JPH1165590A - Voice recognition dialing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce cumbersome key operations for registering names and corresponding telephone numbers on a portable telephone by outputting the recognition result of the names and the telephone numbers by the recognition processing, which uses the monosyllable code data stream of the registered names or the telephone numbers with respect to input voices and standard pattern data in half syllable units. SOLUTION: A voice registration means, such as a general speaker half syllable voice recognizer 6 registers the uttered voices of names and telephone numbers through the voice input of the object names and the corresponding telephone numbers for a voice recognition dialing. The dialing voice means of the recognizer 6 conductors a dialing, using the monosyllablic code data column of the names and the telephone number which have been registered beforehand from the voice analog signals of the names and the telephone numbers. The recognizer 6 conducts the recognition process, which uses the monosyllablic code data column of the names and the telephone numbers that have been registered with respect to input voice and the standard pattern data in half syllable nits. Then, the name of the telephone number monosyllablic code data column of the candidate having a small cumulative distance value is outputted as the recognized result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a voice recognition dial device.

[0002]

2. Description of the Related Art Conventionally, the operation of registering a recognized word requires an operation on a key input device by a user. For example, to try to register one name and phone number for one person,
It is necessary for the user to perform key operation input of [number of characters of hiragana in person's name] + [number of digits of telephone number].

In order to reduce the size of a mobile phone, instead of an input device such as a personal computer having input keys for all characters,
Since the number of input keys is limited, a plurality of characters are duplicated and assigned.

[0004] For example, consider the case of registering a personal name = 7 characters and a telephone number = 10 digits, using an input device of a mobile phone as an example.

In a mobile phone, keys are assigned in the form of a row = 1 key, power row = 1 key to la row = 1 key, and wow = 1 key. When seven characters are registered, the shortest is 1 × 7.
= 7 key inputs, up to 5 x 7 = 35 key inputs,
On average, 2.5 × 7 = 17.5 key inputs.

[0006] All numbers are assigned to keys,
Since the input can be made by changing the operation mode, the key input is 1 + 10 times.

In the above example, if the average operation time is 2 seconds per key input, 2 seconds × (17.5 + 10) times = 5
5 seconds.

[0008]

In the above-mentioned prior art, the registration operation of the name and the telephone number for the voice recognition dial is long, and there are many erroneous operations.

[0009] Japanese Patent Laid-Open No. 2-1 has disclosed a prior art using voice.
Although there is a voice response recognition automatic dial telephone disclosed in Japanese Patent No. 35847, this conventional technique uses a function of recognizing a voice input signal and converting it into character data, but does not specify a realizing means.

An object of the present invention is to provide a voice recognition dial device which reduces the complexity of key operation such as name and telephone number registration operation and dial operation on a portable telephone or the like.

[0011]

A voice recognition dial device of the present invention acquires a monosyllabic code data string by voice recognition processing from a PCM signal of voice uttered a name and a telephone number, and obtains a target name for voice recognition dialing. Voice registration means for registering a telephone number and a telephone number by voice input, voice dial means for dialing from a voice analog signal of a name or a telephone number using a single syllable code data string of a name and a telephone number registered in advance, and input voice. By using the registered name or telephone number monosyllabic code data string and the semi-syllable unit standard pattern data, the recognition result of the candidate or the monosyllabic code data string of the candidate whose cumulative distance value is small And an unspecified speaker half-syllable speech recognizer.

The voice registration means receives an AD converter for converting a voice analog signal to a PCM signal, a DA converter for converting a PCM signal to a voice analog signal, and monosyllabic code data representing one Japanese syllable. Display means for displaying hiragana, numbers and kanji notation characters, voice response means for receiving single syllable code data, converting it into voice PCM data, and outputting it to a DA converter, Japanese names and all single syllables and numbers in Japanese Database RO in which monosyllable strings of phonetic notation are stored as monosyllable code data for each one digit
M and 1 name or 1 belonging to any one of a name item, a monosyllabic item, and a single digit item in the database ROM
Means for reading a single syllable code data string for one list of single syllables or single digit numbers and storing the same in a recognition work RAM, and a PCM signal obtained by AD-converting a voice analog signal uttered in single syllable units or single digit units Is input to an unspecified speaker half-syllable speech recognizer, performs recognition processing in units of single syllables or digits, and obtains a recognition result of the first candidate or a plurality of candidates. Using both the speech recognition result information and the single-syllable code data string of the name item or single-digit number item in the database ROM, the name or single-syllable code data of one digit or more that is estimated to be closest to the utterance content A voice registration searcher that searches and outputs a column, and a list of multiple names and telephone numbers by associating the searched single-syllable code data strings of names or telephone numbers with each name and telephone number It may comprise a recognition work RAM f accumulation storage for unit.

Further, the voice registration searcher receives the voice recognition result information in units of single syllables or single digits together with the accumulated distance value from the unspecified speaker half-syllable recognizer, and stores the information in the database ROM. From a voice signal uttered in units of one single syllable in one list of one name or one single syllable or one digit belonging to any of the following items: A means for retrieving the closest monosyllable code data string and storing it in the recognition work RAM may be provided.

The voice dial means inputs a PCM signal obtained by converting an uttered voice analog signal corresponding to the name or telephone number of the dial destination by an AD converter to an unspecified speaker half-syllable voice recognizer, and performs a recognition work. Means for obtaining a recognition result estimated to be closest to the voice analog signal uttered for the registered name or telephone number list in the RAM as a single syllable code data string is associated with the recognition result if the recognition result is a name. A means for retrieving a monosyllabic code data string of a telephone number from the recognition work RAM and outputting the same, and a means for converting the monosyllabic code data string of the telephone number to a telephone call signal to the telephone terminal body Good.

