JP3187241B2 - Speech speed converter - Google Patents

Abstract

PURPOSE:To eliminate the dispersion of a speech speed conversion effect on hearing caused by the differences in the length of each voiced sound segment. CONSTITUTION:Corresponding to the segment length, the voiced sound segment, which is shorter than a certain value, is expanded by a higher magnification so as to obtain a speach speed conversion effect in accordance with a desired magnification on hearing. For example, for a short voiced sound having a less than 150ms duration, an expansion magnification, which corresponds to the length of a voiced sound, is given along a magnification function g(w) as shown in the diagram. Moreover, when the method is to be applied to a reference magnification which changes a speech speed from a 'slow' to a 'fast' within the segment generated in one breath, the voiced sound segment, which has a duration of less than 150ms and appears within 450ms from a starting point in the segment, is given an expansion magnification corresponding to the length of a voiced sound along the function g(w) which is shown in graphic independent of its appearance time. If the voiced sound which exceeds 150ms or the duration time exceeds 450ms, a conventional expansion magnification f(t) is applied.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech speed conversion device ,
In particular, in hearing aids for the hearing impaired and the elderly, general language learning devices, radios, tape recorders, telephones, etc. Video tape recorder,
The present invention relates to a real-time speech speed conversion device that effectively absorbs a gap between video and audio generated when audio output of a video disc player or the like is converted into speech speed.

[0002]

2. Description of the Related Art A method for extending a voiced section at a uniform magnification when converting speech speed (Akira Nakamura et al., Spring Meeting of Acoustical Society of Japan 1992, "High-quality real-time speech speed conversion system"
2-6-1P. 329-P. 330 (1992-3))
And a speech rate conversion method that varies the magnification as a function of the elapsed time from the start of the utterance (Ryu Ikezawa et al., Spring Meeting of the Acoustical Society of Japan in 1992, “One way to absorb the time extension associated with speech rate conversion” Method "2-6-2 P.331-P.332 (19
92-3)) exist, but in each case, the expansion ratio is uniquely determined as a function of the elapsed time irrespective of the length of each voiced section. It can not be said to give the same “slow feeling”,
"Dispersion" may occur in the effect on hearing. Heretofore, there has been no speech speed conversion technology that solves this and stably and naturally converts in terms of hearing.

[0003]

SUMMARY OF THE INVENTION In order to make the speech speed of an input voice "slow", a voiceless section, a voiceless section, and a voiced section are separated, and a voiced section is maintained without changing the length of the voiceless section and the voiceless section. When performing voice speed conversion by expanding a section, if a plurality of voiced sections in the voice are expanded at a uniform magnification, a difference occurs in the degree of "slow feeling" in auditory sense depending on the section length of each voiced section. (Atsushi Imai et al., Autumn Meeting of the Acoustical Society of Japan in 1993, "Real-time absorption method of time expansion accompanying speech speed conversion" 1-9-10P.
361-P. 362 (1993- 10)).

In a voice, a relatively long voiced section exceeding 300 ms, such as a chain of different vowels or a long vowel, or conversely, less than 100 ms in unvoiced sections or vowels sandwiched between silent sections. There are cases where relatively short voiced sections appear one after another. For example, when a voice mixed with both voices is listened to a voice expanded at the same constant magnification, the long voiced section is expanded in units of one voiced section. In contrast to a large absolute amount of time and a large "slow feeling" in the auditory sense, a short voiced section has a small absolute amount of decompression time, and in some cases, almost no "slow feeling" is felt.

For example, when the section lengths of 350 ms and 80 ms are uniformly extended 1.5 times by the conventional method, 5
It is converted into 25 ms and 120 ms. The absolute extension of the extension time of the former is 175 ms, whereas the latter is extended by only 40 ms, which appears as a difference in the effect on hearing. Therefore, when there are a plurality of voiced section distributions of various lengths in a series of input voices as described above, the voices are converted into unstable voices having undetermined speech speeds, and in some cases this is quite anxious. is there.

Also, a method of absorbing time expansion associated with speech rate conversion, which has already been proposed (Ryu Ikezawa et al., Spring Meeting of the Acoustical Society of Japan in 1992, entitled "One way to absorb time expansion associated with speech rate conversion". Method "2-6-2 P.331-P.332
(1992-3)), the expansion rate of the voiced section at the start point of the section (phrase) uttered in a short breath is set high, and the speech speed is gradually increased, so that the converted speech as a whole " Slowness "and absorption of time expansion as a whole are realized, but in the case of speech in which short voiced sections appear one after another near the starting point of this phrase, a relatively high magnification is used for the above-described reason. Does not provide a "slow feeling", and as a result, only the high-speed part of the latter half becomes noticeable, and the expected effect may not be obtained.

