JP2001265383A - Voice recognizing method and recording medium with recorded voice recognition processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve voice recognizing performance on hardware composed of a CPU which has a small memory and small processing capability. SOLUTION: Temporary phoneme speech continuing times T11, T12... by phonemes constituting an input voice are obtained, the voice section of the input voice data is divided by the mentioned phoneme speech continuing times, and the path from the 1st-stage phoneme model to the final-stage phoneme model of HMM is given limitations on some phoneme model in sections (other than the sections shown by thick-line arrows) from certain time to certain time based upon the temporary phoneme speech continuing times. The limitations of the path is the control for inhibiting the transition from the final state of one phoneme model to the initial state of the next phoneme model from certain time to certain time based upon the temporary phoneme speech continuing times.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method for improving speech recognition performance on inexpensive hardware constituted by a CPU having a small memory capacity and a low arithmetic capacity, and a recording medium for recording a speech recognition processing program. About.

[0002]

2. Description of the Related Art Voice recognition technology has been widely used in various fields. The problem with speech recognition is that even for the same word, the voice pattern changes due to the difference in the speaker or the way of speaking at the same speaker, or the voice due to the phonemic environment existing before and after. Recognition processing is performed in response to fluctuations in voice patterns due to various factors such as fluctuations in patterns (articulation coupling), and fluctuations in voice patterns due to fluctuations in the length of time required to finish speaking the same word. There is a need.

[0003] In consideration of such problems, HMMs have been conventionally used as phoneme models for obtaining high speech recognition performance.
(Hidden Markov Model) is well known.

In this HMM, as a technique for further improving the recognition rate by a simple method, the utterance duration for each phoneme is considered as a duration distribution, and the likelihood obtained from the phoneme utterance duration is recognized. There is known a method of considering the likelihood of the phoneme (the output likelihood of the HMM) in the process of performing the process.

[0005]

However, although such a method certainly leads to an improvement in the recognition rate, a calculation for obtaining the certainty obtained from the phoneme utterance duration is required, and such a calculation is required. It is necessary to enhance the hardware by mounting a CPU or a memory having a processing capability for performing the processing. However, when speech recognition technology is used for products requiring small size, light weight, and low cost, such as toys, the scale of hardware used is also greatly restricted, so that a CPU or memory capable of performing the above calculations cannot be mounted. is the current situation.

On the other hand, however, it is certain that speech recognition taking into account the duration of phoneme utterance can contribute to an improvement in the recognition rate. Therefore, speech recognition taking into account the duration of phoneme utterance without requiring a large amount of calculation is required. It is desired to make it possible.

SUMMARY OF THE INVENTION It is an object of the present invention to improve the speech recognition performance on inexpensive hardware including a small memory capacity and a low CPU.

[0008]

In order to achieve the above-mentioned object, a speech recognition method according to the present invention comprises combining a phoneme model for each phoneme constituting a recognizable word to form a phoneme connection model of the word. , For that phoneme concatenation model,
By providing the time-series speech data of the input speech, the phonemes existing in the final stage after passing through a predetermined path in the state transition algorithm from the first stage phoneme model to the final stage phoneme model constituting the phoneme connection model. A speech recognition method for obtaining an output likelihood from the final state of the model and performing speech recognition from the magnitude of the output likelihood, wherein each of the phonemes constituting the input speech is obtained from time-series speech data of the input speech. On the other hand, the utterance duration of each phoneme is obtained as a tentative phoneme utterance duration based on standard data for each phoneme, and the input speech is determined based on the tentative phoneme utterance duration of each phoneme constituting the input speech. The voice section of the data is divided, and a certain phoneme is passed from the first-stage phoneme model to the last-stage phoneme model in the path in the state transition algorithm. To phoneme models, so that a limit of path section up to a certain time from a certain time based on the phoneme speech duration of the temporary.

In such a speech recognition method, the restriction of a path to be performed in a section from a certain time to a certain time based on the tentative phoneme utterance continuation time is determined from the last state of a certain phoneme model to the first state of the next phoneme model. Is a control for prohibiting the transition to the period from a certain time to a certain time based on the temporary phoneme utterance duration.