Further, the unspecified speaker half-syllable speech recognizer uses a half-syllable unit of a single-syllable code data string stored in a recognition work RAM or a database ROM. A means for converting to a data string, and a standard pattern obtained by generating a standard pattern data for an unspecified speaker half-syllable speech recognition device of a standard pattern generator for a half-syllable code data string of a name or a telephone number are stored. A means for checking which half-syllable unit hidden Markov model is included in the standard pattern ROM, further converting the data into a semi-syllable hidden Markov model state code data string indicating the connection of the hidden Markov model state, and a name or telephone number. The half-syllable hidden Markov model status code data string is associated with the name-telephone number relation, and the other name-telephone number Means for converting into a numbered list so that it can be distinguished from the list and storing it in a recognition work RAM; a speech analysis feature extractor for extracting an input feature vector in frame units from a speech PCM signal obtained by AD-converting a speech analog signal; Recognition work RA for extracted input feature vectors
A state distance calculator for calculating a state distance value between an input feature vector and a standard feature vector of all states of a full syllable hidden Markov model stored in a standard semisyllable hidden Markov model ROM. Means for numbering the calculated state distance value and storing the number in the recognition work RAM, and connecting the state of the half-syllable hidden Markov model state code data string for each name or telephone number of the recognition object stored in the recognition work RAM Cumulative state distance calculator that calculates the cumulative distance value for all frames of the input voice time length by the frame synchronization Viterbi algorithm using the information, each state distance value, and the state transition distance value stored in the standard pattern ROM. And a monosyllabic code for the name or telephone number of the first candidate or the first to Nth candidates with the smallest cumulative distance value It may have a means for outputting over data sequence as a recognition result.

Further, the standard pattern generator converts the analog voice signals of a large number of speakers, which are statistically required, into AD signals.
Means for dividing the converted PCM signal into PCM signals of a predetermined number of single syllable units by waveform display or sample listening, buffering all PCM signals divided by single syllable unit, and displaying waveform or sample listening Means for dividing into a predetermined number of types of semi-syllable unit PCM data signals by
Means for dividing all the PCM signals divided in units of syllables in units of frames, a speech analysis feature extractor for extracting a feature vector for all PCM signals divided in units of frames, Means for setting initial values of a predetermined number of state output probability functions and two times the predetermined number of state transition probabilities in the number of hidden Markov models, and multidimensional when initializing the state output probability functions Means for using the normal distribution probability density function, setting initial values of the mean vector and each component of the covariance matrix as parameters, and the number of statistically required persons obtained for each predetermined type of semisyllable From the frame-based feature vector of
Forward-Back is organized as feature vector samples for the number of people who are statistically required for each syllable.
means for obtaining a predetermined number of mean vector and each component value of the covariance matrix and a state transition probability of twice the predetermined number by an iterative method called a ward algorithm; Means for storing parameter values of a mean vector, a covariance matrix, and a state transition probability of a multidimensional normal distribution probability density function, which is a parameter group of the obtained hidden Markov model, as standard pattern data in a standard pattern ROM. Good.

Further, the voice analysis feature extractor includes means for emphasizing a high frequency band by performing pre-emphasis processing on a frame PCM signal obtained by dividing the PCM signal of the input voice into frames, and a pre-emphasis processed frame PC.
Means for performing window processing on the M signal to perform frame boundary smoothing processing for subsequent FFT processing; and performing N-order FFT conversion processing on the frame-processed PCM signal for N-th order on the linear frequency axis. Means for converting to a complex coefficient vector; means for converting an N-order complex coefficient vector to an N-order amplitude coefficient vector by calculating the absolute value of a complex number;
Means for calculating an order logarithmic amplitude coefficient vector, means for calculating a Pth order cepstrum vector by performing an inverse discrete cosine transform on the time axis for an Nth order logarithmic amplitude coefficient vector, and suppressing a high time component of the Pth order cepstrum vector. Means for calculating a Pth vocal tract characteristic cepstrum vector by separating the vocal tract characteristic and the pitch characteristic and extracting only the vocal tract characteristic, and performing a discrete cosine transform on the frequency axis for the Pth vocal tract characteristic cepstrum vector. And N on the linear frequency axis
Means for converting to a second vocal tract characteristic logarithmic amplitude coefficient vector;
Nth vocal tract adapted to human auditory frequency resolution characteristics by calculating vector components obtained by performing interpolation or smoothing processing on spectral components on equal division points on the mel frequency axis with respect to the logarithmic amplitude coefficient vector of the secondary vocal tract characteristics Means for transforming into a characteristic log magnitude mel scale coefficient vector;
Means for transforming the secondary vocal tract characteristic logarithmic amplitude mel scale coefficient vector into a Qth-order mel-cepstral vector by inverse discrete cosine transformation on the time axis.

Further, the state distance calculator converts the analog signal of the input voice into an analog signal, converts the analog signal of the input voice into frames, and further divides the frame into PCM signals obtained by the voice analysis feature extractor and all the states of the standard semisyllable hidden Markov model. Means for calculating a state distance value with the standard feature vector by a predetermined distance calculation formula.

Further, the cumulative distance calculator calculates the state combination information of the half-syllable hidden Markov model state code data string for the name or telephone number to be recognized, each state distance value, and the state selection stored in the standard pattern ROM. There may be provided means for calculating a cumulative distance value for all frames of the input voice by a frame synchronization Viterbi algorithm using the transfer distance value.

Therefore, according to the present invention, the operation time for registering the name and telephone number for the voice recognition dial is reduced as compared with the conventional example, and the number of operation errors is reduced.

Also, in the recognition automatic dial telephone of the prior art, the voice recognition processing for displaying characters from unspecified voice is actually used in the technical field to which the present invention belongs by specifying means for realizing this technology. It becomes possible.

[0022]

Embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the overall configuration of the embodiment of the present invention. First, 2 as a storage device
ROM (database ROM1, standard pattern R)
OM2) and two types of RAMs (a recognition work RAM3 and a search work RAM4).