[0007] The above-mentioned problems will be shown in more concrete cases.

(1) Section (phrase) uttered in a short breath
Is fixed at 2000 ms, and the expansion ratio r is changed from rs (rs> 1) to re (re <re) according to the curve shown in FIG.
Monotonically decrease to 1).

(2) When the time exceeds 2000 ms, the magnification is appropriately corrected according to the change in the pitch frequency.

This technique is introduced into a real-time speech rate conversion system, and as a result of converting a large number of news voices, the effect expected for some phrases, particularly, a "slow" feeling near the starting point of the phrases is generated. Some effects were not obtained. FIG. 2 shows one example of a particularly effective phrase ((a) in FIG. 2) and two examples of a phrase in which no effect was particularly felt ((b),
Regarding (c)), the distribution of the voiced section length in the phrase on the time axis is shown.

The following points are mentioned as the tendencies typified by these three examples.

(1) When there are a plurality of relatively long voiced sections exceeding 150 ms within 450 ms to 500 ms of the beginning of the sentence, the effect is large even if the expansion ratio r is r = 1.4.

(2) When a relatively short voiced section of 150 ms or less exists at the beginning of a phrase, even if r = 2.0, the effect is small.

As a result of examining other phrases, a similar tendency was found.

SUMMARY OF THE INVENTION The present invention has been made in view of the above-described problems. An object of the present invention is to provide a speech speed conversion effect caused by a difference in the length of a voiced segment of an input voice when speech speed conversion is performed by expanding a voiced segment. It is an object of the present invention to provide a speech speed conversion device which eliminates variations in the sense of hearing and can obtain a natural and stable speech speed conversion effect for any input voice.

[0016]

In order to achieve the above object, the present invention separates a silent section, an unvoiced section, and a voiced section of an input voice, and expands the voiced section to speed up the utterance. When performing a conversion that reduces the voice speed while maintaining the voice pitch, the time length of each voiced section is detected sequentially, and the time length of each voiced section is a uniform value or smooth with the elapsed time. By multiplying the reference magnification that changes
A speech speed conversion device that obtains an audible effect corresponding to the magnification, wherein a time length of a voiced section to be converted is a predetermined length.
Determining means for determining whether or not:
The fruit, when said time length conversion subject to voiced segment exceeds the <br/> predetermined length is multiplied by a reference factor in the appearance time of the voiced, voiced section serving as said converted For a voiced section whose time length is equal to or less than the predetermined length,
Calculating means for multiplying the expansion rate higher than the reference magnification according to the time length of the voiced section.

In the present invention, preferably, the arithmetic means
For a short voiced section equal to or less than 150 ms corresponding to the predetermined length, regardless of the appearance time of the voiced section, the voiced section is provided along a magnification function that provides a higher expansion magnification than the reference magnification. In the case of a voiced section exceeding 150 ms, the time length of the voiced section is multiplied by the reference magnification.

Preferably , the present invention sets a slow speech speed at a start point of this section in units of a section generated in a breath as a reference magnification that smoothly changes with the elapsed time, and terminates the section. When applying a magnification function characterized by gradually increasing the speech speed toward a point,
The arithmetic means is provided within a fixed time from the start time of the section,
Preferably the predetermined interval length which corresponds to the length appearing within about time 450 ms, preferably for a voiced having a section length less than approximately 150 ms, regardless of the appearance time of the voiced segment, the reference magnification A stretching function corresponding to the time length of the voiced section is always set along the scaling function that provides a higher stretching magnification compared to the above, and if the voiced section exceeds 150 ms and the elapsed time exceeds 450 ms, the time length of the voiced section is changed. Is multiplied by the reference magnification.

The above values of 150 ms and 450 ms specifically illustrate one of the preferable values.
The present invention is not limited to this value.

In the present invention, preferably, the predetermined length is a maximum time length of a voiced section in which a “slow feeling” of the converted voice cannot be audibly sensed when a practical value is set as a reference magnification. For a voiced section shorter than the maximum time length, a new magnification function g (w) having the time length w as a variable is introduced, and an expansion magnification is given according to the magnification function. The magnification is a higher value than the reference magnification, and in particular, has a property that the shorter the voiced section, the higher the magnification, and the maximum amplification value and the minimum amplification value of the magnification by the same magnification function are not fixed, Each of the reference magnifications may be proportionally changed by a value of a reference magnification function f (t).