Further, the recording medium storing the speech recognition processing program of the present invention forms a phoneme connection model of the word by combining the phoneme models of the phonemes constituting the recognizable word. By providing the time-series speech data of the input speech, it passes through a predetermined path in the state transition algorithm from the first-stage phoneme model to the last-stage phoneme model constituting the phoneme connection model, and exists at the last stage. A recording medium on which a speech recognition processing program for obtaining an output likelihood from the final state of the phoneme model and performing speech recognition based on the magnitude of the output likelihood is recorded, wherein the speech recognition processing program is used for the input speech. For each phoneme composing the input speech from the sequence speech data, the utterance duration of each phoneme is based on the standard data for each phoneme. Obtaining a tentative phoneme utterance duration, dividing a voice section of the input voice data based on the tentative phoneme utterance duration for each phoneme constituting the input voice, and From the path in the state transition algorithm in the phoneme model of the last stage to the phoneme model for a certain phoneme, the procedure of setting a path restriction in a section from a certain time to a certain time based on the temporary phoneme utterance duration. .

[0011] In the speech recognition processing program in the recording medium on which such a speech recognition processing program is recorded, the restriction on the path to be performed in a section from a certain time to a certain time based on the provisional phoneme utterance duration is limited to a certain phoneme model. This is control for prohibiting transition from the final state to the first state of the next phoneme model from a certain time to a certain time based on the provisional phoneme utterance duration.

As described above, according to the present invention, the tentative phoneme utterance duration for each phoneme constituting the input speech is obtained based on the standard data for each phoneme, and the input speech is converted into the speech section of the input speech data. It is divided by the provisional phoneme utterance duration for each constituent phoneme, and in the path from the first-stage phoneme model to the final-stage phoneme model in the HMM, for a certain phoneme model, The path is restricted in a section from a certain time to a certain time based on the path. The restriction of the path performed based on the temporary phoneme utterance duration means that the transition from the final state of a certain phoneme model to the first state of the next phoneme model is performed at a certain time based on the temporary phoneme utterance duration. This is control to prohibit until a certain time.

As described above, by providing a restriction on the path, the state transition algorithm for the time-series audio data for the input audio is such that the final state can be reached through a path in an unrestricted range. Becomes For this reason, an appropriate output likelihood is obtained for the input speech, and erroneous recognition can be reduced.

For example, when speech data of a word containing many phonemes similar to the word is input to the phoneme connection model of the HMM set to obtain a high output likelihood for a certain word, Since the sound has to pass through a path in an unrestricted range, the final output likelihood can be reduced as compared with a case where the path is not restricted. In other words, when the path is not restricted, a high state probability is obtained for a certain temporary phoneme utterance duration of a certain phoneme model part, which is the output of the final state of the phoneme model existing at the final stage. Affecting the likelihood and consequently setting the output likelihood in the final state to a high value, causing misrecognition. However, by limiting the path as in the present invention, Such a problem can be solved and the recognition rate can be improved.

[0015]

Embodiments of the present invention will be described below. The contents described in this embodiment are as follows.
A description of the speech recognition method of the present invention,
It also includes a specific processing procedure of the voice recognition processing program on the recording medium storing the voice recognition processing program of the present invention.

The present invention performs speech recognition in consideration of the phoneme utterance duration of phonemes in order to improve the recognition performance of words having similar speech patterns. In this embodiment, it is assumed that speech recognition by HMM is performed.

First, the phoneme utterance duration for each phoneme is determined. In this method, an average phoneme utterance duration is obtained for each phoneme from standard data for each phoneme. For example, the average utterance duration of phoneme "a" is 100m
sec, the average of the utterance duration of the phoneme “i” is 120 msec,
The average utterance duration of each phoneme is determined such that the average of the utterance duration of the phoneme "u" is 100 msec and the average of the utterance duration of the phoneme "e" is 110 msec.

Then, for some words recognizable by the system, the ratio of the utterance duration for each phoneme constituting the word is determined. For example, the recognizable word "Sato"
If there is the word “sato”, the ratio of the phoneme utterance durations of the phonemes “s”, “a”, “t”, and “o” constituting the “sato” is obtained. For example, FIG.
It is assumed that there is audio data of "sato" as shown in FIG. The voice section of this voice data is 370 msec
If the individual frame length in the audio feature analysis processing is 20 msec and the frame shift length is 10 msec, this audio data is composed of 36 frames.

According to the speech duration of each phoneme of "s", "a", "t", and "o" obtained from the above-described standard voice data, the ratio of each phoneme speech duration is
Assuming that 1: 2: 1: 2 is obtained, as shown in FIG. 1B, 36 frames are divided into the first to sixth frames (the number of frames is six) and the seventh to eighteenth frames. It can be divided into frames (12 frames), 19th to 24th frames (6 frames), and 25th to 36th frames (12 frames). An approximate utterance duration for each phoneme can be obtained from the obtained number of frames.