The functional blocks include a standard pattern generator 5, an unspecified speaker half-syllable speech recognizer 6, a speech registration searcher 7, an AD converter 8, and a DA converter 9
, A voice synthesizer 10, a mobile phone main unit 11, a key matrix 12, a display device 13, and an I / F unit 14. Hereinafter, the contents of the implementation will be described separately for the voice registration operation and the voice dial operation. Subsequently, the internal detailed operation of the main functional blocks is described by a standard pattern generator 5, an unspecified speaker half-syllable speech recognizer 6, and a speech registration searcher 7
Are described together with the contents of the link operation with each storage device.

The embodiment for the functional blocks shown in FIG. 1 can be basically realized by a semiconductor integrated circuit and a composite device. First, ROM1, 2 and RAM
3, 4, the AD converter 8, the DA converter 9, and the like constitute a semiconductor integrated circuit. Also, the mobile phone body 11
, A key matrix 12 and a display device 13 (this is an LC
D, etc.) and the I / F unit 14 (which can use an expansion connector for connecting the mobile phone and the expansion device) constitute a composite device. Furthermore, the voice registration searcher 7, the unspecified speaker half-syllable voice recognizer 6, and the voice synthesizer 10 are provided on a semiconductor integrated circuit that can incorporate a microprogram called a CPU or a DSP capable of performing high-speed operation of voice signal processing. It can be realized by software in the above.

1. Speech registration operation First, single syllable code data (ASCII or J) in all single syllable units (one hiragana character, that is, 50 tones + α (dull sounds, murmur sounds, etc.)) stored in the database ROM1.
IS or SJIS, etc., and further divides each single syllable code data into half syllable code data (a half syllable unit is a phoneme unit obtained by further dividing one syllable into half. For example, in the case of “ta”, TA → T−, -Divided into -A) and stored in the recognition work RAM 3. The conversion from a single syllable to a half syllable is described in 3. This function is built into the unspecified speaker half-syllable speech recognizer.

Next, by discretely uttering the name and telephone number to be registered in discrete monosyllable units or single digit units,
The monosyllabic code data as the hiragana character information of the registered name and the registered telephone number and the numerical code data as the numerical information can be obtained by the following (1) name registration operation and (2) telephone number registration operation. . Further, the obtained association information between the name and the telephone number and the number in the order of registration are given and stored in the recognition work RAM 3. Thereby, the consecutive voice registration operation ends. In the following (1) and (2), the details of the voice registration operation will be described using an example in which the name is “Takahashi” and the telephone number is “03-123-4567”.

(1) Name Registration Operation FIG. 2 is a flowchart of the voice registration operation mode.

First, the speaker is "ta": "ka": "ha":
A voice analog signal uttered with a silence interval of "S" and a single syllable is converted into a PCM signal by the AD converter 8 (S3). An unspecified speaker half-syllable speech recognizer 6 obtains a single-syllable unit recognition result for each of the four single-syllable PCM signals (S4). Here, it is assumed that the recognition results are output from the closest one to the first to fifth candidates. The reason that a plurality of single syllable recognition result candidates are required is that the recognition accuracy of a minimum phoneme unit is usually low in a speech recognizer of a phoneme unit such as a half syllable. Is the same vowel syllable,
Since a recognition result such as a vowel system such as a murmur or a muddy tone is likely to be output, almost no accurate recognition result can be obtained by using only the first candidate. Therefore, if a plurality of candidates are used as the recognition result, there is a high possibility that the recognition result of the true vocal syllable is included in the plurality of candidates, and further, a voice registration search unit 7 and a database for determining a registration name described below. The recognition accuracy is increased from the information in the ROM 1.

By the way, the names "ta""ka""ha""shi"
For each PCM signal obtained by AD-converting a four-syllable discrete utterance analog signal, all single syllables were regarded as recognition candidates (actually, the recognition work R
All the monosyllable code data is transferred to AM3). As shown in FIG.
A recognition result for syllables × 5 candidates is obtained.

These information are described in 4. It is stored in the registration work RAM as described in the voice registration search device.

Next, the voice registration search unit 7 determines the registered name by a search process from the recognition result and the name list limited to four syllables in the database ROM 1 (since the number of utterances is four) (see FIG. 4). S9).

If a plurality of name search results are to be obtained, a user (= speaker) can be finally selected by key input or the like using a display function or a voice synthesis function (S10). .

(2) Telephone Number Registration Operation FIG. 4 is a flowchart of a telephone number registration operation mode.

The telephone number is registered in the recognition work RAM 3 in a single syllable combination list of numbers "" ++ (0) "to" @ + @ +
(9) ″) is loaded in advance, and the discrete utterance is performed one digit at a time for each digit of the telephone number (S
13) The obtained recognition result is converted into a code directly representing a number as shown in FIG. 5 (S14). Note that in the case of digit recognition, the recognition target list is about 10 compared to single syllable recognition, and in the case of recognition of two or more syllables, the recognition accuracy is improved.
There is no need to prepare for 1. When a plurality of numerical results are selected, the user (= speaker) can be finally selected by key input or the like using a display function or a voice synthesis function (S16).

When registering a telephone number, a key input operation on a mobile phone or the like may be made selectable. This is because, in general, numeric key assignment is often performed one digit at a time, and thus the input of a name is less complicated.

2. Voice Dial Operation FIG. 6 is a flowchart of the voice dial operation mode.

Before starting dialing, an initial setting as to which of the name or telephone number stored in the recognition work RAM 3 is to be recognized is made by designating the voice input of the name or telephone number by key input or the like. (S2
2, S23, S30).

Next, a name or a telephone number is uttered (S24, S31), and the PCM signal obtained by AD-converting the analog voice signal is subjected to recognition processing by an unspecified speaker half-syllable voice recognizer (S25, S25). S32).

