JPH0372992B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention will be explained in the following order. A. Industrial application field B Summary of disclosure C. Conventional technology D. Problem that the invention aims to solve E. Means to solve the problem F Example F1 speech recognition system environment F1a General explanation (Figures 4 to 9) F1b auditory model and speech recognition system sounds
Its realization in the Hibiki processor (Part 10)
Figure ~ Figure 16) F1c precision matching (Fig. 6, Fig. 17) F1d Basic high-speed matching (Figures 18 to 20)
figure) F1e Alternative high-speed matching (Fig. 21, 22
figure) F1f Matching based on first J level
(Figure 22) F1g phoneme tree structure and high-speed matching example
(Figure 23) F1h language model (Figure 4, Figure 24) F1j stack decoder (Figures 7 to 9,
Figure 25) F1k Construction of phonetic basic form F2 Start phoneme machine and end phoneme machine
Formation of a set of phoneme machines including (Fig. 1A)
~Figure 3, Figure 26~Figure 32) F3 table G Effect of invention A. Industrial application field The present invention is based on speech recognition using a Markov model.
As for the method, write your words paying particular attention to the beginnings of words.
basic form (referring to the Markov model of a word)
) and create a word basic form like this.
It was designed to be used. B Summary of disclosure The present invention produces strings of acoustic labels.
(Phonetic phoneme machine (phonetic phoneme machine)
including forming a set of
A device and method for constructing word basic forms that can
Disclose the law. Each phoneme machine has () multiple states.
state, () each indicates a transition from one state to another.
multiple transitions, () stored probabilities for each transition,
and () has a memorized label output probability,
, each label output probability is the corresponding label
corresponding to the probability of each said phoneme machine to generate; said
Setting of phonetic type phoneme machine starts.Support of phoneme machine starts.
If configured to include a bushet, each open
The memorized probabilities of the initial phoneme machine are
at least one phonetic element pronounced at the beginning of the g
Corresponding to; the setting of the phonetic symbol type phoneme machine is completed.
Configured to include a subset of phoneme machines
If so, the memorized probability of each ending phoneme machine
is uttered at the end of the audio segment.
also corresponds to one phonetic element. Word basic form
is the concatenation of phoneme machines selected from the set.
More constructed. C. Conventional technology As an invention providing the background or environment of the present invention,
U.S. Patent Application No. 06/665401 (dated October 26, 1984)
Application), and No. 06/672974 (November 19, 1984)
There is an application). In probabilistic methods for recognizing speech, the acoustic waveform is
For the first time, a label string can be created using a sound processor.
is converted to Alphabet of labels (set)
typically consists of about 200 different labels, and this
Identify the corresponding acoustic type using the label.
The generation of such labels has been reported in various papers as well as
As described in the aforementioned U.S. Patent Application No. 06/665,401.
Ru. Simply put, the acoustic input is divided and continuous.
into time frames and label each time frame
Assign. Labels are usually based on energy characteristics.
It is formed by Marco on using labels for speech recognition
The f-model (stochastic finite state machine) has already been proposed.
It is being proposed. Markov models typically have multiple
Contains states and transitions between those states. Furthermore,
Markov models typically include (a) the probability of occurrence of each transition;
and (b) each generating each label at various transitions.
has a probability assigned to it with respect to the probability of
Ru. Markov model (or equivalently Markov model
IEEE Proceedings: Pattern Analysis and
Computer Information (PAMI) Vol. 5 No. 2 (March 1983)
month) outside of L.R.B. listed on pages 179-190.
Paper “Maximum Likelihood Method for Recognizing Continuous Speech” (L.R. Bahl
et al, “A Maximum Likelihood Approach
to Continuous Speech Recognition”, IEEE
Transactions on Pattern Analysis and
Machine Intelligence, Vol.PAMI-5, No.o.2,
March 1983)
There is. A Markov model machine is a Markov model machine.
Also called model phoneme machine or simply phoneme machine.
cormorant. When recognizing speech, which word(s) in the vocabulary?
) is generated by a sound processor.
has the highest likelihood of yielding a label string
A matching process is performed to determine the
Ru. One such matching procedure is the
As shown in patent application no. 06/672974. in addition
According to Acoustic matching, (a) each word in the vocabulary
into a sequence of Markov model phoneme machines.
(b) the phoneme machine represented by each word;
of the sequence, generated by the acoustic processor.
Determine the likelihood of each yielding a string of labels.
It is executed by setting Each word represents
Phoneme machine sequences correspond to word basic forms
do. When forming a word basic form, first
We define the properties of the phoneme machine used to construct the basic form of
need to be justified. Said U.S. Patent Application No. 06/
No. 672974, a work constructed from phonetic phoneme machines.
The basic form of the card is shown. In this case, each phoneme
The machine corresponds to a phonetic type single note, and has 7 states and 13
Contains transitions. In detail, each corresponds to
The basic form is a set of about 70 phonemes representing phonetic elements.
It has become the basis for building the. Generally word base
In this format, a phonetician identifies words with their respective phonetic symbols.
the corresponding phoneme machine for each phonetic symbol.
Constructed by assigning to segments
Ru. Traditionally, each of the 70 phonemes corresponds to a given grade.
whether the single note is at the beginning, middle, or end of the word.
represents a given phonetic magnitude, regardless of whether it occurs in
did. For example, the single note “k” is similar to “cat”.
at the beginning of a word like “scat”.
in the middle of the word, or in the case of “back”
The phoneme k occurs in any case at the end of the letter k.
was displayed. D. Problem that the invention aims to solve The present invention provides that the predetermined sound is
i.e. precedes or follows the quiet period?
It expresses different energy characteristics depending on the
Based on knowledge. In particular, the present invention is preceded by a quiet period.
Depending on the sound, the energy may be amplified and
If a silent period follows, some sounds may lose energy.
It is based on the fact that it is attenuated. Energy special
The quality generally determines the label that should be generated for the acoustic input.
It is used by the sound processor when
depending on whether the sound occurs at the beginning or end of the code.
This results in energy enhancement or attenuation, resulting in different lamps.
Bells may be generated. Therefore, the invention provides that the sound is uttered at the beginning of the word.
There are several first methods that involve energy enhancement when
With a phoneme machine of the type, a sound is uttered at the end of a word.
If the
Define the second type of phoneme machine. Furthermore, it is uttered
sounds with significant energy enhancement or attenuation.
There is a third type of phoneme machine for cases where
Ru. Start the first type of phoneme machine, the first type of phoneme machine.
Ending the second type of phoneme machine, the third type of phoneme machine
The phoneme machine is called the common phoneme machine. Start phoneme machine statistics reflect transitions from silence.
The end phoneme machine statistics indicate the transition to silence.
reflect. The common phoneme machine starts in the middle of the word.
Generally speaking, the sound that is spoken is a transition to silence.
Transitions from or to silence are added to the statistics of the phoneme machine.
The sound uttered at the word position does not have a large effect.
It is desirable to have corresponding statistics. A given sound is the part of the word in which that sound is uttered.
even if the corresponding energy characteristics do not change significantly.
If there is a common phoneme machine associated with it, then
do. In accordance with the present invention, a plurality of starting phoneme machines and
An ending phoneme machine is provided, and a predetermined sound is used for a period of silence.
Give the energy characteristics when it occurs adjacent to . Thus, according to the invention, a given word
Open with a target sound that has a starting phoneme machine that corresponds to
starts with that starting phoneme machine, and
Basic form followed by a common phoneme machine for the target sound
is configured to obtain. Similarly, according to the present invention
A given word has its corresponding ending phoneme machine.
If the word ends with a target sound that has a
ends with its ending phoneme machine, and its target
We will obtain the basic form preceded by the common phoneme machine of the sound.
It is composed of sea urchins. Therefore, the object of the present invention is to (construct word basic forms)
) The Markov model set includes
for single sounds that occur in transitions or transitions from silence.
including the corresponding Markov model, and
precision in word recognition systems using this form.
The aim is to increase the level of Furthermore, it is an object of the present invention to obtain similar energy enhancement properties.
Group sounds that have , into that group.
Define a single-start phoneme machine for all sounds that belong to
sound with similar energy attenuation characteristics.
group together, and everything belonging to that group.
by defining a single-end phoneme machine for all sounds.
The goal is to limit the total number of phoneme machines. E. Means to solve the problem The method to achieve the above objective includes the following steps:
nothing. (a) Each phoneme machine has () multiple states, ()
Multiple transitions, each transitioning from one state to another
transition, () memorized probability for each transition, () notation
Stored label output probability (each label output probability is
of each phoneme machine to generate the corresponding label.
corresponding to the probability), then the phonetic phoneme
Form a set of machines. (b) The set of phonetic alphabet type phoneme machines is the starting phoneme machine.
formed to include a subset of thin
If, the memorized probability of each starting phoneme machine is
At least the first utterance of the audio segment
Corresponds to one phonetic type element. (c) a given starting phoneme machine to which a word corresponds.
starts with a phonetic element with a given starting sound.
If you have a word elementary form that starts with elementary machine,
Then, each word basic form is sequenced by a phoneme machine.
Build it as a service. The method of the invention further has the following features. Sunawa
In other words, the set of phonetic symbol type phoneme machines is the end phoneme machine.
formed to contain a subset of thin, each end
The phoneme machine's memorized probabilities are the phonetic segment
at least one single phonetic type pronounced at the end of
corresponds to an element. Also, word corresponds to it
Ends with a phonetic type element with a given ending phoneme machine
and the word base ending with a given ending phoneme machine
form, each word base form has a phoneme machine.
constructed as a sequence of events. The device of the present invention which achieves the above object is a Markovian
Contains a set of model phoneme machines. Each phoneme machine
is () multiple states, () each state
Multiple transitions from to a state, () per transition
means to store the probability of () the label output probability
Means for storing (each label output probability is
At the transition, each phoneme machine generates a specific label.
corresponds to the probability that Part of the phoneme machine is the start phoneme
machine, each starting phoneme machine is ()
relating to at least one sound from a set of sounds;
At least one related sound at the beginning of a word ()
The transition probability and label output probability are shaped from the utterance of
be done. ) and the basic form of each word as a phoneme map.
Means for constructing as a sequence of syn (the construction
The means is that the word corresponding to the basic form of the target word is
When starting with a sound associated with a given starting phoneme machine
If the given starting phoneme appears at the beginning of the target word basic form,
Including means of placing the machine. ) is characterized by having
shall be. Furthermore, as a feature of the device of the present invention, a phoneme machine
Part of is the end phoneme machine (each end phoneme machine is
() associated with at least one sound from a set of sounds
and () at least at the end of the audio segment.
Transition probabilities and labels from the utterance of one related sound
Output probabilities are formatted. ), the construction means
furthermore, the word corresponding to the basic form of the target word is
If it ends with a sound related to a given ending phoneme, the
Given ending phoneme machine at the end of the elephant word base form
including means for placing. Furthermore, the device of the present invention each
In response to sounds that are affected by energy enhancement,
A common representation of the sound when it is not affected by Lugi enhancement.
also includes a phoneme machine, in which case the construction means further comprises:
, the starting phoneme machine and the subsequent common phoneme machine
(A particular sound starts a word, and the starting phoneme machine
(corresponding to a particular sound) if relevant.
including means for containing. F Example F1 speech recognition system environment F1a General explanation (Figures 4 to 9) FIG. 4 is an overview diagram of the speech recognition system 1000.
Show the diagram. This system is a stack deco
1002 and the audio processor connected to it.
Setusa (AP) 1004, high-speed approximate acoustic matuchin
Array processor 1006 performs precision
Array processor 1 that performs acoustic matching
008, language model 1010, and works
station 1012. The audio processor 1004 receives audio waveform input as
roughly identify the monotone symbol each of which corresponds to
A string of labels, i.e.
The microphonemes obtained from the front end are expressed like this.
call to. In principle, it is obtained from the front end.
(but not limited to)
It is. In this system, the sound processor 10
04 is based on a unique model of human hearing
No. 06/665,401 (October 26, 1984)
(Applications filed in Japan). Label from sound processor 1004, i.e.
Chifinim is stack decoder 1002.
Sent. FIG. 5 shows a stack decoder 100.
2 shows a logic element. i.e. stack deco
The reader 1002 connects the search device 1020 and
Connected workstation 1012,
Face 1022, 1024, 1026 and
Contains 1028. Each of these interfaces
are the sound processor 1004 and the array processor.
service 1006, 1008 and language model 101
0 respectively. During operation, the sound processor
The finem from the server 1004 is the search device 102
0 causes array processor 1006 (fast match
Sent to Fast matching explained below
The procedure is described in US Pat. No. 06/672,974 (November 1984).
(filed on the 19th). matching eyes
Simply put, the target is a given label string.
The most likely word (or multiple words)
The purpose is to determine the Fast matching searches for words in the vocabulary of words.
the string of the given incoming label.
is designed to reduce the number of candidate words in
There is. Fast matching is a stochastic finite state machine
(also referred to as Markov model in this specification)
It is something that can be developed. After fast matching reduces the number of candidate words,
Stack decoder 1002 uses language model 10
10 and, if possible, the existing Mieji.
Based on each candidate word in the high-speed matching candidate list,
Determine the contextual likelihood of the code. Precision matching allows these words to be
A fast matutin with moderate likelihood as a word
from the candidate list based on language model calculations.
It is recommended that the Precision matching is also the same as above.
National Patent Application No. 06/672974 (dated November 19, 1984)
application). Precision matching is the 6th
A Markov model phoneme machine as shown in the figure
Execute more. After precision matching, call the language model again
It is desirable to determine the word likelihood. Book
The stack decoder 1002 of the invention comprises:-
matching, precision matching, and language models.
Using the information obtained from the use of - generated labels
most likely path for the word in the string
