200 commences recognition processing starting from the new start frame FS identified in step 108. In one embodiment, recognition processing is performed in accordance with step 210 using a standard speech recognition HMM separate from the endpointing HMM.
In step 212, the method 200 detects the end of the speech to be processed. In one embodiment, a speech “end frame” is detected when the recognition process started in step 210 of the method 200 detects a predefined sufficient number of frames of silence following frames of speech. In one embodiment, the number of frames of silence that are required to indicate a speech endpoint is adjustable based on the particular speech recognition application. In another embodiment, the ending/silence frames might be required to legally end the speech recognition grammar, forcing the endpointer not to detect the end of speech until a legal ending point. In another embodiment, the speech end frame is detected using the same endpointing HMM used to detect the speech start frame. Embodiments of methods for implementing an endpointing HMM in accordance with step 212 are described in further detail below with reference to
In step 214, the method 200 terminates speech recognition processing and outputs recognized speech, and in step 216, the method 200 terminates.
Implementation of endpointing HMM's in conjunction with the method 200 enables more accurate detection of speech endpoints in an input audio signal, because the method 200 does not have any internal parameters that directly depend on the characteristics of the audio signal and that require extensive tuning. Moreover, the method 200 does not utilize speech features that are unreliable in noisy environments. Furthermore, because the method 200 requires minimal computation (e.g., processing while detecting the start and the end of speech is minimal), speech recognition results can be produced more rapidly than is possible by conventional speech recognition systems. Thus, the method 200 can rapidly and reliably endpoint an input speech signal in virtually any environment.
Moreover, implementation of the method 200 in conjunction with the method 100 improves the likelihood that a user's complete utterance is provided for speech recognition processing, which ultimately produces more complete and more accurate speech recognition results.
The method 300 is initialized at step 302 and proceeds to step 304, where the method 300 counts a number, F1, of frames of the received audio signal in which the most likely word (e.g., according to the standard HMM Viterbi search criteria) is speech in the last N1 preceding frames. In one embodiment, N1 is a predefined parameter that is configurable based on the particular speech recognition application and the desired results. Once the number F1 of frames is determined, the method 300 proceeds to step 306 and determines whether the number F1 of frames exceeds a first predefined threshold, T1. Again, the first predefined threshold, T1, is configurable based on the particular speech recognition application and the desired results.
If the method 300 concludes in step 306 that F1 does not exceed T1, the method 300 proceeds to step 310 and continues to search the audio signal for a speech endpoint, e.g., by returning to step 304, incrementing the location in the speech signal by one frame, and continuing to count the number of speech frames in the last N1 frames of the audio signal. Alternatively, if the method 300 concludes in step 306 that F1 does exceed T1, the method 300 proceeds to step 308 and defines the first frame FSD of the frame sequence that includes the number (F1) of frames as the speech starting point. The method 300 then backs up to a predefined number B of frames before the speech starting frame for speech recognition processing, e.g., in accordance with step 208 of the method 200. In one embodiment, values for the parameters N1 and T1 are determined to simultaneously minimize the probability of detecting short noises as speech and maximize the probability of detecting single, short words (e.g., “yes” or “no”) as speech.
In one embodiment, the method 300 may be adapted to detect the speech stopping frame as well as the speech starting frame (e.g., in accordance with step 212 of the method 200). However, in step 304, the method 300 would count the number, F2, of frames of the received audio signal in which the most likely word is silence in the last N2 preceding frames. Then, when that number, F2, meets a second predefined threshold, T2, speech recognition processing is terminated (e.g., effectively identifying the frame at which recognition processing is terminated as the speech endpoint). In either case, the method 300 is robust to noise and produces accurate speech recognition results with minimal computational complexity.
The method 400 is initialized at step 402 and proceeds to step 404, where the method 400 identifies the most likely word in the endpointing search (e.g., in accordance with the standard Viterbi HMM search algorithm).
In order to determine the speech starting endpoint, in step 406 the method 400 determines whether the most likely word identified in step 404 is speech or silence. If the method 400 concludes that the most likely word is speech, the method 400 proceeds to step 408 and computes the duration, Ds, back to the most recent pause-to-speech transition.
In step 410, the method 400 determines whether the duration Ds meets or exceeds a first predefined threshold T1. If the method 400 concludes that the duration Ds does not meet or exceed T1, then the method 400 determines that the identified most likely word does not represent a starting endpoint of the speech, and the method 400 processes the next audio frame and returns to step 404 and to continue the search for a starting endpoint.
Alternatively, if the method 400 concludes in step 410 that the duration Ds does meet or exceed T1, then the method 400 proceeds to step 412 and identifies the first frame FSD of the most likely speech word identified in step 404 as a speech starting endpoint. Note that according to step 208 of the method 200, speech recognition processing will start some number B of frames before the speech starting point identified in step 404 of the method 400 at frame FS=FSD−B. The method 400 then terminates in step 422.