In the initial setting, it is known whether the recognition result is a name or a telephone number. Therefore, when the telephone number recognition result is obtained, the dialing can be executed by converting the numeric code or the like of the telephone number into a dial call signal based on the information (S36, S37). When the name recognition result is obtained, the number code of the telephone number is specified from the information in which the name and the telephone number are associated with each other, and the dial can be similarly performed (S29, S36, S37).
In addition, after the telephone number or name recognition process, before executing the dial, the name character and the number of the telephone number are displayed, or the sound is output by a voice synthesizer to perform a single confirmation procedure. Yes (S27, S3
4) By doing so, it is possible to make the voice dial function more compatible.

3. Unspecified speaker half-syllable speech recognition operation (1) Standard pattern generation processing FIGS. 7 and 8 are functional block configuration diagrams of the standard pattern generator.

FIG. 9 is a flowchart of the standard pattern generation processing by the standard pattern generator.

The generation of the standard pattern is performed by using a large number of speakers (Nmax
Hidden Markov model (HMM) for each syllable unit from the uttered voice sample of
Model) probability parameter.

First, a balanced phoneme text including all articulatory connection patterns is prepared, and a number of statistically sufficient speakers = Nmax are uttered (S38), and AD converted (S38).
39), and temporarily store the PCM signal in the recognition work RAM 3. Next, the PCM data is waveform-displayed or D / A-converted, and divided into Hmax-type single syllable units (S40) by work such as visual observation or sample listening, and further, Imax-type semisyllable units classified in consideration of articulation before and after. A semisyllable PCM signal is obtained every time (S41). The state of this processing is shown in the example of FIG.

Here, "-" of "K-" indicates that a sound follows, and "-" of "1A" indicates that the sound follows the preceding sound.

Up to the above, the number of semisyllable types × Nmax
The PCM signal samples for the person are ready. After further dividing (framing) these PCM signal samples into frames of about 12 ms to 16 ms per frame (S42), a feature vector (frame unit) for half syllable types × Nmax is obtained by speech analysis feature extraction processing described later.
Is obtained (S43).

Finally, generating a standard pattern for "K-" means that one HMM standard pattern model is associated with one semisyllable category "K-", and the Fmax of the half syllable is obtained. An object is to obtain a probability parameter and a state transition probability of each state so that a feature vector output for a frame is output with the highest probability in the transition process of the corresponding four states of the HMM.

Next, for each semisyllable type, the number of states = Jmax The parameter of the state output probability function Bj of the Left to Right type HMM and the state transition probability of Jmax × 2 state = α
A method for obtaining each parameter of (j-1, j) and α (j, j) will be described. For example, one semisyllable "
For K- ", a semisyllable H having four states as shown in FIG.
The case where each parameter of MM is obtained from Nmax feature vectors will be described.

Here, α _{00 to} α ₃₃ indicate the state transition probabilities α [*, *] as follows. Α (j−1, j): state transition probabilities from state j−1 to j (α ₀₁ , α ₀₂ , α ₀₃ ). Α (j, j): state transition probability from state j to j (α ₀₀ , α ₁₁ , α ₂₂ , α ₃₃ ).

The output probability = B ₀ -B ₃ is calculated as follows. Bj: a feature vector output probability function of state j in the following equation (using a multidimensional normal distribution (Gaussian) probability density function at the time of initial setting, and initial values of each component of a mean vector and a covariance matrix as parameters) Yes.)

[0050]

(Equation 1) (Note: t is a transpose operation (vertical vector → horizontal vector), V
kmj-1 indicates an inverse matrix. Where j = state number, j = 0 to Jmax-1 k: mixture distribution number. k = 0 to Kmax-1, Kmax: number of mixture distributions m: feature vector type number m = 0 to Mmax-1, Mmax:
Number of feature vector types λkmj: Mixture distribution coefficient that determines the weight of mixture distribution μmj: Feature vector weighting factor that determines the weight between feature vector types Xm: Feature vector of input speech sample in frame units

[0051]

(Equation 2) Vkmj: covariance matrix | Vkmj |: norm of Vm (determinant) In the following, the number of mixture distributions is set to 1 for ease of explanation.
(Kmax = 1, λkmj = 1) and the feature vector type is 1 (Mmax = 1, μmj = 1), and the following expression is used. [Output probability density function when the number of mixture distributions = the number of feature vectors = 1]

[0052]

(Equation 3) The parameter to be calculated here is

[0053]

(Equation 4) Covariance matrix: Vj and state transition probability: α (j-1, j)
And α (j, j), which are the standard patterns of the semisyllable “K−”.

These parameters are obtained from the feature vector samples of Nmax syllables “K−” as described below.
B (Forward Backward) algorithm (also called Baum-welch algorithm, EM
(Expecton Maximization)
(Algorithm based on the method) until the convergence is repeated.

Before describing the FB algorithm, first, the feature vectors of Nmax persons of the syllable "K-" are redefined as follows.

Redefinition: Feature vector of speaker n: X → X (n, f) where n: speaker number, n = 0 to Nmax−1 f: frame number, f = 0 to Fmax (n) − 1 Fmax (n): the number of frames of the half syllable “K−” of the speaker number n (Note: generally, the number of frames sampled for each speaker is different) Further, the following FB algorithm processing is performed (S4).
6).

[FB algorithm] Covariance matrix: Vj,

[0058]

(Equation 5) State transition probabilities: Initial values of α [j−1, j] and α [j, j] are set (S45).

[Initial setting value]

[0060]

(Equation 6) Covariance matrix: Vj → Vj 0 State transition probability: α [j-1, j] → α [j-1, j]
0 and α [j, j] → α [j, j] 0

The target value (= FWD) of the probability value by the forward pass algorithm for the HMM of the semisyllable "K-" t
h) and the maximum number of iterations (= CNTmax) are set.

The processing of (cnt = 1 to CNTma)
Repeat until x.

The above processing is performed for j = 0 to Jmax (Jmax
= 3).

The updated value of each parameter is calculated by the following equation.

[0065]

(Equation 7)

Using the new parameters, the output probability of the HMM model by the forward pass algorithm for the input feature vector = Xm (n, f) is obtained by the following equation.