This approximate utterance duration is referred to herein as a temporary phoneme utterance duration. Further, a temporary boundary of each phoneme (referred to as a temporary phoneme boundary) can be obtained from the temporary phoneme utterance duration obtained in this way. In this case, as shown in FIG. 1B, the provisional phoneme utterance duration of the phoneme "s" is T1, the provisional phoneme utterance duration of the phoneme "a" is 2T1, and the provisional phoneme utterance of the phoneme "t". The duration is T1, and the temporary phoneme utterance duration of the phoneme "o" is 2T1.
And the boundary of each phoneme (temporary phoneme boundary) is p
1, p2, and p3.

Next, at the time of trellis or Viterbi calculation in the HMM, the transition from the final state of a certain phoneme model to the first state of the next phoneme model is performed between a certain time based on the temporary phoneme duration and a certain time. Restrictions are set to prohibit

Now, for the word "Sato", the phoneme models of the phonemes "s", "a", "t" and "o" constituting this word are as shown in FIGS. 2 (a) to 2 (d). Are represented by four states and three loops, respectively. And these,
When the phoneme model of "s", the phoneme model of "a", the phoneme model of "t", and the phoneme model of "o" are respectively connected,
The result is as shown in FIG.

As can be seen from FIG. 2 (e), each of the phoneme models "s", "a", and "t" has a final stage, that is, a state without a loop ( As shown in FIGS. 2A to 2D, the state of the phoneme model “s” is state S14, and the state of the phoneme model of “a” is state S14.
24, in the phoneme model of "t", the state S34) is removed and combined, and the phoneme model of "o" has a state S44 having no loop at the final stage.

That is, as shown in FIG. 2E, the phoneme connection model for the word "Sato" has 12 states S11, S12, S13, S2 having a loop.
1, S22, S23, S31, S32, S33, S4
1, S42, S43 and state S without loop at the last stage
44 can be considered.

Thus, the phoneme connection model passes through the path from the first state to the phoneme model existing at the last stage, and the time of the state S44 existing at the last stage by the Viterbi or trellis algorithm. tn
, The output likelihood as the final state probability is obtained.
The final output likelihood in this case is obtained by giving time-series speech data for a certain input speech to the phoneme connection model as shown in FIG. Is determined by the magnitude of the value of. In this case, since the phoneme connection model of FIG. 2E is a phoneme connection model for "sato", a high output likelihood is obtained when the speech "sato" is input.

Here, in the present invention, based on the tentative phoneme utterance duration, a limit is imposed on the path from the final state of a certain phoneme model to the first state of the next phoneme model from a certain time to a certain time. . For example, in the example of the phoneme connection model shown in FIG. 2E, the initial state S2 of the phoneme model “a” is changed from the final state S13 of the phoneme model “s”.
1. From the final state S23 of the phoneme model “a” to the first state S31 of the phoneme model “t”, etc., the path is restricted from a certain time based on the temporary phoneme utterance duration to a certain time. Thus, an attempt is made to improve the recognition rate. Hereinafter, description will be made with reference to a simple example.

In the above description, each phoneme model has been described by taking a phoneme model of four states and three loops as an example, but here, in order to prevent the drawing from becoming complicated and to simplify the description. A description will be given using a phoneme connection model obtained by connecting three-state two-loop phoneme models.

FIG. 3 is a diagram for explaining a state transition algorithm of a phoneme connection model in the HMM for a certain word composed of three phonemes. As shown in FIG. 4, the phoneme connection model includes six states S1, S2, S
3, S4, S5, S6 and a state S7 without a loop connected to the last stage. In FIG. 4, S1 and S2 correspond to the first-stage phonemes, S3 and S4 correspond to the second-stage phonemes, and S5, S6 and S7 correspond to the third-stage phonemes, respectively. A11 is the transition probability of the loop in the state S1, a12 is the transition probability from the state S1 to the state S2,
a22 is the transition probability of the loop in the state S2, a23
Is the transition probability from the state S2 to the state S3, and a33 is the state S
3, a34 is the transition probability from state S3 to state S4, a44 is the transition probability of the loop in state S4, a45 is the transition probability from state S4 to state S5, and a55 is the transition probability of the loop in state S5. ,
a56 is the transition probability from state S5 to state S6, a66
Is the transition probability of the loop in the state S6, and a67 is the state S
6 represents the transition probability from state 6 to state S7.

