Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Reissue Patent
2001-07-13
2004-11-09
Banks-Harold, Marsha D. (Department: 2654)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
C704S253000, C704S254000, C379S088030
Reissue Patent
active
RE038649
ABSTRACT:
TECHNICAL FIELD
The invention relates to an automatic speech recognition method and apparatus and more particularly to a method and apparatus that speeds up recognition of connected words.
DESCRIPTION OF THE PRIOR ART
Various automatic speech recognition methods and systems exist and are widely known. Methods using dynamic programming and Hidden Markov Models (HMMs) are known as shown in the article Frame-Synchronous Network Search Algorithm for Connected Word Recognition by Chin-Hui Lee and Lawrence R. Rabiner published in the IEEE Transactions on Acoustics, Speech, and Signal Processing Vol. 37, No. 11 November 1989. The Lee-Rabiner article provides a good overview of the state of methods and systems for automatic speech recognition of connected words in 1989.
An article entitled A Wave Decoder for Continuous Speech Recognition by E. Buhrke, W. Chou and Q. Zhou published in the Proceedings of ICSLP in October 1996 describes a technique known as beam searching to improve speech recognition performance and hardware requirements. The Buhrke-Chou-Zhou article also mentions an article by D. B. Paul entitled “An Efficient A* Stack Decoder . . . ” which describes best-first searching strategies and techniques.
Speech recognition, as explained in the articles mentioned above, involves searching for a best (i.e. highest likelihood score) sequence of words, W
1
-W
n
, that corresponds to an input speech utterance. The prevailing search algorithm used for speech recognition is the dynamic Viterbi decoder. This decoder is efficient in its implementation. A full search of all possible words to find the best word sequence corresponding to an utterance is still too large and time consuming. In order to address the size and time problems, beam searching has often been implemented. In a beam search, those word sequence hypotheses that are likely, that is within a prescribed mathematical distance from the current best score, are retained and extended. Unlikely hypotheses are ‘pruned’ or removed from the search. This pruning of unlikely word sequence hypotheses has the effect of reducing the size and time required by the search and permits practical implementations of speech recognition systems to be built.
At the start of an utterance to be recognized, only those words that are valid words to start a sequence based on a predetermined grammar can be activated. At each time frame, dynamic programming using the Viterbi algorithm is performed over the active portion of the word network. It is worth noting that the active portion of the word network varies over time when a beam search strategy is used. Unlikely word sequences are pruned away and more likely word sequences are extended as specified in a predetermined grammar. These more likely word sequences are extended as specified in the predetermined grammar and become included in the active portion of the word network. At each time frame the system compiles a linked list of all viable word sequences into respective nodes on a decoding tree. This decoding tree, along with its nodes, is updated for every time frame. Any node that is no longer active is removed and new nodes are added for newly active words. Thus, the decoding tree maintains viable word sequences that are not pruned away by operation of the beam search algorithm by means of the linked list. Each node of the decoding tree corresponds to a word and has information such as the word end time, a pointer to the previous word node of the word sequence and the cumulative score of the word sequence stored therein. At the end of the utterance, the word nodes with the best cumulative scores are traversed back through their sequences of pointer entries in the decoding tree to obtain the most likely word sequence. This traversing back is commonly known in speech recognition as ‘backtracking’.
A common drawback of the known methods and systems for automatic speech recognition is the use of energy detectors to determine the end of a spoken utterance. Energy detection provides a well known technique in the signal processing and related fields for determining the beginning and ending of an utterance. An energy detection based speech recognition method
200
is shown in FIG.
2
. Method
200
uses a background time framing arrangement (not shown) to digitize the input signal, such as that received upon a telephone line into time frames for speech processing. Time frames are analyzed at step
202
to determine if any frame has energy which could be significant enough to start speech processing. If a frame does not have enough energy to consider, step
202
is repeated with the next frame, but if there is enough energy to consider the content of a frame, method
200
progresses to steps
204
-
210
which are typical speech recognition steps. Next, at step
220
, the frame(s) that started the speech recognition process are checked to see if both the received energy and any system played aural prompt occurred at the same time. If the answer is yes, a barge in condition has occurred and the aural prompt is discontinued at step
222
for the rest of the speech processing of the utterance. Next, either from a negative determination at step
220
or a prompt disable at step
222
, step
224
determines if a gap time without significant energy has occurred. Such a gap time signifies the end of the present utterance. If it has not occurred, that means there is more speech to analyze and the method returns to step
204
, otherwise the gap time with no energy is interpreted as an end of the current utterance and backtracking is started in order to find the most likely word sequence that corresponds to the utterance. Unfortunately, this gap time amounts to a time delay that typically ranges from one to one and a half seconds. For an individual caller this delay is typically not a problem, but for a telephone service provider one to one and a half seconds on thousands of calls per day, such as to automated collect placing services, can add up. On 6000 calls, one and one-half seconds amounts to two and one-half hours of delay while using of speech recognition systems. For heavily used systems this one-to one and one-half second delay causes the telephone service provider to buy more speech recognizers or lose multiple hours of billable telephone service. Further, since the backtracking to find the most likely word sequence does not begin until the end-of-utterance determination has been made based on the energy gap time, the use of partial word sequences for parallel and/or pipelining processes is not possible.
It is an object of the present invention to provide a method for determining an end of an utterance that is faster than speech energy gap timing.
It is another object of the present invention to provide a method for reliably detecting a group of words within an utterance in real time as partial word sequences of the utterance to allow parallel processing of the first portion of the utterance.
It is another object of the present invention to provide reliable barge-in over aural prompts.
SUMMARY OF THE INVENTION
Briefly stated, in accordance with one embodiment of the invention, the foregoing objects are achieved by providing a method having a step of determining if a speech utterance has started, if an utterance has not started then obtaining next frame and re-running this speech utterance start determining step. If an utterance has started, the next step is obtaining a speech frame of the speech utterance that represents a frame period that is next in time. Next, features are extracted from the speech frame which are used in speech recognition. The next step is performing dynamic programming to build a speech recognition network followed by the step of performing a beam search using the speech recognition network. The next step is updating a decoding tree of the speech utterance after the beam search. The next step is determining if a first word of the speech utterance has been received and if it has been received disabling any aural prompt and continuing to the next step, otherwise, if a first word has
Setlur Anand Rangaswamy
Sukkar Rafid Antoon
Banks-Harold Marsha D.
Lerner Martin
Lucent Technologies - Inc.
LandOfFree
Method and apparatus for word counting in continuous speech... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and apparatus for word counting in continuous speech..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and apparatus for word counting in continuous speech... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3309226