[0067]

(Equation 8)

Output probability ≧ FWD th holds (S
47) Or, if cnt> CNTmax (S4
9) End the process.

At this time, the parameter is set to a semisyllable “K−”.
Standard pattern.

Here, FWD (j, f) is the probability (Baum-Welch) obtained by the forward path algorithm.
BCK (j, f)
Is the probability determined by the backward path algorithm.

In practice, when obtaining the parameters of the standard pattern from a large number of samples, the state output probability function may be a mixture distribution of two or more multidimensional normal distributions (for example, male and female), Recognition accuracy can be further improved by increasing the number of types (for example, in addition to the mel cepstrum vector, Δmel cepstrum of inter-frame difference, inter-frame difference of average power of one frame: Δ average power, etc.).

In the above processing, Imax kinds of semisyllable HMMs are obtained as standard patterns, and standard patterns RO
It is stored in M (S50).

(2) Speech Analysis Feature Extraction Processing FIGS. 12 and 13 are functional block diagrams of an unspecified speaker half-syllable speech recognizer.

FIGS. 14 to 16 are flowcharts of the unspecified speaker half-syllable speech recognition processing. FIG. 17 is a chart showing a legend of the flowchart.

The voice analysis feature extraction process of the voice analysis feature extractor is a process in which all the following steps are performed in frame units.

The PCM signal of the input voice is changed from 12 ms to 16
The high frequency band is emphasized by performing pre-emphasis processing (first-order difference processing) on the frame PCM signal obtained by dividing the frame into about ms.

Pre-emphasis-processed frame PC
By performing window processing (such as a Hanning window) on the M signal, smoothing processing of frame boundaries for the subsequent FFT processing is performed.

The frame PCM signal after the window processing is converted to an Nth order F
By performing the FT conversion processing, it is converted into an Nth-order complex coefficient vector on the linear frequency axis.

The N-order complex coefficient vector is converted into an N-order amplitude coefficient vector by calculating the absolute value of a complex number.

A logarithmic operation is performed on the Nth order amplitude coefficient vector to calculate an Nth order logarithmic amplitude coefficient vector.

A P-order cepstrum vector is calculated by performing an inverse discrete cosine transform on the time axis for the N-th logarithmic amplitude coefficient vector.

The P-order cepstrum vector is calculated by extracting the vocal tract characteristics only by extracting the vocal tract characteristics by separating the vocal tract characteristics and the pitch characteristics (vocal fold characteristics) by a process (lifting) that suppresses the high-time component of the P-order cepstrum vector.

The P-order vocal tract characteristic cepstrum vector is subjected to a discrete cosine transform on the frequency axis, and is converted into an N-order vocal tract characteristic logarithmic amplitude coefficient vector on a linear frequency axis.

By calculating vector components obtained by performing interpolation or smoothing processing on spectral components on equal division points on the mel frequency axis (approximately on a logarithmic scale) with respect to the logarithmic amplitude coefficient vector of the Nth vocal tract characteristic, An Nth vocal tract characteristic logarithmic amplitude mel scale coefficient vector is converted to an auditory frequency resolution characteristic (low frequency: high → high frequency: low).

Next, the Nth vocal tract characteristic logarithmic amplitude mel scale coefficient vector is transformed into a Qth mel-cepstral vector by inverse discrete cosine transformation on the time axis.

As described above, the input feature vector is obtained from the PCM signal for one frame (S58).

(3) State distance calculation The state distance calculator converts the analog signal of the input voice from analog to digital, and further obtains the PCM signal divided into frames of about 12 ms to 16 ms by the voice analysis feature extraction processing of the voice analysis feature extractor. Input feature vector and standard semisyllable H
The state distance value of all states of the MM with respect to the standard feature vector is calculated by the following distance calculation formula (S60).

[Distance calculation formula] ΔXi, j, k, m = Xin−Xi, j, k, m

[0089]

(Equation 9) (Note: t is a transpose operation (vertical vector → horizontal vector), V
−1 of i, j, k, m indicates an inverse matrix. Xin: feature vector of input voice Xi, j, k, m: feature average vector of standard pattern Di, j: state distance calculation value of state of semisyllable = i, state = j Vi, j, k, m: standard Pattern feature covariance matrix | Vi, j, k, m |: Norm (dispersion value) of covariance matrix Vi, j, k, m i: Semisyllable number, i = 0 to Imax-1, Imax: All Semisyllable type j: State number in HMM of one syllable, j = 0 to Jma
x-1, Jmax: total number of states of 1 HMM k: mixture distribution number, k = 0 to Kmax-1, Kmax: mixture distribution number m: feature vector type number, m = 0 to Mmax-1, Mma
x: Total number of feature vector types

(4) Viterbi Process (Pattern Matching Process) As an example, consider the recognition of the following name list. The word “takahashi” is first written as “TAKAH
ASI "and converted to vowel and consonant strings, and" T-, -A
-, -K-, -A-, -H-, -A-, -S-, -I "
Is regularly decomposed into a series of syllables. (Single syllable string) "Takahashi" (vowel, consonant string) "TAKAHASI" (half syllable string) "T-, one A-, -K-, -A-, -H-, -A-,-
S−, −I ″ Each semisyllable is a standard pattern HMM as shown in FIG.
Was expressed with.

As a result, the H of the word "Takahashi"
The MM connection model is as shown in FIG.

The Viterbi algorithm calculates the -type cumulative probabilities from the HMM connection model.

The Viterbi algorithm is basically a repetition of the optimum path selection processing of FIG.

As an easy-to-understand example, FIG. 20 shows an example of calculating a Viterbi score of one syllable HMM "T-". The input pattern is for Fmax frames. As in the example of FIG. 20, the Viterbi scores of all states are obtained for each frame input, and the calculated Viterbi scores of state 3 for all frames are the standard values for the input feature vector of “T−”. This is the output probability of the pattern. Now, when this is applied to the HMM connection model of the word “Takahashi”, the number of states is 4 × 8 = 32, the number of input frames = Fmax, and the Viterbi of the state 3 in the HMM of the semisyllable “−I” is set.
By calculating the score, the output probability of the input feature vector “Takahashi” from the HMM connection model is obtained.