Here, it is assumed that a provisional phoneme utterance duration is obtained based on the standard data for each phoneme. For example, as shown in FIG. 3, among the phonemes constituting the word, times t0 to t5 are provisional phoneme utterance durations T1 for the phonemes of the first row, and times t6 to t10 are phonemes of the second row of the word. Tentative phoneme utterance duration T2 for
Times t11 to t17 are provisional phoneme utterance durations T3 for the phonemes of the third row of the word. In FIG. 3, p1 and p2 indicate temporary phoneme boundaries.

It is assumed that in such a state transition algorithm, a path is restricted so that the state probability of the hatched portion becomes zero. That is, in this case, during the provisional phoneme utterance duration T1 (from time t0 to t5), the transition probability a45 = 0 from the state S4 of the second-stage phoneme to the state S5 of the third-stage phoneme is set to 0, and the provisional During the phoneme utterance duration T3 (from time t11 to t17), the state S2 in the first-stage phoneme changes to the state S in the second-stage phoneme.
The transition probability a23 to 3 is set to 0. As a result, the state S in the provisional phoneme utterance continuation time (time t0 to t5) is obtained.
5, the state probability S5 (t) = 0, the state probability S3 of the state S3 in the temporary phoneme utterance duration (t11 to t17)
(T) = 0.

By providing such a path restriction, when time-series input speech data from time t0 to time t17 is given to this algorithm, the first stage of the phoneme connection model as shown in FIG. When the output likelihood is obtained from the state FS (see FIG. 3) of the phoneme model existing at the final stage after passing through the path from the phoneme model of the second stage to the phoneme model existing at the last stage, the path is restricted. Therefore, the state probability due to the restricted path does not affect the state probability of the final state FS (output likelihood of this phoneme model).

By the way, even if the path restriction as shown in FIG. 3 is not provided, as a result, the path restriction is naturally applied as shown in FIG. That is, there is originally a path that does not contribute to obtaining the final output likelihood from the state FS of the last stage (the shaded portion in FIG. 5), but in the present invention, it is shown in FIG. In addition to the paths that do not contribute to obtaining the final output likelihood in the first place as described above, restrictions on the paths indicated by oblique lines are added as shown in FIG.

That is, the shaded portion in FIG. 6 is a path which does not contribute to obtaining the final output likelihood (same as that shown in FIG. 5), and is a path which is naturally restricted. You can say that. And in addition to this,
By adding the path restriction of the present invention as described with reference to FIG. 3, as a result, the path of the hatched portion and the hatched portion in FIG. 6 is restricted.

FIG. 5 shows an example of a path obtained by the Viterbi algorithm between the algorithm that does not impose the restriction as shown in FIG. 5 and the algorithm that restricts the path according to the present invention (FIG. 6). 7 and FIG. FIG. 7 corresponds to FIG. 5, and the way of limiting is exactly the same as FIG. FIG. 8 corresponds to FIG. 6, and the way of limiting is exactly the same as that of FIG.

The example of FIG. 5 does not impose an active restriction on the path, and the restriction on the path is loose. Therefore, in an extreme example, there is a possibility of passing the path shown by the thick line in FIG. is there. That is, in this case, the state probabilities in the first-stage phoneme model are large values around the temporary phoneme utterance duration T2 for the second-stage phoneme and near the temporary phoneme utterance duration T3 for the third-stage phoneme. This is an example in which erroneous recognition has occurred.

That is, for example, when there is a phoneme model of the word "Aki" and a speech "Aki" is input to the phoneme model, the final output likelihood of the phoneme model becomes a high value. . That is, if the phoneme connection model shown in FIG. 7 is a phoneme connection model for recognizing the word “Aki”, if the time series data of the input voice “Aki” is given, the final From the state FS, a high output likelihood is obtained, whereby the input speech is recognized as being "vacant".

On the other hand, when a voice such as "Aka" is given to this phoneme connection model, the output likelihood lower than when "Aki" is input is obtained from the final state FS of the "Aki" phoneme model. Degree must be obtained. However, in a path with a moderate restriction as shown in FIG.
The phoneme part of “a” that is included in the phonemes “a”, “k”, and “a” that make up “red” has a high state probability in the phoneme model of “a” of “aki”, This may increase the output likelihood while leaving the output likelihood in the final state FS unaffected.