i.e. designed to determine the sequence
There is. Find the most promising word sequence
The two conventional methods are Viterbi decoding.
decoding and single stack decoding. These people
Each of the above-mentioned Barr outsider papers “Recognize Continuous Speech”
Viterbi decoding
In Section V of the paper, single stack decoding
is also mentioned in the section. In the single stack decoding method, the
The cases are listed according to their likelihood in a single stack, and the
Decoding is performed based on the stack. single star
The likelihood of Tsuku decoding is somewhat dependent on the path length.
This is due to the fact that
is used. Viterbi's method does not require normalization and is generally
Suitable for small tasks. Another alternative is to use the possible combinations of words
each as a possible word sequence.
Check which combinations of label strings were generated.
to determine which has the highest probability of causing
perform decoding with a small vocabulary system by
be able to. The amount of computation required for this method is
In the case of a similar vocabulary system, it becomes huge and impractical.
be. Stack decoder 1002 actually
act to control elements of the
There aren't many calculations. Therefore, stack decoder 1
002 is VM (virtual machine)/system program
Duct Introduction Release 3
(1983), as described in publications such as
IBM VM/370 Operating System Controls
Desired to include a 4341 processor running under control.
Delicious. Array processors that perform a significant amount of calculations
Setsa is a floating point system
This is achieved using a commercially available 190L manufactured by (FPS).
Ru. Figures 7, 8 and 9 are
Multiplex system invented by L.R. Bahl et al.
Novel techniques including tack method and unique judgment method
Show the law. Figure 7 shows the duplicates generated with continuous label spacing.
continuous label y of number₁y₂……It is shown. Figure 8 also shows multiple word passes, i.e.
path A, path B and path C are shown.
Ru. In the context of Figure 7, path A is the entry “to
be or”, path B is entry “two b”, path C is
May correspond to the entry “too”. Target Wa
For each word pass, “The target word pass is
There is a label with the highest probability of
Such a label is called a boundary label. of a word path W representing a sequence of words.
Most likely end time (labeled as boundary label)
(shown on the bell string) is an IBM
Technology Disclosure Bulletin Vol. 23 No. 4 (September 1980 issue)
Le R. Barr et al.'s paper “Acoustic matching calculations”
“Faster Acoustic” (L.R. Bahl et al, “Faster Acoustic
Match Computation”, IBM Technical
Disclosure Bulletin, Vol.23, No.4,
September 1980).
Known ways to explore likely boundaries between
can be found by Simply put, this
The paper addresses two important points: (a) How many label strings Y are in the word
(or word sequence)
(b) At what label spacing (label string
the partial sentence (corresponding to the part of the sentence) ends, or
This article discusses ways to address this issue. Any given word pass has a label strip.
from the first label of the string to the border label.
The “likelihood” associated with each label or label spacing
There is a “likelihood value” for a given word path.
All degree values are the “likelihood vector” for a given word path.
”, so for each word pass,
There is a likelihood vector. Likelihood value L_tis shown in Figure 8.
It is. Word Pass W¹,W²,...,W^Sof a gathering of
“Likelihood envelope” Λ at label interval t_tis mathematically
It is defined as follows. Λ_t=max(L_t(W¹),...,L_t(W^S)) That is, for each label interval, the likelihood envelope is
the latest associated with any word pass in said collection.
Contains a high likelihood value. Figure 8 shows the likelihood envelope 1040
It is shown. If the word path corresponds to a complete sentence,
considered “complete”. For the complete path, please enter
When a speaker reaches the end of a sentence, e.g.
It is recommended that you press the button to identify it. entered
The input is synchronized with the label interval that marks the end of the sentence.
It will be done. A complete word pass can be added by appending the word
It cannot be extended further. Partial work
The de pass corresponds to an incomplete sentence, so it cannot be extended.
I can do it. A partial path can be a “living” path or a “dead” path.
Word Pass is classified as
“dead” when it has already been extended
However, when it has not yet been extended, it is “alive”.
This classification has already been extended by at least one
form a word pass that extends beyond the
The path will not be considered for extension again at
do not have. Each word path is
Characterized as a “good” or “bad” path.
can be used. The word pass is
The label corresponding to the bell whose word pass is
Good if the likelihood value is within the maximum likelihood envelope.
password. In other cases, the work
A do pass is a bad word pass. maximum likelihood hull
It is good (bad) to reduce each value of the connecting line by a certain value.
It is not desirable to change the limit level
Yes, but not necessarily. There is a stack element for each label interval.
Ru. Each living word pass looks like this
label corresponding to the boundary label of the living path
Assigned to the stack element corresponding to the interval. vinegar
Tack elements (listed in order of likelihood values)
0, 1 or more word pass entries
It may have. Next, the stack decoder 1002 in FIG.
The steps to be executed will be explained below. Form a likelihood envelope and determine which word path is better
Determining the stack decoding method shown in Figure 9
There is a mutual relationship as shown in the flowchart. In the flowchart of FIG. 9, block 1050
So, first, the null path enters the first stack 0.
Ru. At block 1052, the previously confirmed completion
If the (complete) stack element containing the complete path is
If available, it will be provided. in a (complete) stack element
Each complete path has a likelihood vector associated with it
has a file. has the highest likelihood for that boundary label
The likelihood vector of the complete path is first calculated by the maximum likelihood envelope
Decide on the line. If a (complete) stack element has a complete
If there is no path, the maximum likelihood envelope is
It is initialized to −∞. Additionally, the complete path is specified
Even if the maximum likelihood envelope is not
The period may be set. The initial setting for the envelope is
Locks 1054 and 1056 take place. The maximum likelihood envelope is initialized and then adjusted by a predetermined amount Δ
of the Δ predetermined value over the reduced likelihood.
Δ that forms a good region and is below the reduced likelihood
Form a poorly defined area. The larger Δ, the larger
The longer the number of word passes, the more likely they are to be extended.
becomes larger. L_tlog to determine_Tenuse
In this case, a satisfactory result is obtained if the value of Δ is 2.0.
It will be done. The value of Δ is uniform along the length of the label interval
Although it is desirable, it is not necessary to
isn't it. The word path has a good delta specification on its boundary label.
If the word has a likelihood that lies within the
The path is marked "good". In other cases
, the word pass is marked as “bad”. As shown in Figure 9, the likelihood envelope is updated and the word
A “good” (extendable) path, or
or as a “bad” path.
Blog about finding the longest unmarked word pass
It starts with 1058. Not marked with 2 or more
short word path corresponds to the longest word path length
If the border label is on the stack
A word path with a likelihood of is selected. War
block 1060 if a path is found.
, the likelihood at that boundary label is a region with good Δ specification
Check whether it is inside. If it's in a good area
If so, in block 1062, the region with poor Δ regulation is detected.
In block 1058, mark the path in
Look for unmarked living paths. if
If it is within the good range, block 1064
Mark the path within a well-defined area and block 10.
66, update the likelihood envelope and mark it as “good”.
Contains the likelihood value of the detected path. That is, la
For each bell interval, the updated likelihood value is (a) the current likelihood value within its likelihood envelope; (b) related to word passes marked “good”;
likelihood value is defined as the larger likelihood value between. this
Operations are performed in blocks 1064 and 1066.
It will be done. After the envelope has been updated, block 1058
Back to the longest unmarked, best alive
Find the password again. This loop uses unmarked word
Iterates until there are no more spaces left. Not marked
If you run out of passwords, block 1070
The shortest word path marked as “good” in
is selected. If 2 or more have the shortest length
If you have a “good” word pass, block 10
72, find the word with the highest likelihood for that boundary label.
The selected shortest path is
It will be extended. i.e. at least one prospect
If a following word is a fast match, as described above,
language models, precision matching, and languages
determined by successful execution of the model procedure.
Ru. For each likely subsequent word, extended
A word pass is formed. In detail,
The lengthened word path is the shortest word path selected.
Append a promising trailing word to the end of the decoded path.
Formed by adding The selected shortest word path is
After forming the code path, the selected word
The path is removed from the stack in which it was an entry.
instead, each extended word
The path is inserted into the appropriate stack. In particular, extension
The word path that has been created will be
entry into the stack (block 107).
2). An extended path is formed and its stack is re-formed.
Once formed, return to block 1052 and continue the process.
The process is repeated. Therefore, at each iteration, the shortest, best “good”
selected and extended. in some iteration
Word paths marked as “bad” paths are
Repetition can result in a “good” path, so
Is the word/pass you are using “good” or “bad”?
The path characteristic is uniquely given at each iteration.
It will be done. In reality, the likelihood envelope is between one iteration and the next
The word path does not change significantly between iterations.
Efficiently perform calculations to determine whether something is good or bad
This eliminates the need for normalization. If you want to identify a complete sentence, block 1074
It is desirable to include. i.e. alive
What remains unmarked in Word Pass
If there is no “good” word pass that should be extended
If so, decryption ends. each of its boundary labels
The complete word path with the highest likelihood
the most likely word for the power label string.
identified as a code sequence. For continuous speech where sentence ends are not identified, path extension
The process is continuous, i.e. the system
Regarding the predetermined number of words desired by the system user.
It is done. F1b auditory model and speech recognition system acoustics
Its realization in the processor (Figures 10 to 1)
Figure 6) FIG. 10 shows an audio processor 11 as described above.
A specific example of 00 is shown. Acoustic wave input (e.g.
natural sounds) are sampled at a predetermined rate.
Enters A/D converter 1102. representative sample
The processing speed is 1 sample every 50 microseconds.
Ru. time to shape the edges of the digital signal.
A window generator 1104 is provided. Time window occurs
The output of the
FFT (Fast Fourier Transform) device that provides torque output
Enter 1106. Then, the output of the FFT device 1106 is labeled
L₁L₂...L_fis processed to generate . Features
Selection device 1108, cluster device 1110, prototype
The device 1112 and the encoder 1114 cooperate.
generate a label. When generating labels, the original
places points (or vectors) in space based on selected features.
formed as a vector). The audio input is selected
A counterpart that can be compared to the original due to the same characteristics
special to supply points (or vectors) into space.
It is marked. In detail, when defining the prototype, the cluster
Each set of points is clustered by setting 1110.
summarized as The way to form clusters is to
A probability distribution (such as a Gaussian distribution) applied to voices
Based on. The prototype of each cluster is (class
(with respect to the center locus or other features of the data)
is generated by the location 1112. The generated prototype
and acoustic input (both with the same features selected).
) enters the encoder 1114. Symbolization device 1
114 performs a comparison procedure so that a particular sound
Assign a label to the acoustic input. By the way, the method for assigning labels to acoustic input is
It is designed for applications other than speech recognition. subordinate
Therefore, such a method and its symbolization are
The device can generally be used in speech recognition systems.
can. The selection of appropriate features represents the acoustic (speech) wave input.
This is an important element when removing the wasu label. here
The acoustic processor described has improved feature selection.
Includes device 1108. Follow this sound processor
Then, the auditory model is retrieved and used in the speech recognition system.
Used in sound processors. Listen according to Figure 11.
Explain the sensory model. FIG. 11 shows a portion of the human inner ear. detailed
Inner hair cells 1200 and fluid-containing grooves
The distal end 1202 extending to 1204 is shown in detail.
ing. Furthermore, upstream from the inner hair cells 1200,
Outer hair cell 1206 and distal end extending into groove 1204
1208 is shown. 1200 inner hair cells and outer hair cells
Hair cells 1206 contain nerves that transmit information to the brain.
are combined. In particular, Nieuron is an electrochemical
electrical pulses are carried along the nerves to the brain.
and will be processed. Electrochemical changes are
Stimulated by mechanical movement of the basement membrane 1210. The basement membrane 1210 serves as a frequency analyzer for acoustic wave input.
The area along the basement membrane 1210
Is it conventional to respond to each critical frequency band?
is known. responds to the corresponding frequency band.
Each portion of the basement membrane 1210 has an acoustic waveform
Affects the perceived volume of the input. That is,
The volume of the tone can be determined by combining two tones of similar power intensity.
than if the bands occupy the same frequency band.
If the two tones are in separate critical frequency bands
It is perceived as larger when the By the basement membrane 1210
There are 22 classes of critical frequency bands defined by
I know that. Match the frequency response of the basement membrane 1210
Therefore, the present invention is advantageous in that the critical frequency band is
Physically convert the input acoustic waveform into some or all parts.
and then for each defined critical frequency band.
The signal components are examined separately. This feature is
Applying the signal from FFT device 1106 (Figure 10)
Each critical frequency band is carefully filtered and tested.
Providing a separate signal to feature selection device 1108
This is done by A separate input is also provided by the time window generator 1104.
Blot on a time frame (preferably 25.6 ms)
is blocked. Therefore, the feature selector 1108 has 22
It is desirable to include the following signals. of these signals
each for a given frequency band per time frame
represents the strength of the sound. The signal is a conventional critical band filter in Figure 12.
Preferably, the filter is filtered by filter 1300. Next
, the signal is individually expressed as a function of frequency and the change in volume as a function of frequency.
Processed by a volume equalization converter 1302 that perceives
do. By the way, given dB level at one frequency
The perceived loudness of the first tone of the
the volume of the second tone of the same dB level at the frequency of
It may be different. Volume equalization converter 1302
are based on empirical data and each frequency
Convert the band signals so that each is measured on the same loudness scale.
be determined. For example, volume equalization converter 1
302 is a 1933 Fletscher and Munsson
(Fletcher and Munson) with some changes.
By
can be imaged. Figure 13 has been changed to the above study.
Shows the result of adding . According to Figure 13, at 40dB
A 1KHz tone has a volume level of 60dB compared to a 100Hz tone.
It turns out that it corresponds to Bell. The volume equalization converter 1302 converts the music shown in FIG.
Adjust the volume according to the line, equal regardless of frequency
produces a volume of In addition to frequency dependence, Figure 13 shows that
As is clear from examining the wave number, the change in power is
Does not respond to changes in volume. That is, the intensity of the sound,
That is, amplitude fluctuations are perceived at all points.
is not reflected in similar changes in volume. for example,
At a frequency of 100Hz, 10dB around 110dB