Actually, since there are a plurality of recognition target lists other than "Takahashi", for example, the name "Ito" is converted into the same syllable string as Vit
An erbi score is calculated (S75). Then, one recognition target list having the highest probability value (smallest distance value) from the Viterbi score in the entire recognition target list is set as a recognition result (S86). The processing procedure of the Viterbi algorithm will be described below.

[Viterbi algorithm] (S57 to S81) to f = 0 to Fmax-
Repeat until 1. (S73 to S79) to w = 0 to Wmax-
Repeat until 1. (S73 to S77) to s = 0 to Smax
Repeat until (w) -1. (S73) i ← [semisyllable number to which state = s belongs] j ← [H at semisyllable number = i to which state = s belongs]
State number in MM] (S74) Path (j-1, j) = α (j-1,
j) + G (w, s-1) Path (j, j) = α (j, j) + G (w, s) (S75) Cumulative distance value: G (w, s) = Max [Pa
th (j-1, j), Path (j, j)] + Dij where f: input frame number, Fmax: total number of frames w: number of recognition target (name or telephone number) list, Wma
x: total number of lists s: serial number of state connection in the list, Smax (w): recognition target list = w total number of connected states α (j−1, j): state j in the i-th semisyllable HMM
State transition distance value from −1 to j (distance value = 0 when j−1 <0) α (j, j): State transition distance value from state j to j in the i-th semisyllable HMM Dij: State distance value from input feature vector of state of semisyllable = i, state = j

The Viterbi score has been described as a probability value. However, in order to avoid a problem such as an underflow due to the limitation of the number of digits, the Viterbi score is obtained by performing a logarithmic operation on the Virbi score.
Sometimes iterbi scores are calculated.

There is also a device for reducing the calculation amount by clustering the average vector and the covariance matrix of the standard pattern in order to reduce the calculation amount and the memory amount.

For example, if there are 250 types of half-syllable HMMs, 250 × 4 = 1000 types of average vectors and covariance matrices must be prepared. For example, the average vectors are classified into 512 categories (in this case, the covariance matrix is also 512). Type) and the norm of the variance matrix into 256 categories,
When clustering is performed by representing a vector value by a vector quantization method (such as a centroid vector), the amount of calculation and the amount of memory are reduced to ２〜 to ４. Actually, it is possible to reduce the calculation amount and the memory amount without deteriorating the recognition performance by such a device.

4. Search unit for voice registration The search unit for voice registration (hereinafter referred to as a search unit) is a unit for recognizing recognition results in units of one syllable from an unspecified speaker half-syllable recognizer (hereinafter referred to as a recognizer) together with distance values from the first candidate to the N candidate. When the information is received, the character code of the closest name is searched from the hiragana character data list of all Japanese names stored in the database ROM 1 from the voice signal of the name uttered in syllable units. Is stored in the recognition syllable string work RAM. As a result, the name is registered by inputting the audio signal. Hereinafter, a specific example of the voice registration operation will be described.

First, the voice signal to be uttered is "ta" +
The case where “ka” + “has” + “shi” and discretely uttered in units of one syllable is taken as an example. A plurality of candidate character codes and distance values of the recognition result are output from the recognizer to the searcher for each syllable utterance, as shown in the example of FIG.

Here, the first candidate is the actual uttered voice =
The reason why “na” is set to “ta” is that the utterance method for one syllable is similar. This is basically the same as that in which only consonants differ in the structure of one syllable of consonant + vowel, and when all vowels are the same, even if a person's auditory discrimination ability is erroneously heard. The recognizer extracts a feature amount of a voice signal called a mel-cepstrum, which has the same meaning as extracting a feature amount of a vocal tract.

In the case where the PCM signal of “ta” is actually recognized by the recognizer, the probability of occurrence of the semi-syllable HMM combination of “T −” + “− A” is the highest (the distance value is small). ), Actually, the "consonant half-syllable" + "-A" also has substantially the same recognition distance due to the influence of human subtle vocal deformation. For example, if we consider the same vowel “na”, the semisyllable is “N” + “−
A ", the vowel part is the same, and the consonant part is" ta "and" na ", because at the moment when both consonants are uttered, the tongue is put on the upper jaw and then" -A "is uttered. However, in the case of a name that is usually composed of multiple syllables, the recognition results of the remaining syllables of “ka”, “wa”, and “shi” are also different. Although it is easy to make a misrecognition, it is possible to compensate for the misrecognition of the monosyllabic recognition by calculating the cumulative probability only for the syllable string of the existing name, which will be described below.

First, "ta", "ka", "ha", "shi"
It is assumed that the monosyllable recognition result for each utterance is as shown in FIG.

First, the search device temporarily stores this information in the syllable string work RAM for registration. Next, since it is possible to count the number of utterances to four by using a counter or the like, a four-character name list is searched from the database ROM and stored in the work RAM. Next, the character codes in the four-character name list are checked to narrow down the names including at least one character code in the recognition result. Further, for the narrowed-down four-character name list, the cumulative distance value is calculated by the cumulative addition operation process using the distance value of the recognition result. Here, some character codes with a certain name have no distance, but are calculated by the following method.

Here, it is assumed that the narrowed down four character name list is as shown in FIG. 23 as an example.

Next, a cumulative distance value is calculated for these candidate names. At this time, distance values are added to candidates in the recognition result list in consideration of the order of utterance, and for characters not in the recognition result list, the maximum distance value is set to 5.0 and cumulative addition is performed. Perform the operation. From the list of examples in FIG.
“Ta” + “ka” + “ha” + “shi” is the smallest value, and this is the registered name.