In order to prevent this, according to the present invention, the paths are restricted as shown in FIG.
Thus, a restricted path is formed. According to this, when uttering "red", the voice has to pass through an unrestricted range path,
Among the respective phonemes for the voice data "aka", the state probability in the vicinity of the provisional phoneme utterance duration T2 for the second stage phoneme and the vicinity of the provisional phoneme utterance duration T3 for the third stage phoneme Has a lower value than that in FIG. 5, which also affects the output likelihood in the final state FS, and can reduce the output likelihood.

On the other hand, if "Aki" is uttered to such an "Aki" phoneme model, it is possible to pass through a path as shown in FIG. , The optimal state probabilities are obtained at the tentative phoneme utterance duration for each phoneme without being significantly affected by the restriction, thereby increasing the output likelihood from the final state FS. Can be obtained.

As described above, the recognition rate can be improved by providing restrictions on the paths. As one specific example, for example,
As shown in FIG. 9, the time series data (time t0 to tn) of the voice of “Good morning” is based on the standard data for each phoneme as described above.
Each phoneme "o", "h", "a", "y",
Temporary phoneme utterance duration T11, T12, T for each "o"
13, T14, T15, temporary phoneme boundaries p0, p
1, p2,..., P5 are determined.

In this case, in the phoneme models of the phonemes “o”, “h”, “a”, “y”, and “o”, the phoneme “h” is changed from the final state of the phoneme model of the phoneme “o”. To the first state, the path between the provisional phoneme boundary p0 and the provisional phoneme boundary p2 (portion indicated by a thick line) is allowed to pass, and the other path (portion indicated by a thin line arrow) is restricted. (The state probability between them is set to 0). Further, the transition from the final state of the phoneme model of the phoneme “h” to the first state of the phoneme “a” is performed by a path (the thick arrow) from the intermediate point between the temporary phoneme boundaries p0 and p1 to the temporary phoneme boundary p3. Part) is allowed to pass, and the other paths (fine arrow parts) are restricted (state probability between them is 0). Further, the transition from the final state of the phoneme model of the phoneme “a” to the first state of the phoneme “y” is performed by a path (the thick arrow) from the intermediate point between the temporary phoneme boundaries p1 and p2 to the temporary phoneme boundary p4. Part) is allowed to pass, and the other paths (fine arrow parts) are restricted (state probability between them is 0). Further, the transition from the final state of the phoneme model of the phoneme “y” to the first state of the phoneme “o” is performed by a path (the thick arrow) between the intermediate point between the temporary phoneme boundaries p2 and p3 and the temporary phoneme boundary p5. Part) is allowed to pass, and the other paths (fine arrow parts) are restricted (state probability between them is 0).

Experiments have confirmed that good recognition results can be obtained when such path restrictions are applied.

Although the above description is based on the Viterbi algorithm, the same concept can be applied to the trellis algorithm.

The above-described speech recognition processing program for performing the processing of the present invention can be recorded on a recording medium such as a floppy disk, an optical disk, or a hard disk. The present invention also includes the recording medium. Further, the processing program may be obtained from a network.

[0045]

As described above, according to the present invention, the tentative phoneme utterance duration for each phoneme constituting the input voice is obtained based on the standard data for each phoneme, and the tentative phoneme utterance duration is set in the voice section of the input voice data. On the other hand, the input speech is divided by the provisional phoneme utterance duration for each phoneme constituting the input speech, and in the path from the first-stage phoneme model to the last-stage phoneme model in the HMM, the temporary Is limited in a section from a certain time to a certain time based on the phoneme utterance continuation time. The restriction of the path performed based on the temporary phoneme utterance duration means that the transition from the final state of a certain phoneme model to the first state of the next phoneme model is performed at a certain time based on the temporary phoneme utterance duration. This is control to prohibit until a certain time.

As described above, by providing a restriction on the path, the state transition algorithm for the time-series audio data for the input audio can be such that the final state is reached through a path in an unrestricted range. Becomes For this reason, an appropriate output likelihood is obtained for the input speech, and erroneous recognition can be reduced.

For example, when speech data of a word including many phonemes similar to the word is input to the phoneme connection model of the HMM set to obtain a high output likelihood for a certain word, Since the sound has to pass through a path in an unrestricted range, the final output likelihood can be reduced as compared with a case where the path is not restricted. In other words, when the path is not restricted, a high state probability is obtained for a certain temporary phoneme utterance duration of a certain phoneme model part, which is the output of the final state of the phoneme model existing at the final stage. Affecting the likelihood and consequently setting the output likelihood in the final state to a high value, causing misrecognition. However, by limiting the path as in the present invention, Such a problem can be solved and the recognition rate can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating an example of obtaining temporary utterance duration for each phoneme included in an input voice data with respect to certain input voice data.