The perceived volume change is around 20dB.
Much larger than the 10dB perceived volume change.
This difference is due to volume compression, which compresses the volume in a predetermined way.
Processed by device 1304. Volume compression device 13
04 is the volume amplitude measurement value in phon units.
By replacing the power P with its cube root
P^1/3can be compressed into Figure 14 shows known phone pairs determined empirically.
Shows the relationship between Thorns. With the use of sone units,
The model of the present invention is almost correct even with large audio signal amplitude.
Maintain a solid state. 1 sone is a tone of 1KHz.
The volume level is specified as 40dB. FIG. 12 shows a new time-varying response device 13.
06 is shown. This device has each critical frequency
Band-related volume equalization and volume compression signals
Works better. In detail, the tested frequency
For every few bands, the neural firing rate f changes each time frame.
It can be determined by The firing rate f is the acoustic process of the present invention.
It is defined as follows according to the f=(So+DL)n (1) However, n is the amount of neurotransmitter; So is the acoustic wave
Spontaneous firing related to neural firing independent of shape input
fire constant; L is the volume measurement; D is the displacement constant.
So・n is a spontaneous phenomenon that occurs regardless of the presence or absence of acoustic wave input.
Corresponds to the spontaneous neural firing rate, and DLn corresponds to the acoustic wave input.
This corresponds to the firing rate of An important point is that in the present invention, the value of n is determined by the following formula:
By having the characteristic of changing over time,
be. dn/dt=Ao−(So+Sh+DL)n (2) where Ao is the recruitment constant; Sh is the spontaneous neurotransmission
is the material attenuation constant. The new relationship shown in equation (2)
is that neurotransmitters are not produced at a certain rate at Ao.
(a) Decay (Sh・n), (b) Spontaneous firing (So・
n), and (c) neural firing due to acoustic wave input (DL・
n) is taken into consideration. these
The modeled phenomenon is located at the location shown in Figure 11.
Assume that this occurs. As is clear from equation (2), the following amounts of neurotransmitters and
and the next firing rate is at least as high as the current amount of neurotransmitter.
It is proportional to the square of the sound processor of the present invention.
This shows the fact that it is nonlinear. Sunawa
The amount of neurotransmitter in state (t+Δt) is
Amount of neurotransmitter in state (t+dn/dt・Δt)
be equivalent to. Then, n(t+Δt)=n(t)+(dn/dt)・Δt (3) holds true. Equations (1), (2) and (3) describe the operation of the time-varying signal analyzer.
represent A time-varying signal analyzer shows how the auditory system changes over time.
It is adaptive and the auditory nerve signal is non-directive with the acoustic wave input.
It shows the fact that they are linearly related.
By the way, the acoustic processor of the present invention is based on the nervous system.
to better follow the apparent temporal changes in
The best way to perform nonlinear signal processing in speech recognition systems
This is the first model offered. Reduce the number of unknown terms in equations (1) and (2)
Therefore, in the present invention, the following
Use the formula. So+Sh+DL=1/T (4) However, T is the audio wave input generated
After that, the auditory response drops to 37% of its maximum value.
This is the measured value of the time it takes to complete the process. T is a function of volume
With the sound processor of the present invention, various
Known display of volume level response attenuation
Take it out from the graph. That is, a tone of constant volume
When a response is generated, initially a high level response is generated.
After that, the response is determined by the time constant T.
and decays toward steady-state levels. sound
If there is no acoustic wave input, T=T₀It is. This is 50
It is about milliseconds. Volume is L_naxIf T=
T_naxIt is. This is about 30 milliseconds. Ao=
By setting it to 1, 1/(So+Sh) becomes
If L=0, it is determined to be 5 centiseconds. L is
L_naxSo, L_nax= In the case of 20 zones, the following formula holds true. So+Sh+D(20)=1/30 (5) According to the above data and formula, So and Sh are
It is determined by equations (6) and (7) shown below. So＝DL_nax/ [R+(DL_naxT₀R)-1〕 (6) Sh=1/T₀−So (7) however, R=f steady state-L_nax/f steady state-L=0 (8) f Steady state is when dn/dt is 0, given volume
represents the firing rate at R is the only variable left in the sound processor
It is. Therefore, the performance of this processor is R
Just change it and it will change. That is, R is the performance
One parameter that can be adjusted to change the
meter, typically stable against transient effects
This means minimizing the effects of similar
Output pattern is inconsistent when using audio input
This is generally due to differences in frequency response,
difference, background noise as well as (steady state of the audio signal)
Distortion (affecting parts but not transient parts)
Minimize the steady-state effect because it is caused by
It is desirable to do so. The value of R is perfect speech recognition.
Configure your system to optimize error rates
This is desirable. The optimum found in this way
The value is R=1.5. In that case, the values of So and Sh
are 0.0888 and 0.11111 respectively, and the value of D
gives 0.00666. Figure 15 shows the operation of the audio processor according to the present invention.
This is a flowchart. Preferably sample at 20KHz
daisi during a 25.6 ms time frame.
The talized audio passes through a Hanning window 1320 and its
The output of the DFT1322 is at 10 ms intervals.
It is desirable to perform a double Fourier transform. conversion
The output is filtered in block 1324 and includes at least
One frequency band (preferably all critical frequencies)
several bands or at least 20 bands)
Each provides a power density output. Next, power
- Density is block 1326, recorded size
is converted to volume level. This operation is the first
This can be easily carried out by modifying the graph in Figure 3. So
Summary of the process after (limitations of block 1330)
(including value updates) is shown in FIG. In FIG. 16, first, the filtered frequency
Each sensory limit T of band m_fand audible limit T_o
are set to be 120dB and 0dB, respectively.
(block 1340). Then the voice cowl
total frame register and histogram
Reset system registers (block 134)
2). Each histogram contains bins,
(in a given frequency band)
or similar measurements within their respective ranges.
represents the number of samples, or count. present invention
Then, the histogram (for each given frequency band)
) the volume is within each of multiple volume ranges
Preferably, it represents the number of centiseconds of the period. example
For example, in the third frequency band, the 10dB and 20dB
The time between waves may be 20 centiseconds. Similarly, the
20 frequency bands, between 50dB and 60dB, total
If there are 150 centiseconds out of a total of 1000 centiseconds
There is. Total number of samples (i.e. centiseconds) and
and the percentile is taken from the count contained in the bin.
be done. Block 1344 identifies each frequency band.
The frames at the output of the filter are examined and blocked.
1346, the appropriate histogram (per filter)
1) The bins inside are incremented. block 1348
, the total number of bins whose amplitude exceeds 55 dB is the filter
data (i.e., frequency band), and
Determine the number of filters that indicate the presence of voices. block
1350, with a minimum number (e.g. 20
If there is no filter in 6), block 1
At 344, the next frame is inspected. existence of voice
If there are enough filters to show, block 135
2, increment the voice counter. voice counter
The audio appears in block 1354 for 10 seconds and the block
New T with Lock 1356_fand T_hThe value of
is incremented until determined for each router. new T for given filter_fand T_hThe value of is
It is determined as follows. T_f, the maximum of 1000 bins
dB of the bin holding the 35th sample from the top
The value (i.e. the 96.5th percentile of volume) is
BIN_His defined as T_fis T_f＝BIN_HSet to +40dB
be done. T_hIf , from the lowest bin (0.01)
Hold the (total number of bins - audio count)th value
The dB value of the bin is BIN_Lis defined as That is,
BIN_Lis classified as audio in the histogram.
bins of 1% of the number of samples excluding those
Ru. T_his T_h＝BIN_LDefined as −30dB. Blocks 1330 and 1332 of FIG.
, the amplitude of the sound updates the limit value as mentioned above
and changes per son based on the updated limit value.
converted and compressed. Introducing the son unit and compressing
An alternative method is to use the filter (after the bin has been incremented).
Take the amplitude "a" and convert it to dB using the following formula. a^dB=20log_Ten(a)−10 (9) Next, each of the filter amplitudes is equivalent to
Range between 0dB and 120dB to give a volume of
compressed into a^eql=120(a^dB−T_h)/(T_f−T_h) (Ten) Then a^eqlis the volume level (phone unit) using the following formula.
) to an approximate value of the volume in son units.
It is desirable to map a 1KHz signal to 1 at 40dB.
Delicious. L^dB=(a^eql-30)/4 (11) Next, the approximate amount L of the volume in units of son_sis given by the following equation
available. L_s=10(L^dB)/20 (12) At step 1334, L_sis the input of equations (1) and (2)
block 1355.
Determine the output firing rate f for each command. 22 frequency bar
In the case of a 22-dimensional vector
Characterize the acoustic wave input over a frame. deer
Generally, 20 frequency bands are available by email while
Uses a normal scaled filter bank.
and inspect it. Block 1336 processes the next time frame.
In block 1337, the “next state” of n is
Determined according to equation (3). The acoustic processor described above has a firing rate f and a neural
If the DC pedestal has a large amount of transmitter n,
Improvements are needed regarding the use of Sunawa
Therefore, the dynamic range of the terms in the equation of f and n is
If important, derive the formula below to determine the pedestal height
lower. Steady state and no acoustic wave input signal present
(L=0), Equation (2) is expressed as follows in the stable state.
can be solved for the partial state n′. n′=A/(So+Sh) (13) The internal state of the amount of neurotransmitter n(t) is as follows
are shown as steady state part and fluctuating part as
Ru. n(t)=n′+n″(t) (14) Combining equations (1) and (14) yields
Fire rate is obtained. f(t)=(So+D・L)(n′+n″(t))(15
) The term So・n′ is a constant, but all other terms
is the variable part of n or expressed by (D・L)
contains the passed input signal. Subsequent processing is done using the output vector.
Since it is only related to the square of the difference between tors, the constant term is
It will be ignored. From equations (15) and (13), we get the following equation:
It will be done. f″(t)=(So+D・L)・[{n″(t)D・L・A} /(So+Sh) (16) Considering equation (3), the “next state” is as follows.
Ru. n(t+Δt)=n′(t+Δt) +n″(t+Δt) (17) n(t+Δt)=n″(t)+A−(So+ Sh+D・L)・(n′+n″(t))(18) n(t+Δt)=n″(t)−(Sh・n″(t) −(So＋Ao・L^A)・n″(t) −(Ao・L^A・D)/(So+Sh) +Ao−(So・Ao)+(Sh・Ao)) /(So+Sh) (19) Equation (19) becomes as follows if all constant terms are ignored.
I'm going to growl. n″(t+Δt)=n″(t)(1−So・Δt) −f″(t) (20) Equations (15) and (20) are calculated for each 10 ms
Output expression applied to each filter during the time frame
and configure the state update expression. Using these expressions
The result is a vector of 20 elements every 10 ms,
Each element of this vector is scaled by mel
each frequency in the filter bank
Corresponds to the firing rate of the band. Regarding the embodiment described above, the flowchart in FIG.
Special case of fire rate f and “next state” n(t+Δt)
Equations (11) and (16) that define the equations of
Therefore, we can put the expressions of f, dn/dt and n(t+Δt).
This applies except for changing it. A unique value for each equation term (i.e., t₀=
5csec, t_L=3csec, Ao=1, R=1.5 and L_nax
= 20) can be set to other values, So, Sh
and D terms are set to different values when other terms are set to different values.
Then, the respective desired values are 0.0888, 0.11111,
and will be a different value from 0.00666. The present invention can be applied to various software or hardware.
It can be implemented by a. F1c precision matching (Fig. 6, Fig. 17) Figure 6 shows a precision matching phoneme machine as an example.
2000 is shown. Phonetic type matching machines
is a stochastic finite state machine, (a) Multiple states S_i; (b) Multiple transitions tr(Sj−Si): A certain transition has different
Between states, some transitions transition between the same states, and each
Transitions have corresponding probabilities; (c) Corresponding labels for each label that can be generated by a specific transition.
actual label probability It is characterized by having the following. In Figure 6, seven states S₁~S₇and 13 history
Transfer tr1 to tr13 are precision matching phoneme machine 20
00, and the three transitions tr11, tr
The paths of tr12 and tr13 are shown with dashed lines.
At each of these three transitions, the phoneme produces a label.
can change from one state to another without achieving
Ru. Therefore, such a transition is called a null transition.
Ru. Generate labels along transitions tr1 to tr10.
can be done. In detail, transition tr1 to tr
At least one label along each of the 10
There may be unique probabilities that are generated.
Ru. For each transition, the system can generate
There is a probability associated with each label. In other words, too
can be selectively generated by an acoustic channel.
If there are 200 possible labels, each transition (not null)
The associated “actual label probability” is 200
each of which has a corresponding label for a particular transition
corresponds to the probability generated by a phoneme. transition tr
The actual label probability of 1 is given by the symbol P, as shown.
and 1 to 200 surrounded by the following brackets.
Represented in columns. Each of these numbers is
Represents a bell. For label 1, precision butting sound
The elementary machine 2000 generates label 1 at transition tr1.
There is a probability P[1]. Various actual label probabilities
is stored in association with the label and the corresponding transition.
It is. label y₁y₂y₃The string of ... is the given phoneme
Presented to the precision matching phoneme machine 2000 corresponding to
Then, a matching procedure is performed. precision butt
The steps related to the phoneme machine are shown in Figure 17.
I will explain. Figure 17 is a trellis diagram of the phoneme machine in Figure 6.
be. As in the case of the phoneme machine, this training
Squirrel diagram is also in state S₁from state S₇null transition to state S₁
from state S₂transition to, and state S₁from state S_Fourfart
shows the transition of Transitions between other states are also shown.
Ru. In addition, the trellis diagram is horizontally
Indicates the time. Starting probability q₀, and q₁is the phoneme
Time t=t of that phoneme₀or t=t₁to each of
represents the probability of having a start time at . At the start of each
The respective transitions in time are also shown. China
For example, the interval between consecutive start (and end) times
must have a length equal to the label time interval.
desirable. Using precision matching phoneme machine 2000
how much a given phoneme is in the label of the incoming string
When determining whether to match closely,
Search the end time distribution of a phoneme and find the end time distribution of that phoneme.
Used to determine the timing value. End time distribution
The method to perform precision matching depends on
All sounds described in this invention with respect to the singing procedure
This is common to the bare machine embodiments. Precise pine chin
When generating the end time distribution for performing the
Dense matching phoneme machine 2000 is accurate and complex.
requires a lot of calculation. First, according to the trellis diagram of FIG. 17, time t
=t₀required to get the start and end times in
Find out about calculations. The phoneme map shown in Figure 6
In the case of the example of the
Ru. Pr(S₇,t=t₀)=q₀・T (1 → 7) +Pr(S₂,t=t₀)・T(2→7) +Pr(S₃,t=t₀)・T(3→7) (21) However, Pr represents the probability, and T is the 2 in parentheses.
represents the transition probability between two states. This formula is t
=t₀There are three states in which the end time can be
The probability of each is shown. Furthermore, t=t₀end time of
is state S₇limited to current occurrences. Next, end time t=t₁When we examine the state S₁other than
calculations for all states of
No. state S₁starts at the end time of the previous phoneme.
Ru. For convenience of explanation, state S_FourShow only calculations for
vinegar. S_FourIn this case, the calculation becomes: Pr(S_Four,t=t₁)=Pr(S₁,t=t₀)・ T(1→4)・Pr(y₁,1→4) +Pr(S_Four,t=t₀)・T(4→4)・ Pr(y₁,4→4) (22) Equation (22) is t=t₁The phoneme machine is in state S_FourAnd