[0021]

According to the present invention, attention is paid to the length of a voiced section which affects the degree of the effect of speech rate conversion, and a certain length or less is used so that a speech rate conversion effect that is natural and stable in terms of audibility can be obtained. A new function is applied to a voiced section having a shorter length corresponding to the section length, such that a shorter one gives a higher expansion ratio. As a result, it becomes possible to naturally convert various input voices to a desired speech speed with a stable effect. In particular, a method of absorbing time expansion by speech rate conversion proposed in the past (Ryu Ikezawa et al., Autumn Meeting of the Acoustical Society of Japan in 1992, "A study of real-time absorption of time expansion in speech rate conversion" 2-9-2P .349-P.350
(1993-10)), the disadvantage that the "slow feeling" near the starting point of the utterance was unstable in the sense of hearing can be solved, and a stable and more effective converted voice can be obtained. Become.

[0022]

Embodiments of the present invention will be described below in detail with reference to the drawings.

In a section predicted to be uttered in a short breath,
A "real-time speech rate conversion method" in which the speech speed is set to be "slower" than the original speech at the start point of this section, and the speech speed is increased according to a certain rule toward the end point. (Ryu Ikezawa et al., Autumn Meeting of the Acoustical Society of Japan in 1992, "Examination of real-time absorption of time expansion in speech speed conversion," 2-9-2 P. 349-P. 350 (199
An example in which the present invention is applied to 3-10)) will be described. This "real-time speech speed conversion method" is a method of absorbing time expansion in a speech speed conversion device that operates in real time, and this requires a "slow feeling" at the start point of each phrase. is there. It can be said that applying the method according to the present invention to this is particularly effective.

FIG. 3 shows an outline of the operation of one embodiment of the present invention. For a short voiced sound of 150 ms or less that appears within 450 ms from the beginning of the sentence, the length of the voiced sound corresponds to the length of the voiced sound along the scaling function g (w) as shown in FIG. 3 regardless of the appearance time of the voiced sound. Give the stretching magnification. If the voiced sound exceeds 150 ms and the elapsed time exceeds 450 ms, the conventional expansion factor curve f (t) (FIG. 1) is applied.

4 to 7 show an embodiment of the present invention in more detail.

FIG. 4 is a block diagram showing the overall circuit configuration of one embodiment of the present invention.

The real-time speech speed converter shown in FIG.
Voice input circuit 1 and CPU (central processing unit) circuit 2
A PROM (programmable ROM) circuit 3, an input buffer circuit 4, a processing buffer circuit 5, a file circuit 6, an audio output circuit 7, and a bus 8. Then, the voice (original voice) to be subjected to speech speed conversion is fetched by the voice input circuit 1, and a change in the pitch (pitch frequency) of the voice of the original voice is detected by real-time processing. By changing the speech speed according to the rule of slowing down the speech speed in the high voice part and increasing the speech speed in the low voice portion, the speech time of the original voice is maintained,
Converts the original sound into good sound that is easy to listen to.

The audio input circuit 1 is a circuit having a general configuration for inputting an original audio, for example, a microphone, a tone control circuit, an A / D (analog / digital) converter, an audio storage / reproduction circuit, an audio recording medium (for example, , An IC memory, a hard disk, a floppy disk or a VTR (video tape recorder)), an interface circuit, and the like. The audio to be converted is converted into a digital audio signal and converted. The digital audio signal is supplied to the input buffer circuit 4 in frame units based on an instruction from the CPU circuit 2.

The input buffer circuit 4 has a required capacity RA.
M (random access memory), etc.
This portion is used as a work area of the CPU circuit 2 and takes in and stores an audio signal output from the audio input circuit 1 and processes the stored audio signal based on an instruction from the CPU circuit 2. The data is transferred to the buffer circuit 5.

The processing buffer circuit 5 has a required capacity RA.
M, and is used as a work area of the CPU circuit 2. The audio signal output from the input buffer circuit 4 is fetched and stored.
The stored audio signal is transferred to the file circuit 6 or the like based on an instruction from the PU circuit 2.