FIG. 2 is a diagram schematically showing a state transition of a phoneme connection model obtained by connecting a plurality of phoneme models (four states and three loops) to a plurality of phonemes constituting a certain word;

FIG. 3 is a diagram showing a state in which three-state two-loop phoneme models are connected.
It is a figure explaining the example which gave the restriction of the pass of the present invention to the phoneme connection model of the phoneme word.

FIG. 4 is a diagram showing a concatenation of three-state two-loop phoneme models.
It is a figure which shows the phoneme connection model of a phoneme word typically.

FIG. 5 is a diagram showing a concatenation of three-state two-loop phoneme models.
FIG. 11 is a diagram illustrating the existence of a path that does not affect the output likelihood in the final state for a phoneme link model of phoneme words.

FIG. 6 is a diagram illustrating a path which does not affect the output likelihood in the final state by providing the path restriction of the present invention shown in FIG. 3;

FIG. 7 is a diagram illustrating an example of a path obtained by an algorithm in the phoneme connection model illustrated in FIG. 5 and illustrating an example in which erroneous recognition has occurred.

FIG. 8 is a diagram illustrating an example of a path obtained by an algorithm in the phoneme connection model illustrated in FIG. 6, and illustrating an example in which proper recognition is performed.

FIG. 9 is a diagram illustrating an example of a restriction on a path that enables proper recognition of a certain word.

[Explanation of symbols]

 p0, p1, ... provisional phoneme boundary T1, T2, ... provisional phoneme utterance duration FS Final state in phoneme model at final stage

Claims (4)

[Claims] 1. A phoneme connection model of a word is constructed by combining phoneme models for each phoneme constituting a recognizable word, and time-series speech data of an input speech is given to the phoneme connection model to provide the phoneme connection model. The output likelihood is obtained from the final state of the phoneme model existing at the final stage through a predetermined path in the state transition algorithm from the first stage phoneme model to the final stage phoneme model constituting the phoneme connection model. A speech recognition method for performing speech recognition based on the magnitude of output likelihood, wherein for each phoneme constituting the input speech from time-series speech data of the input speech, the utterance duration of each phoneme is determined for each phoneme. Of the input speech data based on the provisional phoneme utterance duration of each phoneme constituting the input speech. In the path of the state transition algorithm in the first-stage phoneme model to the last-stage phoneme model from the first-stage phoneme model, the phoneme model for a certain phoneme is at a certain time based on the temporary phoneme utterance duration. A speech recognition method, wherein a path is restricted in a section up to a time. 2. The method according to claim 1, wherein the restriction of a path to be performed in a section from a certain time to a certain time based on the temporary phoneme utterance continuation time is based on a transition from a final state of a certain phoneme model to a first state of the next phoneme model. 2. The speech recognition method according to claim 1, wherein control is performed to prohibit a period from a certain time to a certain time based on the phoneme utterance continuation time. 3. A phoneme concatenation model of the word is constructed by combining phoneme models for each phoneme constituting a recognizable word, and time-series speech data of an input speech is given to the phoneme concatenation model. The output likelihood is obtained from the final state of the phoneme model existing at the final stage through a predetermined path in the state transition algorithm from the first stage phoneme model to the final stage phoneme model constituting the phoneme connection model. A recording medium that stores a speech recognition processing program that performs speech recognition from the magnitude of output likelihood, and the speech recognition processing program stores, for each phoneme constituting the input speech from time-series speech data of the input speech. On the other hand, the procedure for obtaining the utterance duration of each phoneme as a temporary phoneme utterance duration based on the standard data for each phoneme, A step of dividing a speech section of the input speech data based on a provisional phoneme utterance duration for each phoneme to be composed; A step of providing a path restriction for a section from a certain time to a certain time based on the provisional phoneme utterance duration with respect to the phoneme model corresponding to. 4. Limiting a path to be performed in a section from a certain time to a certain time based on the tentative phoneme utterance continuation time is performed by changing a transition from a final state of a certain phoneme model to a first state of the next phoneme model. 4. A recording medium storing a speech recognition processing program according to claim 3, wherein the control is for prohibiting a period from a certain time to a certain time based on the phoneme utterance continuation time.