The probability that (a) Time t=t₀In state S₁With probability that state S₁mosquito
state S_FourMultiplying the transition probability to
A given label −y− in the string is in state S₁mosquito
state S_FourThe value obtained by multiplying the probability of transition to (b) Time t=t₀In state S_FourWith probability that state S_Fourmosquito
is multiplied by the transition probability to itself, and further, the state S_Four
as something that transitions from to itself.
obtained by multiplying the probability of generating the given label −y−
value and Show that it is determined by the sum of Similarly, (state S₁(excluding) other conditions
calculation is also performed, and the phoneme is determined at time t=t₁in certain conditions
generate the corresponding probability that is the state. In general, given
When determining the probability of being in the target state at the time of
The matching is (a) Each state and state before the transition leading to the target state
and the respective probabilities of each of the previous states; (b) for each of said previous states, its label string;
Each of the previous states and the current
Labels that must be generated on transitions between states
Recognize the value representing the probability of (c) that representing the probability and label probability of each previous state;
Combine each value and target by corresponding transition
gives the probability of the state. The overall probability of being in the target state is
is determined from the target state probabilities for all transitions.
state S₇The calculations for the three null transitions
term, whose phoneme is in state S₇For phonemes ending in
time t=t₁Allows you to start and end with
Ru. Time t=t₀and t=t₁determine the probability for
Determination of probabilities for other pairs of end times, such as when
The determination shall be made in such a way as to form an end time distribution.
is desirable. The value of the end time distribution of a given phoneme is
How well a given phoneme matches the arriving label
Displays whether the How well words match incoming labels
the phoneme that represents the word.
are processed sequentially. Each phoneme is the end time of the probability value
Generate a distribution. The phoneme matching value is
By summing the time probabilities and taking the logarithm of that sum,
can be obtained. The start time distribution of the next phoneme is the end time
It is derived by normalizing the distribution. This positive
In normalization, for example, we define each of those values as
scale by dividing by the sum of
Make sure that the sum of the ringed values is 1. Examining a given word or word string
There are at least two ways to determine the number of phonemes h that should be
be. In the depth-first method, the computation follows the basic form
Let's do it. (Successively by each successive phoneme)
calculate subtotals). Given that this subtotal follows
is found to be below a predetermined limit value for the phoneme position of
If so, the calculation ends. Another method, breadth-first method
The method calculates similar phoneme positions in each word.
Do this. The calculation is the count of the first phoneme of each word.
calculation, followed by calculation of the second phoneme of each word.
Do it one after the other. In the breadth-first method, each
The calculated values along the same number of phonemes in words are relative
Compare at the same phoneme position. Either way,
The word with the largest sum of matching values is
That's the target word. Precise matching is done using APAL (Array Processor).
It is realized in the software (assembly language). this
is Floating Point Systems, Inc.
(Floating Point Systems, Inc.) assembly
It is 190L. By the way, precision matching is
The actual label probability (i.e., given the phoneme
probability of producing a given label y at the transition of ), the phoneme
The transition probabilities for each machine and for a given phoneme are
The certainty of being in a given state at a given time after the start time of
Requires considerable memory to remember each of the rates
shall be. The above 190L is the end time, preferably the end time.
Matching value based on the logarithmic sum of completion time probabilities,
start time based on the end time probability generated in
and the matching values of sequential phonemes in the word.
Calculate each word matching score
It will be set up as shown below. In addition, precision pine
Ching calculates the tail probability of the matching procedure
This is desirable. The tail probability is independent of the word.
Measure the likelihood of consecutive labels. Simple example
Then the given tail probability follows one more label
Corresponds to the likelihood of the label. This likelihood is, for example,
A string of labels generated by a sample audio
Easily determined from the ring. Therefore, in precise matching, the basic form,
Coff model statistics, including tail probabilities.
Equipped with sufficient storage devices. Each word has about 10 sounds
In the case of a vocabulary of 5000 words including elements, the basic form is
Requires 5000x10 memory. (Makes a mark for each phoneme.)
70 distinct phonemes (with Lukov model), 200
separate labels, and any labels that generate
If there are 10 transitions with probability, the statistic is 70×
Requires 10 x 200 storage locations
Become. However, the phoneme machine has three parts.
(start part, middle part and end part)
It is desirable that the statistical tables correspond to this.
(one of the three self-loops is included in the continuous part)
It is desirable that ) Therefore, the memory requirement is 70×
Reduced to 3 x 200. For the tail probability, 200×
200 storage locations are required. this array
So 50K integer and 82K floating point storage
It works fine if you have the space. Furthermore, traditional systems contain 70 different phonemes.
However, the present invention uses each phoneme machine.
provides approximately 96 phonemes. F1d Basic high-speed matching (Figures 18 to 20)
figure) Is precision matching calculations expensive?
reduces the required computations without sacrificing too much accuracy.
Basic high-speed matching and alternative high-speed matching
Do ching. High-speed matching is precision matching
It is recommended to use it in conjunction with fast pine
Ching selects promising candidate words from the vocabulary.
and put it on the list, precision matching is often
If so, it is executed on the candidate word for this lift. The fast approximate acoustic matching method is based on the above-mentioned U.S. patent.
Described in Application No. 06/672974 (filed on November 19, 1984)
has been done. In fast approximate acoustic matching, each sound
A phoneme machine is a phoneme machine with all phoneme machines in a given phoneme machine.
Transition sets the actual label probability for each label at a particular position.
It is desirable to simplify it by replacing it with a replacement value.
Delicious. A specific replacement value uses that replacement value.
If the matching value of a given phoneme is
Precision when the replacement value does not replace the actual label probability
Excessive matching value obtained by dense matching
It is desirable to choose to evaluate. This article
One way to guarantee that in a given phoneme machine
If the probability corresponding to a given label of is any of its permutations
Choose each replacement value so that it is no larger than the set value.
This is a method of choosing. Actual labels in the phoneme machine
By replacing the probabilities with the corresponding replacement values,
requirements for determining word matching scores.
The amount of calculation can be significantly reduced. More replacement
Since it is desirable to overestimate the value of
The matching scores obtained are determined without previous substitution.
less than the specified case. Acoustic with language decoder with Markov model
In a particular embodiment of performing matching, each
Through shaping, phonemes are (a) multiple states and transition paths between states; (b) probabilities T(i→j) - each of which is the current state S_i
is given, the state S_jrepresents the probability of transition to
(However, S_iand S_jmay be in the same state or different.
The transition tr(S_j
-S_i), (c) Actual label probabilities (each actual label probability
p(y_k-i→j) by the given phoneme machine,
In a given transition from one state to the next
label y_k(k is a symbol that identifies the label)
(representing the probability that
Ru. Each phoneme machine is (a) Each y in each of the above phoneme machines_kone specific value for
p′(y_k), (b) At each transition in a given phoneme machine, each actual
Output probability p(y_k-i→j), the corresponding y_kdivided into
One particular value p′(y_k) replace it with
means including. The replacement value must be at least a specific phoneme map.
y corresponding to any transition in thin_klabel actual
It is desirable that the magnitude is the maximum label probability.
Fast matching implementation corresponds to incoming labels
10 words chosen as most likely to occur in the vocabulary
We will form a lift of about 100 candidate words.
used in sea urchins. Candidate words are determined by language model and
It is desirable to submit to precise matching. spirit
The number of words considered in tight matching is determined by
The calculation cost is reduced by cutting down to about 1% of the word.
is significantly reduced while maintaining accuracy. Basic high-speed matching is used for all transitions.
Set the actual label probability for a given label to one value.
Simplify by replacing the given label with the given
It can be generated by the phoneme machine. Sunawa
That is, for a given phoneme machine with a probability that the label occurs
Regardless of the transition in the
Replace with a specific value. This value is at least
of the label that occurs at any transition in a given phoneme machine.
It is desired to be an overestimation of the magnitude of the maximum probability.
Yes. Let the label probability replacement value be
as the maximum value of the actual label probability for a given label
By setting, basic high-speed matching
The generated matching values are at least accurate
Matching values such as those resulting from the use of matching
guaranteed to be the same size. like this
Basically, high-speed matching generally involves matching each phoneme.
overestimates the value of the
Generally selected as a candidate word. precision ma
The words considered as candidates by tuching are basically
Pass according to fast matching. Figure 18 shows the basic high-speed matching phoneme machine 30.
Indicates 00. Labels (symbols and
(also called
Enter the matching phoneme machine 3000. Start time distribution
Enter the name and label string using the precision map described above.
Similar to the input of a tuching phoneme machine. Start time
is sometimes not distributed over multiple times.
but instead, e.g. following a silent interval
It may also represent the exact (phoneme start) time. death
However, if the audio is continuous, the end time
The distribution is based on the start time (as explained in more detail later)
Used to form a distribution. Basic high speed matsutchi
The ng phoneme machine 3000 generates an end time distribution.
At the same time, the characteristics from the generated end time distribution are
Generates a matching value for a given phoneme. Word Ma
The tuching score is based on the constituent phonemes (at least
The sum of the matching values of the first h phoneme of the word)
Defined as FIG. 19 shows the basic high-speed matching calculation. base
This high-speed matching calculation uses start time distribution, phoneme
the number or length of labels generated by
each label y_kThe replacement value p′(y_k)only
is connected with. a given label in a given phoneme machine
all the actual label probabilities of and the corresponding replacement values
By replacing , the basic high-speed matching is
Since we replace the transition probability with the length distribution probability, (given
(can be different for each transition in the phoneme machine)
label probability at a given time and a given state at a given time.
It is no longer necessary to include probabilities in the state. By the way, the length distribution is a precise matching model.
Determined from To explain in detail, the length distribution
For each length, this procedure examines each condition individually.
and determine each transition path for each state.
This is desirable. As a result, the actual state of the inspection
The situation is (a) Given a particular label length, (b) May occur independently of the output along the transition.
Ru. all of a specific length to each destination state
The transition path probabilities are summed and then all
The sums of the objective states are summed and the given length in the distribution
represents the probability of The above steps are for each length
is executed repeatedly. Good matching procedure
Following the form, these calculations are Markov-like
Trelli as the modeling technique is known
This is done with respect to the diagram. along the trellis structure
For transition paths that share a branch, each common branch
You only need to calculate it once, and the result is
Added to each path that contains a common branch. In Figure 19, two restrictions are included as an example.
It is. First, the label generated by the phoneme
The length of each has a probability of 1₀,1₁,1₂and 1₃have
may be 0, 1, 2 or 3. start
Time is also limited, each with probability q₀,q₁,q₂Oyo
biq₃Only four start times with . vinegar
That is, L(1₀,1₁,1₂,1₃) and Q(q₀,q₁，
q₂,q₃) is assumed. These limitations make it difficult for the eye to
The ending distribution of target phonemes is defined as the following formula:
Ru. Φ₀=q₀1₀ Φ₁=q₁1₀+q₀1₁p₁ Φ₂=q₂1₀+q₁1₁p₂+q₀1₂p₁p₂ Φ₃=q₃1₀+q₂1₁p₃+q₁1₂p₂p₃+q₀1₃p₁p₂p₃ Φ_Four=q₃1₁p_Four+q₂1₂p₃p_Four+q₁1₃p₂p₃p_Four Φ_Five=q₃1₂p_Fourp_Five+q₂1₃p₃p_Fourp_Five Φ₆=q₃1₃p_Fourp_Fivep₆ Examining these equations, we find that Φ₃are four start times
It can be seen that it contains terms corresponding to each of the following. So
The first term of is the phoneme at time t=t₃start with and length
0 label (phoneme ends at the same time as it begins)
represents the probability of generating . The second term is that the phoneme is at time t
=t₂, and the label length is 1, and
represents the probability that label 3 is generated by that phoneme.
Was. The third term is that the phoneme is at time t=t₁Start with and
With label length 2 (i.e. labels 2 and 3)
Yes, and labels 2 and 3 are produced by that phoneme.
represents the probability that the result will be achieved. Similarly, the fourth term is the phoneme
Time t=t₀and has a label length of 3.
and three labels 1, 2 and 3 correspond to that phoneme.
represents the probability generated by . Calculations and precision matching required for basic high-speed matching
Comparing the calculations required for
It turns out that it is also relatively easy. By the way,
The value of p′(y) is
remain the same value as in the case of label length probability
It is. Additionally, length and start time limitations
This makes later end time calculations easier. example
If, Φ₆So, the phoneme is at time t=t₃Start with 3 levels.
Bells 4, 5 and 6 all sound at their end time.
It must be generated and used by the elements. When generating matching values for target phonemes,
The end time probabilities along the given end time distribution are summed.
It will be done. If necessary, take its logarithm as shown in the following equation. Matching value = log_Ten(Φ₀＋……＋Φ₆) As mentioned above, word matching scores are
The matching values of consecutive phonemes in a given word are combined.
easily determined by measuring Next, we will explain the generation of start time distribution using Figure 20.
I will explain. In Figure 20a, the word
THE₁is broken down into its constituent phonemes and repeated.
In Figure 20b, the string of labels is on the time axis.
shown along. Figure 20c is the first start
Shows time distribution. The initial start time distribution is (silence
(in preceding words, which may contain words)
drawn from the end time distribution of the new preceding phoneme.
Ru. Label input and start time distribution in Figure 20c
Based on the end time distribution of phoneme DH Φ_DHis generated
(Figure 20d). Start time of next phoneme UH1
The distribution is such that the previous phoneme ending distribution is the limit value in Figure 20d.
A is determined by recognizing the time when it happened.
Ru. A is determined individually for each end time distribution.
A is a function of the sum of the end time distribution values of the target phoneme.
be. Therefore, the interval between time a and time b is the phoneme
Represents the time at which the UH1 start time distribution is set.
In Figure 20e, the interval between time c and time d is
The end time distribution of the phoneme DH exceeds the limit value A, and