To determine the speech ending endpoint, referring back to step 406, if the method 400 concludes that the most likely word identified in step 404 is not speech (i.e., is silence), the method 400 proceeds to step 414, where the method 400 confirms that the frame(s) in which the most likely word appears is subsequent to the frame representing the speech starting point. If the method 400 concludes that the frame in which the most likely word appears is not subsequent to the frame of the speech starting point, then the method 400 concludes that the most likely word identified in step 404 is not a speech endpoint and returns to step 404 to process the next audio frame and continue the search for a speech endpoint.
Alternatively, if the method 400 concludes in step 414 that the frame in which the most likely word appears is subsequent to the frame of the speech starting point, the method 400 proceeds to step 416 and computes the duration, Dp, back to the most recent speech-to-pause transition.
In step 418, the method 400 determines whether the duration, Dp, meets or exceeds a second predefined threshold T2. If the method 400 concludes that the duration Dp does not meet or exceed T2, then the method 400 determines that the identified most likely word does not represent an endpoint of the speech, and the method 400 processes the next audio frame and returns to step 404 to continue the search for an ending enpoint.
However, if the method 400 concludes in step 418 that the duration Dp does meet or exceed T2, then the method 400 proceeds to step 420 and identifies the most likely word identified in step 404 as a speech endpoint (specifically, as a speech ending endpoint). The method 400 then terminates in step 422.
The method 400 produces accurate speech recognition results in a manner that is more robust to noise, but more computationally complex than the method 300. Thus, the method 400 may be implemented in cases where greater noise robustness is desired and the additional computational complexity is less of a concern. The method 300 may be implemented in cases where it is not feasible to determine the duration back to the most recent pause-to-speech or speech-to-pause transition (e.g., when backtrace information is limited due to memory constraints).
In one embodiment, when determining the speech ending frame in step 418 of the method 400, an additional requirement that the speech ending word legally ends the speech recognition grammar can prevent premature speech endpoint detection when a user utters a long pause in the middle of an utterance.
Alternatively, the digital scheduling engine, manager or application (e.g., speech endpointer 505) can be represented by one or more software applications (or even a combination of software and hardware, e.g., using Application Specific Integrated Circuits (ASIC)), where the software is loaded from a storage medium (e.g., I/O devices 506) and operated by the processor 502 in the memory 504 of the general purpose computing device 500. Thus, in one embodiment, the speech endpointer 505 for endpointing audio signals described herein with reference to the preceding Figures can be stored on a computer readable medium or carrier (e.g., RAM, magnetic or optical drive or diskette, and the like).
The endpointing methods of the present invention may also be easily implemented in a variety of existing speech recognition systems, including systems using “hold-to-talk”, “push-to-talk”, “open microphone”, “barge-in” and other audio acquisition techniques. Moreover, the simplicity of the endpointing methods enables the endpointing methods to automatically take advantage of improvements to a speech recognition system's acoustic speech features or acoustic models with little or no modification to the endpointing methods themselves. For example, upgrades or improvements to the noise robustness of the system's speech features or acoustic models correspondingly improve the noise robustness of the endpointing methods employed.
Thus, the present invention represents a significant advancement in the field speech recognition. One or more Hidden Markov Models are implemented to endpoint (potentially augmented) audio signals for speech recognition processing, resulting in an endpointing method that is more efficient, more robust to noise and more reliable than existing endpointing methods. The method is more accurate and less computationally complex than conventional methods, making it especially useful for speech recognition applications in which input audio signals may contain background noise and/or other non-speech sounds.
Although various embodiments which incorporate the teachings of the present invention have been shown and described in detail herein, those skilled in the art can readily devise many other varied embodiments that still incorporate these teachings.
This application claims the benefit of U.S. Provisional Patent Application No. 60/606,644, filed Sep. 1, 2004 (entitled “Method and Apparatus for Obtaining Complete Speech Signals for Speech Recognition Applications”), which is herein incorporated by reference in its entirety.
This invention was made with Government support under contract number DAAH01-00-C-R003, awarded by Defense Advance Research Projects Agency and under contract number NAG2-1568 awarded by NASA. The Government has certain rights in this invention.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5596680 | Chow et al. | Jan 1997 | A |
| 5692104 | Chow et al. | Nov 1997 | A |
| 6324509 | Bi et al. | Nov 2001 | B1 |
| 7139707 | Sheikhzadeh-Nadjar et al. | Nov 2006 | B2 |
| 7260532 | Rees | Aug 2007 | B2 |
| Number | Date | Country | |
|---|---|---|---|
| 20060241948 A1 | Oct 2006 | US |
| Number | Date | Country | |
|---|---|---|---|
| 60606644 | Sep 2004 | US |