[0108]

As described above, according to the present invention, a recognition process using a single syllable code data string of registered names or telephone numbers and standard pattern data in semi-syllable units for an input voice is used in a portable telephone or the like. It is possible to reduce the complexity of key operation such as name and phone number registration operation and dial operation, shorten the registration operation time of name and phone number for voice recognition dial compared with the conventional example, There is an effect that it decreases.

[0109] In the recognition automatic dial telephone of the prior art, the voice recognition processing for displaying characters from unspecified voice is actually used in the technical field to which the present invention belongs by specifying means for realizing this technology. There is an effect that it becomes possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an overall configuration of an embodiment of the present invention.

FIG. 2 is a flowchart of a voice registration operation mode.

FIG. 3 shows 4 obtained by an unspecified speaker half-syllable speech recognizer.
It is a figure which shows the recognition result for syllable * 5 candidate.

FIG. 4 is a flowchart of a telephone number registration operation mode.

FIG. 5 is a diagram showing conversion of a telephone number recognition result into a code that directly represents a numeral.

FIG. 6 is a flowchart of a voice dial operation mode.

FIG. 7 is a functional block configuration diagram of a standard pattern generator.

FIG. 8 is a functional block configuration diagram of a standard pattern generator.

FIG. 9 is a flowchart of a standard pattern generation process.

FIG. 10 is a diagram illustrating an example of obtaining a semi-syllable PCM signal from a single-syllable PCM signal.

FIG. 11 is a diagram showing a semisyllable HMM having four states.

FIG. 12 is a functional block diagram of an unspecified speaker half-syllable speech recognizer.

FIG. 13 is a functional block configuration diagram of an unspecified speaker half-syllable speech recognizer.

FIG. 14 is a flowchart of an unspecified speaker half-syllable speech recognition process.

FIG. 15 is a flowchart of an unspecified speaker half-syllable speech recognition process.

FIG. 16 is a flowchart of an unspecified speaker half-syllable speech recognition process.

FIG. 17 is a chart showing a legend of a flowchart.

FIG. 18 is a diagram illustrating an HMM connection model of the word “Takahashi”.

FIG. 19 is a diagram illustrating an optimal path selection process.

FIG. 20: Viterb of one semisyllable HMM “T-”
It is a figure showing the example of i score calculation.

FIG. 21 is a diagram showing a character code and a distance value of a recognition result for each utterance of one syllable, which is output as a plurality of candidates from an unspecified speaker half-syllable speech recognizer to a speech registration searcher.

FIG. 22 is a diagram showing the results of single syllable recognition for each utterance of “ta”, “ka”, “ha”, and “shi”.

FIG. 23 is a diagram for calculating an accumulated distance value for the narrowed-down four-character name list.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Database ROM 2 Standard pattern ROM 3 Recognition work RAM 4 Search work RAM 5 Standard pattern generator 6 Unspecified speaker half-syllable speech recognizer 7 Voice registration searcher 8 AD converter 9 DA converter 10 Voice synthesizer 11 Mobile Telephone main unit 12 Key matrix 13 Display device 14 I / F unit 15 Voice registration voice dial operator 16 Control signal bus 17 Memory bus

Claims (9)