Corresponds to the time when the start time distribution of the next phoneme is set.
do. The value of the start time distribution is, for example, the limit value A.
Divide each end time value by the sum of end times that exceed
Obtained by normalizing the time distribution. The basic high-speed matching phoneme machine 3000 is
Written by Floating Point Systems, Inc.
Executed using assembler 190L using APAL program.
It is being expressed. Also, according to the description herein,
books using other hardware and software.
It is also possible to develop specific forms of the invention. F1e Alternative high-speed matching (Fig. 21, 22
figure) Alone or preferably with precision matching and
Basic high-speed Matsuchin used with language models
This greatly reduces computational requirements. Calculation required
In order to further reduce the amount, the present invention further includes two
Length (minimum length L_nioand maximum length L_nax) evenly between
Precise pine printing by forming a label length distribution
Simplify Ching. In basic high-speed matching,
Labels of given length (i.e. 1₀,1₁,1₂etc)
The probabilities of producing generally obtain different values. alternative high
By fast matching, the probability of each length of the label
Replace with one uniform value. The minimum value has a non-zero probability in the initial length distribution.
is preferably equal to the minimum length to
Other lengths can also be selected. maximum
The choice of length is more arbitrary than the choice of minimum length, but
The probability of the length being less than the smallest and greater than the maximum is 0
is set to The probability of the length is between the minimum and maximum length
uniformly by setting it to exist only in
The pseudo-distribution of can be shown. one way
Therefore, the uniform probability is expressed as the average probability due to the pseudo distribution.
Can be set. As an alternative, uniform accuracy
The rate is set as the maximum value of the length probability, and the uniform value and
Can be replaced. By making the probabilities of all label lengths equal
The effect of
This can be easily recognized from the equation for the end time distribution. In detail
In other words, the probability of length can be taken as a constant.
can. L_nioset to 0 and the probabilities of all lengths
at the end by replacing the value of one constant with the value of one constant
The time distribution is displayed as follows. Θ_n=Φ_n/1=q_n+Φ_n+p_n (twenty three) However, "1" is one uniform replacement value,
p_nThe value of is the location generated at time m for a given phoneme.
It is desirable to correspond to a replacement value for a given label.
stomach. The aforementioned Θ_nFor the expression, the matching value is
defined as Matching value = log_Ten(Θ₀+Θ₁+…+Θ_n) +log_Ten(1) (24) Comparison of basic high-speed matching and alternative high-speed matching
Compared to the alternative, the number of additions and multiplications required is
By using the matching phoneme machine,
Width decreases. L_nioIf = 0, basic high speed pine
Ching must consider the probability of length
So, it required 40 multiplications and 20 additions, but
For alternative fast matching, Θ_nis determined repeatedly
Therefore, the continuous Θ_nonce for each of
It turns out that we only need one multiplication and one addition. Figures 21 and 22 show the second high-speed Matsuchin.
The simplification of the calculation by the algorithm is shown in detail. Figure 21a
is the minimum length L_nioPhoneme machine 310 corresponding to =0
An example of 0 is shown below. The maximum length is equal to the length distribution.
Assume that it is infinite so that Figure 21b shows the sound
A trellis diagram resulting from an elementary machine 3100 is shown.
q_oAssuming subsequent start times are outside the start time distribution
Then, if m<n, continuous Θ_neach decision of
All require one addition and one multiplication. So
If you want to determine the end time after
is sufficient, and no addition is necessary. Figure 22a shows the minimum length L_nioThe specific case when = 4
An embodiment of the phoneme machine 3200 is shown in FIG. 22b.
shows the corresponding trellis diagram. L_nio=4
Therefore, the trellis diagram in FIG. 22 has symbols U,
It produces a zero probability along the V, W and Z paths.
Θ_Fourand Θ_oFor an end time between 4 times and 1
It is necessary to add times. End greater than n+4
For time, only one multiplication is required, no additions.
It is essential. This example is on the FPS 190L.
This is realized using APAL code. The desired additional state is shown in Figure 21 or Figure 22.
It can be added to the example. For example, L_nioThe value of the
Any number of states with null transitions without changing
can be included. F1f Matching based on first J level (first
Figure 22) Basic high-speed matching and alternative high-speed matching
In order to further improve the
Consider only the matching of the first J label of the
Do it like this. Sound Pro whose label is Sound Channel
Rate of 1 label every centisecond by Setsa
A reasonable value for J is
It is 100. In other words, it corresponds to approximately 1 second of audio.
A label is supplied to the phoneme and a label that enters the phoneme machine.
Confirm matching with Bell. Label to inspect
By limiting the number, two advantages can be obtained.
Ru. The first is a reduction in decoding delay, and the second is a reduction in decoding delay.
Compare long word scores with long word scores
The problem can be avoided to a large extent. of course,
The length of J can be changed as desired. Effects of limiting the number of labels inspected
is observed using the trellis diagram in Figure 22b.
I can do it. Without the improvement according to the present invention,
The fast matching scores are in the bottom row of this drawing.
Along Θ_nis the sum of the probabilities of That is, t=t₀
(L_nio= 0) or t=t_Four(L_nio= 4)
State S at each time starting at_FourThe probability that Θ_nand
and then all Θ_nare summed
Ru. L_nioIf = 4, t_Fourstate at any previous time
S_FourThe probability that it is is 0. With the above improvement, Θ_n
The summing ends at time J. Figure 22
In b, time J is time t_o+2corresponds to The inspection of J labels beyond the interval up to time J is completed.
When determining the matching score by
yields the sum of the following two probabilities. First, before
As mentioned, along the bottom row of this trellis diagram
There is a line calculation. However, this calculation is performed at time J-1
That's it. State S at each time up to time J-1_Fourin
Certain probabilities are summed to obtain a row score. Second, that
The phoneme of S at time J₀~S_Fourare in each state of
There is a column score that corresponds to the sum of probabilities. This column score is
It is calculated as follows. Column score =_Four 〓^f=0 Pr(S_f, J) (25) The phoneme matching score is the sum of the row score and column score.
It is obtained by summing the sum and taking the logarithm of the sum.
Ru. To continue fast matching of the next phoneme,
Along the bottom row (preferably including time J)
Using the values obtained, extract the start time distribution of the next phoneme.
vinegar. Determine the matching score for each of the J consecutive phonemes.
Then, as mentioned above, the sum of all phonemes is that phoneme
is the sum of all matching scores. Basic high-speed matching and alternative high-speed matching described above
How to generate end time probabilities in the Ching example
When we examine
It turns out that it does not easily fit the calculation. Label to inspect
The above-mentioned high-speed matching is an improvement to limit the number of files.
and to better accommodate alternative matching.
Therefore, the present invention allows replacing column scores with additional row scores.
and make it possible. That is, (in Figure 22b) when
State S between time J and J+K_FourAdding a phoneme that is
The line score is determined. However, K is any phoneme
This is the maximum number of states in thin. Therefore, optional
If the phoneme machine 10 has the state of
Along the bottom row of that trellis diagram, the refinement
10 end times are added, each of which is confirmed.
rate is determined. In the lowest row up to time J+K
All probabilities along the line (including the probability at time J+K)
) are added to give the matching score for a given phoneme.
generate. As mentioned above, matching consecutive phonemes
The matching values are summed to obtain the word matching score.
Ru. This example is an APAL on the FPS 190L mentioned above.
This is realized in code, but other
other components on other hardware, such as when
It can also be realized by a code. F1g phoneme tree structure and high-speed matching example
(Figure 23) Basic fast matching or alternative fast matching
(with or without a max label limit
) to determine the phoneme matching value.
The calculation time required to determine the results is significantly reduced.
Furthermore, use the words in the list obtained through high-speed matching to
Even when performing dense matching, the computational complexity is
Significant savings. Once the phoneme matching value is determined,
As shown in FIG. 23, the branches of the tree structure 4100
A comparison is made along the path to determine which path of phonemes occurs most.
Determine if it is possible to break. In Figure 23, (point
4102 to branch 4104)
The phoneme DH and UH1 of the phoneme “the”
The sum of the values for each phoneme branching from the phoneme MX is
must be significantly higher than in the case of the other sequences.
Must be. By the way, the phoneme of the first phoneme MX
The matching value is calculated only once and then spread out.
used for each basic form. (branch 4104 and
and 4106. ) Furthermore, the beginning of the branch
The total score calculated along the sequence of
lower than the limit or other seams in the branch.
It can be seen that the total score is lower than Kensu's total score.
and all base forms extending from the first sequence
may be removed from the candidate words at the same time.
Ru. For example, related to branches 4108-4118
The basic form is that MX is not a promising path.
If it is determined, it will be discarded at the same time. High-speed matching example and tree structure make it possible to
A list of candidate words in fixed order is created and
The computation involved is significantly saved. Regarding memory requirements, phoneme tree structure, phoneme organization,
The measured value and the tail probability are to be memorized.
There is. For tree structure, 25000 arcs and each arc
There are four data words that characterize the first
The data word represents the index of the subsequent arc or phoneme.
Was. The second data word is the subsequent data word along the branch.
Represents the number of phonemes. The third data word is a tree structure
Indicates which node the arc is placed on. Fourth
The data word represents the current phoneme. Therefore,
This tree structure requires 25000 x 4 storage spaces.
It is. 100 different sounds with fast matching
There are over 200 different finems. fee
A neem is a probability of being produced somewhere in a phoneme.
has a storage space of 100×200 statistical probabilities
is necessary. For the trailing structure, 200×200
Requires storage space. Therefore, fast matching
In this case, 100K of space to store integers and 60K of float
It is sufficient to have space to store the floating decimal point. F1h language model (Figure 4, Figure 24) As mentioned above, regarding the word in context (triple
Contains a language model that stores information (such as letters)
This increases the probability of correctly selecting the word.
You can The language model is stored in the paper
has been done. Language model 1010 (Figure 4) is a unique character
It is desirable to have i.e. modified triple
law is used. According to the invention, the sample sample
Examine the text and identify triple words in a fixed order in the vocabulary
and the likelihood of each word pair and single word
Confirm. And the most promising Mie work
A list of codes and word pairs is formed. Furthermore,
Triple words not in the list of triple words, and
It is the likelihood of a word pair that is not in the list of word pairs.
Each will be determined. According to the language model, the target word is reduced to 2 words.
If it continues, this target word and the previous two words are
It is determined whether the word is in the list of triple words.
Ru. If it is in the list of triple words, the triple word
Specifies the remembered probability assigned to the
be done. The target word and the two preceding words are triple words.
If it's not in the list, add the target word and
Are adjacent preceding words in the list of word pairs?
judge whether in the list of word pairs.
, the probability of the word pair and the triple word
Multiply the list of words by the probability that there are no triple words and
Assign the product of to the target word. Contains target word
The triple word and word pair are each triple word and word pair.
not in the list of words and the list of word pairs
In this case, we add the probability of only the target word to the triple word
The probability that a word is not in the list of triple words, as well as the word
Multiply the probability that a word pair is not in the list of word pairs and
Assign the product of to the target word. The flowchart 5000 in Figure 24 is used in acoustic matching.
This shows the formatting of the phoneme machine. block 5002
Word vocabulary (generally on the order of 5000 words)
Define. Each word is then assigned a sequence in the phoneme machine.
Displayed by Kens. For example, phoneme machine
is displayed as a phonetic phoneme machine, but
Alternatively, the
Sometimes it happens. Sequence of phonetic phoneme machine,
Or to the sequence of the Finim-type phoneme machine.
The display of words according to the following will be explained below. War
The phoneme machine sequence of ``do'' is a word basic form.
say. In block 5006, the word basic form is
Arrange in a tree structure. Phoneme machining for each basic form of each word
The statistical values for each section are from IEEE Bulletin Volume 64 (1976).
F. Zielinek's paper “Statistical methods
F. Jelinek, “Continuous Speech Recognition” (F. Jelinek, “Continuous Speech Recognition”
“Speech Recognition by Static Methods”
Proceedings of the IEEE, Vol.64, 1976)
Well-known forward and backward als shown
Determined by algorithmic shaping (block
5008). Block 5009 is used for precision matching.
the actual parameter values, i.e. the statistical values
decide. For example, a value instead of the actual label output probability
Confirm. In block 5010, the determined value
will replace the actual memorized probabilities and
Phonemes in the basic form now contain approximate replacement values.
do. All estimates regarding basic high-speed matching
Execution occurs at block 5010. Next, in block 5011, the acoustic matching is
Decide whether you need more. no improvement required
If the base approximation is determined for matching
set the value for use and another regarding other approximations.
No estimated value is set (block 5012). improvement
If necessary, proceed to block 5018.
nothing. In block 5018, the string lengths are equalized.
In block 5020, further improvement is made.
Decide whether you need a top. Need to further improve
If you do not need the label output probability value and
approximate the matching length probability value and use it in acoustic matching.
Set it so that If further improvement is required
produces acoustic matching in block 5022.
limited to the first J label of the created string.
Ru. Whether to select one of the improved embodiments
Determined parameter values are blocked regardless of
5012, so that each word base form
Each phoneme machine in the state is shaped by the desired approximation.
is used to enable fast approximate matching. F1j stack decoder (Figures 7 to 9,
Figure 25) Next, the main speaker used in the speech recognition system shown in Figure 4.
Describe a clear stack decoder.
Ru. Figure 7 shows successive label intervals, i.e.
multiple consecutive labels y generated at file positions<sub>1</sub>y<sub>2</sub>…
…It is shown. Also, FIG. 8 shows a plurality of generated word patterns.
path A, path B and path C are shown.
It is. In relation to Figure 7, path A is the entry
“to be or”, path B is in the entry “twob”,
Path C may correspond to entry "too".
For each target word pass, the target word pass
The label that has the highest probability of being terminated (i.e.
In other words, there is equivalent label spacing), such as
The label is called a boundary label. A word pass W representing a sequence of words.
and the most likely end time - label string
The boundary label between two words in the ring and
− is IBM Technology Disclosure Bulletin Volume 23
L.R.B. No. 4 (September 1980 issue)
Paper outside the scope, “Acoustic matching calculation speed-up”
(L.R.Bahl et al, “Faster Acoustic Match
IBM Technical Disclosure
Bulletin, Vol.23, No.4, September 1980)
found by known methods as described
be able to. Briefly, this paper
Two similar concerns, namely: (a) Which word (or word sequence)
result in as many label strings as (b) At which label spacing, the partial sentence-label space
Does the − corresponding to the string part end? This article discusses ways to address this issue. Any given word pass has a label strip.
each from the first label of the string to the border label
The “likelihood” associated with the label, i.e., the label interval
value ”, i.e. the likelihood of a given word path
All values are the “likelihood vector” for a given word path.
”, so for each word pass,
There is a likelihood vector. Likelihood value L<sub>t</sub>is shown in Figure 8.
It is. Word Pass W<sup>1</sup>,W<sup>2</sup>...W<sup>s</sup>A collection of labels
“Likelihood envelope” Λ with interval t<sub>t</sub>is mathematically the following
It is defined as follows. Λ<sub>t</sub>=max(L<sub>t</sub>(W<sup>1</sup>),...,L<sub>t</sub>(W<sup>s</sup>) That is, for each label interval, the likelihood envelope is
Highest associated with any word pass in the collection
Contains the likelihood value of . The likelihood envelope 1040 is shown in FIG.
It is shown. If the word path corresponds to a complete sentence,
considered “complete”. Please enter the complete path
When a speaker reaches the end of a sentence, e.g.
It is recommended that you press the button to identify it. The input is a sentence
Synchronized with the label interval that marks the end. Perfect
Extend the word password by appending the word
I can't. Partial word passwords are not allowed.
It corresponds to a complete sentence, so it can be extended.
Ru. A partial path can be a “living” path or a “dead” path.
Word Pass is classified as
“dead” when it has already been extended
However, when it has not yet been extended, it is “alive”.
This classification has already been extended by at least one
form a word pass that extends beyond the
The path will not be considered for extension again at
do not have. Each word path is
Characterized as a “good” or “bad” path.
I can do it. The word pass is
The label corresponding to the bell whose word pass is
If you have a likelihood value that is within the maximum likelihood envelope of Δ
is a good word pass. In other cases, the
A bad word pass is a bad word pass. maximum likelihood
It is good (or bad) to reduce each value of the envelope by a certain value Δ.
b) Although the limit level is determined, changing this Δ
Although it is desirable, it is not always necessary.
It's not important. There is a stack element for each label interval.
Ru. Each living word pass looks like this
label corresponding to the boundary label of the living path
Assigned to the stack element corresponding to the interval. vinegar
Tack elements (listed in order of likelihood values)
0, 1 or more word pass entries
It may have. Next, the stack decoder 1002 in FIG.
The steps to be executed will be explained below. As shown in the flowchart in Figure 9, the shape of the likelihood envelope is
The decision on what is a good password and password is mutual.
Involved. In the flowchart of FIG. 9, block 1050
So, first, the null pulse goes to the first stack 0.
enter. Previously confirmed in block 1052
Contains the complete path (if the complete stack element is
If available, it will be provided. in a (complete) stack element
Each complete path of has a likelihood vector associated with it.
It has a tor. has the highest likelihood for that boundary label.
The likelihood vector of the complete path to
Decide on the line. If a (complete) stack element is
If there is no complete path, the maximum likelihood envelope is
is initialized to −∞. Furthermore, the complete path is
Even if the maximum likelihood envelope is not
The period may be set. The initial setting for the envelope is
Locks 1054 and 1056 take place. The maximum likelihood envelope is initialized and then adjusted by a predetermined amount Δ
Δ stipulation above the reduced maximum likelihood
form a good region of and below the reduced likelihood
Forms a region with poor Δ regulation. The value of Δ is the width of the search
control. The larger Δ, the longer the extension.
The number of word passes considered possible increases.
L<sub>t</sub>log to determine<sub>Ten</sub>When using , the value of Δ is
2.0 gives satisfactory results. Δ value
is uniform along the length of the label spacing.
Although desirable, it is not necessary. If the word path is within a good region of delta regulation.
If the field label has a likelihood, its word path
is marked as “good”. In other cases,
Word pass is marked as "bad". As shown in Figure 9, the likelihood envelope is updated and the word
A “good” (extendable) path, or
or as a “bad” path.
Blog about finding the longest unmarked word pass
It starts with 1058. Not marked with 2 or more
short word path corresponds to the longest word path length
If the border label is on the stack
A word path with a likelihood of is selected. War
block 1060 if a path is found.
, the likelihood at that boundary label is a region with good Δ specification
Check whether it is inside. If it's in a good area
If so, in block 1062, the region with poor Δ regulation is detected.
In block 1058, mark the path in
Look for unmarked live paths. too
If it is within the good range, then in block 1064, Δ
Mark the path within the defined good area and block 1
In step 066, update the likelihood envelope and mark it as “good”.
Contains the likelihood value of the tracked path. That is,
For each label interval, the updated likelihood value is (a) the current likelihood value within its likelihood envelope; (b) Associated with word passes marked “good”
likelihood value is determined as the larger likelihood value between. child
The operations are performed in blocks 1064 and 1066.
be exposed. After the envelope has been updated, block 105
Return to 8 and find the longest, best unmarked
Find the password again. This loop uses unmarked word
Iterates until there are no more spaces left. Not marked
If you run out of passwords, block 1070
The shortest word path marked as “good” in
is selected. If 2 or more have the shortest length
If you have a “good” word pass, block 10
72, find the word with the highest likelihood for that boundary label.
The selected shortest path is
It will be extended. That is, at least one prospective
If a trailing word with
matching, language models, precision matching, and
determined by a successful execution of the word model procedure.
It will be done. For each likely subsequent word, the
A word pass is formed. In detail,
The extended word path will be
Append a promising trailing word to the end of the decoded path.
Formed by adding The selected shortest word path is
After forming the code path, the selected word
The path is removed from the stack in which it was an entry.
instead, each extended word
The path is inserted into the appropriate stack. In particular, extension
corresponds to that boundary label.
Entry to stack (block 107)
2). The selected pulse in block 1072
Explain the extending operation in relation to the flowchart in Figure 25.
do. After the path is found in block 1070,
Perform the following steps, which will allow you to
or paths are extended based on appropriate approximate matching.
be done. At block 6000 of FIG. 25, (of FIG. 4)
The sound processor 1004 processes the label as described above.
Generate a string. The label string is
Provided as an input to block 6002,
6002, basic or improved approximate pine
One of the checking steps is performed and the constant
Get a list of candidate words for order. After that, Bro.
6004, convert the language model as described above.
use. After using the language model, block 6
At 006, the remaining target words are generated
Sent along with the label to a precision matching processor.
It will be done. Precise matching with block 6008
yields a list of remaining candidate words and
Well presented to Dell. (approximate matching, precision
determined by dense matching and language model.
) Promising words are in block 1 of Figure 9.
Used to extend the path discovered at 070. Blotsu
The expected
Each word is added to the discovered word path.
Separately appended, multiple extended word paths
(block 6010). In Figure 9, an extension path is formed and the stack is re-established.
Once formed, the process returns to block 1052.
repeat the steps. Therefore, at each iteration, the shortest, best “good”
selected and extended. in some iteration
Word paths marked as “bad” paths are
Repetition can result in a “good” path, so
Is the word/pass you are using “good” or “bad”?
The path characteristic is assigned independently at each iteration.
be done. In fact, the likelihood envelope is one iteration and the next
The word pattern does not change significantly between iterations of
The calculations that determine whether a performance is good or bad should be done effectively.
We also no longer need normalization. If you want to identify a complete sentence, block 1074
It is desirable to include. i.e. alive
What remains unmarked in Word Pass
If there is no “good” word pass that should be extended
If so, the combination ends. each of its boundary labels
The complete word path with the highest likelihood
the most likely word for the power label string.
identified as a code sequence. For continuous speech where sentence ends are not identified, path extension
The length of the system may be continuous or
up to a predetermined number of words desired by the system user.
It will be done. F1k Construction of phonetic basic form Markov models that can be used to form basic forms.
One type of model phoneme machine is one based on phonetic symbols.
It is. That is, each phoneme machine is
Corresponds to single notes. For each given word, its corresponding
A sequence of phonetic phonemes with a phoneme machine of
be. Each phoneme machine has several states and
transitions between them, and some of them include
Those that can produce neem output and those that cannot.
There is also a null transition. As mentioned above, each
Statistics regarding the phoneme machine are: (a) the probability of a given phoneme occurring, and (b) A given transition produces a particular finem.
including the likelihood of For each non-null transition, each
It is desirable to have a probability associated with the neem.
stomach. Finim Alphabets shown in Table 1
There are approximately 200 finems in the set. No.
Figure 6 shows the phonetic alphabet used to form the basic form.
Showing the phoneme machine. The design of such a phoneme machine is
-kens are given for each word. Statistics,
i.e. the probability of uttering a known word is
It can be put into the phoneme machine in Phase. various sounds
Transition probabilities and fees in standard phoneme machines
Neem probability is calculated by using the known phonetic alphabet during shaping.
Fee generated when uttered at least once
Neem string(s),
Well-known forward-backward algorithms
determined by applying the system. Statistics for one phoneme identified as phoneme DH
A sample of is shown in Table 2. As a rough estimate
Therefore, the transitions tr1, tr2 and
Label output probability distribution of tr8, transitions tr3, tr4, tr
5 and tr9 label output probability distribution and transition
Label output probability distribution for moves tr6, tr7 and tr10
are each represented by a single distribution.
this is. In Table 2, each column's arc (i.e. transition
indicated by the assignment of labels 4, 5 or 6 to
has been done. Table 2 shows the beginning, middle and middle of the phoneme DH.
or the probability of each transition generated respectively at the end.
and the probability of the label (i.e., finem).
vinegar. For DH phonemes, for example, the state S<sub>1</sub>from state S<sub>2</sub>
The probability of transitioning to state S is calculated as 0.07243, and the probability of transitioning to state S<sub>1</sub>mosquito
state S<sub>Four</sub>The probability of transitioning to is 0.92757. (what
Then, there are two possible transitions from the initial state.
Therefore, the sum of both probabilities is equal to 1. )
Regarding the label output probability, the DH phoneme is
The last part of the element, that is, the column labeled 6 in Table 2
Confirmation of generating Finim AE13 (see Table 1) with
It has a rate of 0.091. Table 2 also shows each node.
(i.e. state) related counts are shown
There is. The node count is the number of times that phoneme is
Represents the number of times the corresponding state occurred. Table 2
These statistics exist for each phoneme machine.
Ru. Arranging phonetic phoneme machines into word basic form
is generally performed by a phonetician, so
Usually not done automatically. The phonetic type basic form is suitable for precision matching and high-speed outline.
It is successfully used in calculation matching. Phonetic type base
This form depends on the phonetician's judgment and is not automatic.
Therefore, the phonetic basic form is not accurate.
Sometimes. F2 Contains a start phoneme machine and an end phoneme machine.
Formation of a set of phoneme machines (Figures 1A to 3)
Figures 26-32) Phonemes used when constructing basic forms explained in the previous section
The machine is selected from a collection of phoneme machines. Before
As mentioned above, the generation method of conventional speech recognition systems
Now, each sound (or more specifically, each sound
(typical elements) are only relevant to single phoneme machines.
Ta. Each phoneme machine has transitions and their
the probability associated with that transition, as well as the latitude associated with that transition.
Contains the bell output probability. Therefore, the phoneme machine
When the corresponding phonetic sound is uttered,
generate a given label at a given transition of the phoneme machine of
Contains a statistic that indicates the likelihood that This statistic is
The sound processor 1004 (No.
Figure 4) and enter the known forward
Extracted during the shaping period to apply the de-algorithm
It will be done. Most of the statistics retrieved during shaping are known
When the sound is uttered, the sound processor 1004
determined by the label generated by the sound
Label generated by the sound processor 1004
is the energy-related property corresponding to the spoken input.
Determined by The word WILL shown in Figure 26
Spectral photograph and words shown in Figure 27
With the WILL waveform, the “w” sound is stored from a silent state.
The energy characteristics during the product follow the energy storage.
It was found that the energy characteristics are significantly different from the “w” sound.
Ru. Prior to this invention, sounds or phonetic elements were
Occurs at the beginning of or within a word following a period
whether it occurs in the middle or at the end of the word
There was no distinction made between squids. According to the invention, this
A distinction has come to be made between them. Words shown in Figures 26 and 27
The first 0.1 seconds of “WILL” is an accumulation of “w” sounds.
The waveform part immediately after is the shadow caused by silence.
It corresponds to a “w” sound with less resonance. The energy accumulation and subsequent part of the “w” sound is −
Like the generation method of the conventional system - single phoneme and
Treating them as one group may cause errors in the system.
This will result in a loss of performance. In other words, the “w” sound
A one-phoneme machine uses the “w” sound at the beginning of a word,
at the end of a word, and in all cases occurring within a word.
were mixed into the statistics. Therefore, a single note
Elementary machines can store energy as well as attenuate energy.
It contained mixed statistical values. In accordance with the present invention, a given single note - like the "w" sound
has multiple phoneme machines associated with it.
Sometimes. For example, the “w” sound is uttered
Statistics of “w” sound, unaffected by silence
It has a common phoneme machine that contains values. common phoneme
The machine will emit a “w” sound that is not adjacent to a period of silence.
Contains statistics generated by voice. Therefore, common
Phonemes contain effects related to energy storage or decay.
energy characteristics are not mixed in. Furthermore, the “w” sound is
The regulation regarding the production of the “w” sound during the transition from the silent period
Starting phoneme machine that reflects readings, as well as silence
Reflects statistical values regarding the pronunciation of the “w” sound just before the period
It also includes an ending phoneme machine. The “w” sound starting phoneme machine is also ONSETLX
is displayed as ONLX, and the final phoneme machine for the “w” sound is
Display as TRAILLX or TRLX. common phoneme
The machine displays WX. Each phoneme machine is separate
formed into individuals, each with its own probability and
has a label probability. Three things related to the “w” sound
The different statistical values of the phoneme machine are shown in Tables 3 and 4.
and shown in Table 5. In Table 3, the phoneme machine ONLX is