[Claims] 1. PCM of a voice uttering a name and a telephone number
Voice registration means for acquiring a monosyllable code data string from the signal by voice recognition processing and registering the target name and telephone number for voice recognition dialing by voice input, and a name registered in advance from a voice analog signal of the name or telephone number Voice dial means for dialing using a single syllable code data sequence of a telephone number and a telephone number, and a recognition process for input voice using a single syllable code data sequence of a registered name or telephone number and standard pattern data of a half syllable unit And an unspecified speaker half-syllable speech recognizer that outputs a single syllable code data string of a candidate name or telephone number having a small cumulative distance value as a recognition result. 2. The audio registration means includes: an AD converter for converting an audio analog signal to a PCM signal; and a DA for converting the PCM signal to the audio analog signal.
A converter; display means for receiving single syllable code data representing one Japanese syllable and displaying notation characters of hiragana and numbers and kanji; receiving the single syllable code data and converting it to voice PCM data and outputting it to a DA converter A voice response means, a Japanese ROM, a database ROM storing monosyllable strings of phonetic notation for each single syllable and one digit of Japanese as single syllable code data, and a name item or a single name in the database ROM. Means for reading out a single syllable code data string for one list of one name or one single syllable or one digit belonging to any one of a syllable item and a one-digit number item, and storing it in a recognition work RAM; Or inputting the PCM signal obtained by AD converting the voice analog signal uttered in units of one digit to the unspecified speaker half-syllable voice recognizer, Means for performing recognition processing in units of syllables or in units of one digit to obtain recognition results of the first candidate or a plurality of candidates; voice recognition result information in units of single syllables or in units of one digit and a name item or 1 in the database ROM A voice registration searcher that searches for and outputs a single syllable code data string of one or more digits that is assumed to be closest to the utterance content using both information of the single syllable code data string of the digit number item and And means for storing a plurality of each name and telephone number list in a recognition work RAM by associating the retrieved monosyllabic code data string of the name or telephone number with each name and telephone number. A voice recognition dial device as described. 3. The speech registration searcher receives the speech recognition result information in units of single syllables or single digits together with an accumulated distance value from the unspecified speaker half-syllable recognizer, and receives the information from the database. It is uttered in units of one single syllable in one list of one name or one single syllable or one digit belonging to either the name item or one single syllable or one digit number item stored in the ROM. Recognition work RA by searching the closest syllable code data string from the speech signal
3. The voice recognition dial device according to claim 2, further comprising means for storing the data in M. 4. The voice dial means inputs a PCM signal obtained by converting an uttered voice analog signal corresponding to a name or a telephone number of a dial destination by an AD converter to the unspecified speaker half-syllable voice recognizer, Means for acquiring, as a monosyllabic code data sequence, a recognition result estimated to be closest to the uttered voice analog signal with respect to the registered name or telephone number list in the recognition work RAM, and corresponding when the recognition result is a name. Means for retrieving and outputting a single syllable code data string of the attached telephone number from the recognition work RAM, and means for converting the single syllable code data string of the telephone number into a telephone call signal to the telephone terminal body. The voice recognition dial device according to claim 1, further comprising: 5. The unspecified speaker half-syllable speech recognizer according to a single-syllable unit data sequence of a single syllable unit of a name or a telephone number stored in a recognition work RAM or a database ROM. Means for converting into a code data sequence, and a standard pattern obtained by generating a standard pattern data for an unspecified speaker half-syllable speech recognition device of a standard pattern generator with respect to the half-syllable code data sequence of a name or a telephone number. Means for examining which half-syllable unit hidden Markov model is included from the stored standard pattern ROM, and further converting to a semi-syllable hidden Markov model state code data string indicating the connection of the hidden Markov model states; Or, the half-syllable hidden Markov model status code data string of the phone number is associated with the name-phone number relationship,
Means for converting into a numbered list so that it can be distinguished from another name-telephone number list and storing it in the recognition work RAM; and extracting an input feature vector in frame units from a voice PCM signal obtained by AD-converting a voice analog signal. A speech analysis feature extractor that extracts the input feature vector
A state distance calculation for calculating a state distance value between the input feature vector and a standard feature vector of all states of the all-syllable hidden Markov model stored in the standard half-syllable hidden Markov model pattern ROM. And a recognition work R by numbering the calculated state distance values
Means for storing in the AM, state coupling information of the half-syllable hidden Markov model state code data string for each name or telephone number of the recognition object stored in the recognition work RAM, each state distance value, and in the standard pattern ROM. A cumulative state distance calculator that calculates a cumulative distance value for all frames for the input voice time length by a frame synchronization Viterbi algorithm using the stored inter-state transition distance value, and a first candidate having the smallest cumulative distance value 5. The voice recognition dial device according to claim 1, further comprising: a unit that outputs a single syllable code data string corresponding to the first to N-th candidate names or telephone numbers as a recognition result. 6. 6. The standard pattern generator according to claim 1, wherein a predetermined number of single syllables of a PCM signal obtained by AD-converting voice analog signals of a number of speakers required for a statistical number are displayed by waveform display or sample listening. Means for dividing the PCM signal into unit PCM signals, means for buffering all the PCM signals divided into single syllable units, and dividing into a predetermined number of types of PCM data signals in semi-syllable units by waveform display or auditioning Means for dividing all the PCM signals divided in units of syllables in units of frames, a voice analysis feature extractor for extracting a feature vector for all PCM signals divided in units of frames, Initialize the predetermined number of state output probability functions and the number of state transition probabilities of twice the predetermined number in the hidden Markov model of a predetermined number. Means for setting, using a multidimensional normal distribution probability density function when initializing the state output probability function, and means for initializing each component of a mean vector and a covariance matrix as a parameter, a predetermined type From the feature vector in frame units for the statistically required number of people obtained for each half syllable of
For each syllable, the feature vector samples for the statistically required number of persons are arranged as Forward-Ba
means for obtaining each component value of the predetermined number of mean vectors and covariance matrices and the state transition probabilities of twice the predetermined number by an iterative technique called a ckward algorithm; Means for storing parameter values of a mean vector, a covariance matrix, and a state transition probability of a multidimensional normal distribution probability density function, which is a group of parameters of a hidden Markov model obtained in the above, as standard pattern data in a standard pattern ROM. Item 6. The voice recognition dial device according to Item 5. 7. The speech analysis feature extractor includes a frame PC obtained by dividing an input speech PCM signal into frames.
Means for emphasizing a high frequency band by pre-emphasis processing the M signal; means for performing window processing on the pre-emphasis-processed frame PCM signal to perform smoothing processing of frame boundaries for subsequent FFT processing; Means for converting the frame-processed PCM signal after the window processing into an N-order complex coefficient vector on the linear frequency axis by performing an N-order FFT transform processing; and an N-order amplitude coefficient vector by calculating an absolute value of a complex number from the N-order complex coefficient vector Means for performing a logarithmic operation on the Nth-order amplitude coefficient vector to calculate an Nth-order logarithmic amplitude coefficient vector; and performing inverse discrete cosine transformation on the time axis for the Nth-order logarithmic amplitude coefficient vector. Means for calculating a P-order cepstrum vector, and suppressing a high-time component of the P-order cepstrum vector Means for separating a vocal tract characteristic and a pitch characteristic to extract a vocal tract characteristic cepstrum vector by extracting the vocal tract characteristic cepstrum vector, Means for converting into an Nth-order vocal tract characteristic logarithmic amplitude coefficient vector on the linear frequency axis; and interpolating the Nth-order vocal tract characteristic logarithmic amplitude coefficient vector to spectral components on equal division points on the mel frequency axis. Means for converting the vector component subjected to the smoothing process to an Nth vocal tract characteristic logarithmic amplitude mel scale coefficient vector adapted to the human auditory frequency resolution characteristic; Means for transforming into a Qth-order mel-cepstral vector by inverse discrete cosine transform on the time axis. The placing of voice recognition dialing devices. 8. The state distance calculator calculates an input feature vector obtained by the speech analysis feature extractor from a frame PCM signal obtained by AD-converting an analog signal of an input speech and further dividing the frame, and a standard semisyllable hidden Markov model. The voice recognition dial device according to claim 5, further comprising means for calculating a state distance value with respect to a standard feature vector of all states by a predetermined distance calculation formula. 9. The accumulative distance calculator calculates a state combination information of a half-syllable hidden Markov model state code data string for a name or a telephone number to be recognized, each state distance value, and a state stored in a standard pattern ROM. 6. The voice recognition dial device according to claim 5, further comprising means for calculating a cumulative distance value for all frames of the input voice by a frame synchronization Viterbi algorithm using the selected distance value.