It has statistics configured similarly to the statistics shown in the table. sound
In the first, middle and last sections of the bare machine
The probabilities of generating various labels are shown in three columns.
Ru. The transition probabilities from one state to another are also shown.
ing. Figure 28 shows (like the phoneme machine in Figure 6)
Una) How to classify the transitions of the phoneme machine 3
Indicates whether there will be two sections. The statistics in Table 3 were taken during the plastic surgery period and
This applies to speakers of During formatting, sample known text is used for this story.
uttered by the person. From known text, that
The sequence of phonemes corresponding to the text is determined.
Ru. When a known word is uttered, a label (e.g.
A string of ``Fineem'' is generated.
Labels can be added in the usual way, such as Viterbi alignment.
aligned against the phoneme machine in a sequence according to
It will be done. Generated labels and known text phonemes
The correspondence between is that found in each phoneme machine.
This becomes the basis for determining each probability. example
For example, the “w” sound preceded by silence has already been produced during the shaping period.
It may occur multiple times at certain intervals. “w” sound
a specific label when preceded by silence - e.g.
Processed as many times as WX7− is generated, Table 3
The probability shown in is given. In detail,
The starting phoneme machine for the “w” sound is
The probability of generating label WX7 at the center is 0.036, then
generate a label WX7 at the end of that phoneme machine
has a probability of 0.197. Also, in Table 3,
Transition between state 1 and state 4 of the start phoneme ONLX
The probability is 0.67274, but between state 1 and state 2
The transition probability is 0.32370. The importance of the invention is shown in Tables 3, 4 and 5.
This is obvious if you compare the tables. Tables 4 and 5
is the label output WX7 - shown in Table 3 -
is not included as the primary label output. Change
In Table 5, the transition from state 1 to state 4 is
state 1 to state 2 with rate 1.0 and parallel to it
The probability is 0. These points are the third point mentioned above.
The statistics are significantly different from those in the table. Statistical values shown in Tables 3, 4 and 5
The significant difference between all occurrences of the “w” sound
regardless of its position in the code – all together as a single
By using the phoneme machine's statistical values, errors can be avoided.
Indicates that there is a possibility that the Word bases each containing a sequence of phonemes
When forming a form, these phonemes are divided into predetermined phoneme segments.
selected from. Using the single phoneme machine method
Traditional generation methods (as mentioned above) generate approximately 70
The phoneme was hot. According to the invention, the phoneme set is
26 sounds consisting of 14 starting phonemes and 12 ending phonemes
It is desirable to add elements. Table 6 shows these
Indicates additional phonemes. In Table 6, each sound (i.e. phonetic type element)
has its own start and end phoneme machines.
It has no syn. Such an arrangement is within the scope of the present invention.
210 phoneme machine – 3 phonemes per sound
− The catalog is used when a large amount of formatted data cannot be obtained.
considered too large. Therefore, the sink
Statistics, whether or not adjacent to a quiet period
A constant sound that does not show a significant change in corresponds to
It has only a common phoneme machine. A simple like this
Sound classes include PX, TX and KX. Eliminate these
It's called audio closure. Silent closures are
unaffected, so it is displayed by a single phoneme.
Ru. Furthermore, groups of fixed note magnitude store energy.
Since they have very similar statistics regarding the product, this
One starting phoneme machine for each group such as
I can give it to you. One of them is Table 6, which has eight sounds.
(i.e., the phonetic type element) is associated with the starting phoneme
The machine is ONSETAA, or ONAA. similar
In other words, a certain group of sounds has extreme energy attenuation.
Groups like this have similar statistics.
One ending phoneme machine is given for each group.
Ru. For example, in Table 6, seven sounds are ending phonemes.
TRAILAA, or related to TRAA.
With this classification, the acoustic statistics for that purpose can be determined.
Phoneme machine and shaping data needed to generate
less. From this classification, there are 210 phonemes.
Significant performance compromises for systems using
has not occurred. Table 6 shows the standards corresponding to the identifiers used in the present invention.
The subphonetic symbols are also shown. Special points to note here
, the present invention (identified by the symbols shown)
It is desirable to include some of the phonetic elements of
Other types of sounds other than the International Phonetic Alphabet
is also taken into consideration. In Table 6, phonemes with the suffix “0” are
A sound with the suffix “1” that refers to a vowel without a point.
Plain refers to accented vowels. Next, Table 7 shows that the phoneme machine is a good implementation of the present invention.
Identify all of the phonemes formed according to the examples.
Ru. From the set of phonemes shown in Table 7, the basics of words are
A format is constructed. About the word “WILL” again
Considering this, the basic form is the sound shown in Figure 29.
a sequence of elements (i.e., equivalently a phoneme machine)
formed as a Phonetic symbol of the word “WILL”
Pelling is shown in FIG. phoneme machine
ONLX represents the starting phoneme machine for the "w" sound.
(ONLX phoneme machine uses the “l” or “hw” phoneme
It is also the first phoneme machine of the basic form starting with the element
Ru. ) After the ONLX phoneme machine for the word “WILL”
Phoneme machine corresponding to common phoneme machine for “w” sound
WX continues. After that, IX1 phoneme machine and LX
Followed by a phoneme machine, and a TRLX phoneme machine. Each word in the vocabulary is similarly (shown in Figure 29)
(such as the basic form of the word “WILL”)
Displayed depending on the status. When forming each word, pair
The phonemes containing the elephant words are determined, and then they
The phoneme machines corresponding to the phonemes are concatenated. Inventory stored in computer, each word
is determined by the corresponding sequence of phoneme machines.
statistics are displayed for each phoneme machine.
is memorized. To reduce storage requirements, each
Display phoneme machines by their corresponding identifiers and
The basic form of the code is expressed as a sequence of phoneme machine identifiers.
It can be formed by For example, word
The basic form of “WILL” is a sequence of identifiers: 43
Corresponds to -27-81-12-56. Identifier 43 is a phoneme
It corresponds to the machine ONLX, and the identifier 27 is a phoneme machine.
Compatible with Windows WX. The same applies hereafter. phoneme
For each machine, after a shaping period, a portion of memory is
The statistical values shown in Tables 2 to 5 are memorized.
It will be done. Once the target word is considered, the component phoneme machine
The statistical value of the link identifier is retrieved. Examples of the other two basic forms are shown in Figures 31 and 31.
and the word “BOG” and the word “BOG” in Figure 32.
Each basic form of “DOG” is shown. degree
The basic form also begins with the starting phoneme machine ONBX.
Ru. The word “BOG” is the phoneme following ONBX.
Includes machines BX, AW1, GX and TRBX. Wa
The code “DOG” is a phoneme machine following ONBX.
Includes DX, AW1, GX and TRBX sequences.
nothing. The energy storage of the “B” and “D” sounds is similar.
The same starting phoneme machine is used since
Ru. When shaping the ONBX phoneme machine,
any of the phonemes (i.e. phonetic elements) that are displayed
It is desirable to incorporate utterances into the generation of statistical values.
This condition applies to multiple sounds (i.e. phonetic elements).
various other start and end phoneme machines that correspond to
It is desirable to apply this to machines as well. The basic configuration can be explained using the flowcharts in Figures 1A and 1B.
Explain the steps to build. In block 8002,
Starting Phoneme Machine, Common Phoneme Machine and Ending Phoneme Machine
From the machines a set of phoneme machines is formed.
Next, block 8004 selects words from the word vocabulary.
selected. In block 8006, word
are multiple phonetic elements, or sounds in general,
A predetermined order such as W-I-l for the word “WILL”
Characterized by order. Next, in block 8008,
Examine the first phonetic element in a given order and find its corresponding
Determine whether there is a starting phoneme machine. correspondence
If there is a starting phoneme machine to
10, search for the corresponding starting phoneme machine and block
With Tsuk8012, the first two phoneme machines are
The beginning phoneme machine of the first phonetic element followed by the common
Set it as a phoneme machine. to the first phonetic element
If there is no corresponding starting phoneme machine, block
At 8013, the common phoneme machine is searched. child
The common phoneme machine of represents the beginning of the basic form. Next, in block 8014, there is no next phonetic element.
If so, proceed to block 8015 and select the first phonetic mark.
whether the element has a terminal element associated with it
decide. If there is no final phoneme, its basic form
is the starting phoneme (machine) followed by the common phoneme
(machine). related to the first phonetic element
If there is a final phoneme, block 8016
Then, the final phoneme is added to the common phoneme, so
The word base form is the starting phoneme map of the first phonetic element.
including the common phoneme machine and the end phoneme machine.
nothing. In block 8014, if the following phonetic element is present:
If so, block 8017 examines the next phonetic element.
and determine whether the next phonetic element is last in order.
decide. If it is the last, block 80
18, the end phoneme to which the phonetic element is associated
Decide whether to have a machine. ending phoneme better
If there is a component, block 8020 describes its basic
The morphology is the common phoneme machine corresponding to the first phonetic element.
by adding a final phoneme machine followed by a final phoneme machine.
more complete. There is no associated ending phoneme machine.
If the last phonetic element is
The common phoneme machine is a
It will be done. In block 8017, the next phonetic element is the last
If not, in block 8024, the phonetic element is
Attach a common phoneme to the previously arranged phoneme machine.
Add. Phoneme machine corresponding to the last phonetic element
(there may be more than one) are added consecutively.
A phoneme machine is added, and the sequence of the phoneme machine is
Extend the time. Next, in FIGS. 2A and 2B, according to the present invention,
This section explains the formation of a phoneme machine. First, block
8100, for example, the International Phonetic Alphabet
The sound is defined as a phonetic element selected from
Ru. A collection of sounds is the seeds of single sounds formed by speech.
represents a class. In block 8102, each
a plurality of devices having means for storing statistics regarding the
A phoneme machine is formed. Next block 8104
, we define a given sound as an initial set of sounds, each of which is
to get the starting phoneme machine assigned to it
Select ``Natsuru''. Shadow due to energy storage
The sounds that receive a lot of sound will form the first set.
is desirable. (As mentioned above, sufficient plastic surgery data
If the data is obtained, the first set is
can be formed. ) then block 81
In 06, a starting phoneme machine is assigned to a given sound.
Ru. At block 8108, the assigned starting phoneme
Machine statistics for audio segments (e.g. work
from the first utterance of d) - this utterance is given
A single note corresponding to the note of, i.e. a similar energy storage
It is a single note with product characteristics. Block 8110 then extracts the common sound from the given sound.
Form a raw machine and use block 8112 for that purpose.
Generate statistics for Block 8114 opens
Each note that is supposed to get an initial phoneme machine is
the second phonetic magnitude after being treated as a given note
for a given sound in the set of
) is formed. In block 8116, the given
Select the sound of the ending phoneme in block 8118.
Assign a syn to it. With block 8120,
The statistics of the guessed ending phoneme are added to the speech segment.
Vocalization occurring at the end of - a single sound corresponding to a given sound,
That is, for single notes with similar energy attenuation characteristics,
Generate from utterances. Then block 8122
, a common phoneme machine for a given phone is assigned,
Block 8124 indicates that the statistic was previously determined.
If not, statistics are generated. Block 8
126, all sounds (the ending phoneme machine
) is to be assigned a given phoneme and
to determine whether it has been selected. selected
If so, all phoneme machines are formed.
Ru. If not selected, the previously unselected
Select the above sound as the given sound and select the above blot.
8118-8126 are repeated. FIGS. 2A and 2B show that according to the present invention
It can be modified in various ways. First, open
If you want to search only the first phoneme machine, go to block 8.
116 to 8126 can be omitted. similar
If you want to explore only the ending phoneme machine, use Blotsu.
8104 to 8114 can be omitted.
Second, if you wish, the first set of notes and the second
A set of sounds can be generated simultaneously. In addition, a single starting phoneme machine or ending phoneme machine
The operation step of assigning a sound to two or more sounds is
Relating to these examples. In this case, the statistical value
only needs to be generated once and used appropriately for each sound.
do. First, start which sound is assigned to it
Should I get a phoneme machine and an ending phoneme machine?
and thereby block 8104 and
and the first and second set of 8116
It is desirable to form each. In speech recognition, the present invention reduces the number of phoneme machines to
An apparatus for forming an increased basic form is provided. this
An example of the device is shown in FIG. FIG. 3 shows a plurality of phoneme machines 8202 to 821.
2 is shown. Each phoneme machine is phoneme machine 8
202, each with (a) transition certainty.
rate memory 8214, (b) label probability memory 821
6, and (c) state identifier and transition identifier memo.
Contains 8218. Phoneme machines 8202 and 8
Multiple phoneme machines including 204 are common phoneme machines
and includes phoneme machines 8206 and 8208.
The multiple phoneme machine that includes is the starting phoneme machine. Ma
In addition, a complex system including phoneme machines 8210 and 8212
The number phoneme machine is a termination phoneme machine. each phoneme
In each memory of machines 8202 to 8212
is the statistical value by the phoneme machine shaping device 8220.
is memorized. Each word is predefined as a sequence of phonemes.
and these sequences are stored in storage device 823.
0. Basic format construction device 8240
is phoneme sequence information from storage device 8230
and extracted by the phoneme machine shaping device 8220.
By combining the statistical values obtained, the phoneme machine sequence
Build. Sequence of phoneme machine for given word
The sound represents the basic form of the word and
Tsuching (described in sections F1c to F1f above)
used for. In other words, the unknown voice to be recognized.
When uttered, the sound processor 1 (shown in Figure 4)
004 generates a string of labels accordingly
do. The present invention provides an improved set of voice machines.
The basic form formed by the phoneme machine from
Enables matching with labels during stringing.
Ru. Using the phoneme machine added by the invention
This significantly improves the accuracy and speed of speech recognition.
Improved. Incidentally, the present invention is based on speech recognition of separated words.
recognition system and continuous speech speech recognition system
It can be used for. Separated word place
There is a short pause after each word. obey
The beginning and end of each word are often
In this case, there is energy storage and decay. present invention
is particularly well suited for such systems. Communicating
In continuous speech, multiple words are combined, usually
There are short pauses between phrases, allowing energy to accumulate and
Characterize each word basic form with a decaying part
Instead, the interphrase onset phoneme machine and decay phoneme machine
Indicates the supply of syn. Separated words and sequences
Speech phrases are categorized under the umbrella term “speech segments”
included. The audio segment is between two periods of silence.
It is considered to be the audio part of

【table】

【table】

【table】

【table】

【table】

【table】

【table】

【table】

【table】

【table】

【table】 >>
sz

【table】

【table】

[Table] SP〓'BX'〓b SP〓'ONAW'〓 SP〓'UH0'
〓〓
SP〓'DH'〓〓 SP〓'ONBX'〓 SP〓'UU0'
〓u
SP〓'DX'〓d SP〓'ONDH'〓 SP〓UX0'〓
U
SP〓'D$'〓 SP〓'ONEE'〓 SP〓'AA1'
〓α
SP〓'FX'〓f SP〓'ONER'〓 SP〓'AE1'
〓〓
I
SP〓'GX'〓g SP〓'ONFX'〓 SP〓'AI1'
〓a
<
U
SP〓'HX'〓h SP〓'ONIX'〓 SP〓'AU1'
〓a
<
SP〓'JX'〓j SP〓'ONLX'〓 SP〓'AW1'
〓〓
h
SP〓'KQ'〓 SP〓'ONMX'〓 SP〓'EE1'
〓i
<
SP〓'KX'〓k SP〓'ONSH'〓 SP〓'EH1'
〓ε
SP〓'LX'〓l SP〓'ONSX'〓 SP〓'EI1'
〓e
SP〓'MX'〓m SP〓'ONUH'〓 SP〓'ER1'
〓з
SP〓'NG'〓y SP〓'TRAA'〓 SP〓'IX1'
〓I
I
SP〓'NX'〓n SP〓'TRAW'〓 SP〓'OI1'
〓o
<
SP〓'NXV'〓 SP〓'TRBX'〓 SP〓'OU1'
〓o
SP〓'PQ'〓u SP〓'TRDH'〓 SP〓'UH1'
〓Λ
SP〓'PX'〓p SP〓'TREE'〓 SP〓'UU1'
〓u
SP〓'RX'〓r SP〓'TRER'〓 SP〓'UX1'
〓U
SP〓'R$'〓〓 SP〓'TRFX'〓 SP〓'AA2'
〓α
〉
SP〓'SH'〓 SP〓'TRKQ'〓 SP〓'AE2'
〓〓
s
I
SP〓'SX'〓s SP〓'TRLX'〓 SP〓'A12'
〓a
<
U
SP〓'TH'〓θ SP〓'TRMX'〓 SP〓'AU2'
〓a
<
n
SP〓'TQ'〓 SP〓'TRSH'〓 SP〓'AW2'
〓〓
<
SP〓'TX'〓t SP〓'TRSX'〓 SP〓'EE2'
〓i
SP〓'VX'〓v SP〓'AA0'〓α SP〓'EH2
'〓ε
SP〓'WX'〓w SP〓'AE0'〓〓 SP〓'EI2
'〓e
I
SP〓'W@'〓hw SP〓'AI0'〓a SP〓'ER2'
〓з
<
U
SP〓'XX'〓 SP〓'AU0'〓a SP〓'IX2'
〓I
<
〉I
SP〓'ZH'〓 SP〓'AW0'〓〓 SP〓'O12'〓
o
z〈
SP〓'ZX'〓z SP〓'EE0'〓i SP〓'OU2'
〓o
SP〓'? X'〓? SP〓'EH0'〓ε SP〓'UH
2'Λ
i
SP〓'EEG'〓 SP〓'EI0'〓e SP〓'UU2'
〓u
<
I
SP〓'IXG'〓 SP〓'ER0'〓〓 SP〓'UX2'
〓U
<
U
SP〓'UXG'〓 SP〓'IX0'〓I
<
I
SP〓'ONAA'〓 SP〓'OI0'〓o
<
G. Effects of the Invention According to the present invention, an improved word basic format can be constructed and the accuracy and speed of speech recognition can be improved.

[Brief explanation of drawings]

1A and 1B are flowcharts illustrating the method of constructing the basic form according to the present invention, and FIG.
Figures 2A and 2B show the arrangement of Figures A and 1B, and Figures 2A and 2B show the starting phoneme machine, common phoneme machine, and ending phoneme machine according to the present invention for use in constructing an improved basic form. FIG. 2C is a diagram showing the arrangement relationship between FIGS. 2A and 2B, FIG. 3 is a starting phoneme machine,
A block diagram showing an apparatus for constructing an improved word base form formed from a common phoneme machine and an end phoneme machine, FIG. 4 is a schematic block diagram of a system environment in which the present invention can be implemented, and FIG. FIG. 6 is a detailed block diagram of the stack decoder in the system environment of FIG. 8 shows the likelihood vector and likelihood envelope of each word path. FIG. 9 is a flowchart showing the steps of the stack decoding procedure. FIG. 10 shows the steps of the stack decoding procedure. 11 is a diagram showing the parts of a typical human ear representing the locations forming the components of the acoustic model. FIG. 12 is a block diagram showing the parts of the audio processor. Figure 13 is a diagram showing the relationship between sound intensity and frequency, which is used in the design of an audio processor. Figure 14 is a diagram showing the relationship between horns and horns. Figure 15 is a method for characterizing sound using the audio processor shown in Figure 10. FIG. 16 is a flowchart showing how to update the limit value in FIG.
17 is a diagram showing the trellis or lattice of the precision matching procedure, FIG. 18 is a diagram showing the phoneme machine used to perform the matching, FIG. 19 is a time distribution diagram used in the matching procedure with specific conditions, and FIG. Figures 20a-e are diagrams showing the interrelationships between phonemes, label strings, and start and end times determined by the matching procedure; Figures 21a and b are specific phoneme machines with a minimum length of 0; and its corresponding start time distribution, Figures 22a and b are illustrations of a specific phoneme machine of minimum length 4 and its corresponding trellis, and Figure 23 allows processing of multiple words at the same time. Figure 24 is a flowchart showing the steps taken in forming a formatted word base form; Figure 25 is a flowchart showing the steps taken in extending a word path; Figure 26 is a spectrum photograph of the word WILL spoken in isolation, No. 27
Figure 28 shows the waveform of the word WILL spoken in isolation, Figure 28 shows three statistical parts: initial, middle,
FIG. 29 is a diagram showing the finally divided phonetic symbol type phoneme machine, in which the word “WILL” is divided into 5 consecutive words according to the present invention.
Figure 30 shows the standard phonetic spelling of the word "WILL" as a continuous three-phonetic spelling; Figure 31 shows the phoneme sequence of the word "BOG" according to the present invention. FIG. 32 is a diagram showing the sequence of phonemes of the word "DOG" according to the present invention. 1000...Voice recognition system, 1002...
Stack decoder, 1004...Acoustic processor, 1006, 1008...Array processor, 1010...Language model, 1012...Workstation, 1020...Search device, 102
2,1024,1026,1028...interface.

Claims (1)

[Claims] 1. From a set of labels each representing an acoustic type that can be assigned to a minute time interval, a corresponding string of labels is generated according to an unknown audio input, and this string of labels is used as a vocabulary. In a speech recognition method that performs speech recognition by matching word Markov models of words in a set, the word Markov model extracts corresponding partial word Markov models from a set of partial word Markov models and concatenates them. The partial word Markov model is composed of a partial word Markov model of the start part corresponding to the start part of the word and a partial word Markov model of the other partial words.
The partial word Markov model of the above starting part is used only for matching the starting part of the word, and the partial word Markov model of the above starting part is used only for matching the starting part of the word. A speech recognition method characterized in that it is used. 2 For at least some of the partial word Markov models, a partial word Markov model of the ending part corresponding to the ending part of the word is further retained, and the partial word Markov model of the ending part is used only for matching the ending part of the word. A speech recognition method according to claim 1. 3. The speech recognition method according to claim 1 or 2, wherein the probability value of the partial word Markov model of the start part is determined based on the utterance of the start part of the word.