The file circuit 6 includes, in addition to the RAM, an IC
A memory configured by a sound recording medium such as a memory or a floppy disk and storing an expanded sound signal of a voiced section according to the present invention, a signal subjected to a process of shortening a silent section, and the like. 5, when a processed audio signal is output, capture and store it,
Thereafter, the stored audio signal is supplied to the audio output circuit 7 based on an instruction from the CPU circuit 2.

The audio output circuit 7 is a circuit having a general configuration for outputting the audio signal in the file circuit 6 to the outside, for example, an interface circuit, a D / A (digital / analog) converter, a speaker, a recording device ( Or when the audio signal is output from the file circuit 6, while taking the audio signal and converting it into audio,
Output to the outside.

The CPU circuit 2 is a part constituted by a one-chip microcomputer or the like.
Based on the programs shown in FIGS. 5 and 6 stored in the OM circuit 3, control of the entire apparatus and various data processing are performed.

The PROM circuit 3 is a portion used as a storage location for programs that define the operation of the CPU circuit 2 and constant data used in various processes.
In accordance with a read command from the CPU circuit 2, the stored program or constant data is read and supplied to the CPU circuit 2.

Next, the operation of one embodiment of the present invention will be described with reference to FIGS.

FIGS. 5 and 6 are flow charts showing the flow of processing, and FIG. 6 shows details of the voiced section processing routine in ST9 of FIG.

Here, for the sake of explanation, the breathing section in the voice signal is referred to as “pause”, the section generated in one breath is referred to as “phrase”, and the average value of the time length of “phrase” is referred to as “predicted phrase length”. And is defined as follows.

Pause: Of the sections determined to be silent sections, the section length is Th1 (Th1 = 200 in this embodiment).
ms) or more silent section. In addition, Th means a threshold value.

Phrase: A section between a pose and the next pose. The start point of this section is Ph_st.

Predicted phrase length: The average time length of the phrase, which is T (unit: ms). (In this embodiment, T = 2
000 ms) f (t) and g (w) in FIG.
Is a function that determines the expansion ratio of a voiced section, and has the following characteristics.

F (t): a magnification function used to absorb the time expansion accompanying the speech speed conversion, wherein the magnification is set with respect to the appearance time t (0 ≦ t ≦ T) of the voiced section within the predicted phrase length. It is a monotonically decreasing function to be determined.

The predetermined magnification at t = 0 is r
_s, a predetermined magnification in t = T r _e (r _s
When ≧ r _e) to, f (t) satisfies _{r s ≧ f (t) ≧} r e, 0 ≦ t ≦ T.

G (w): A voiced section shorter than a fixed section length W ₁ (W ₁ = 150 ms in this embodiment) is scaled according to the section length w to a magnification higher than a standard magnification determined by f (t). Is a scaling function for expanding the voiced section length w
This is a monotonically decreasing function that determines the magnification for (0 <w ≦ W ₁ ).

Here, for a voiced section [t _k , t _k + W ₁ ] (w _k <W ₁ ) that satisfies the application condition of g (w), by definition of g (w), always f The relationship of (t _k ) ≦ g (w _k ) holds.

Next, the processing procedure of FIG. 5 will be described. In addition,
ST means a step.

(ST0) First, the maximum magnification r _s of f (t)
And to set the minimum magnification r _e.

(ST0-1) Next, the frame number i is set to 0
Set to.

(ST0-2) Subsequently, the above i is incremented to i + 1.

(ST1) Then, the input voice fetched by the voice input circuit 1 is divided into portions of a fixed length called frames, and the result is stored in the input buffer circuit 4.

In this embodiment, H of a frame width of 6.66 ms
Amming windows are cut out and stored while being shifted by 3.3 ms.

(ST2) The input voice signal is processed for each frame by a method such as the autocorrelation method or the zero-cross method to determine voiced, unvoiced, or silent. Input sounds other than voiced and unvoiced voices (for example, low-level noise and background sounds) uttered by humans are identified as silent in principle.

(ST3) The voiced / unvoiced / silent determination result (the current determination result) for the i-th frame and i-th frame
It is determined whether or not the voiced, unvoiced, and silence determination results (the previous determination results) for the first frame are the same.
If the two judgment results are the same, return to (ST0-2),
If they are not the same, move to the next (ST4). Where i = 1
In this case, the process returns to (ST0-2).

In this embodiment, in order to minimize the delay time of the processing of the entire system, voiced, unvoiced, and silent sections are not processed as a whole, but as long as possible. Divided into short sections (in this embodiment,
(The voiced section was divided into 150 ms.)

(ST4) Up to the (i-1) -th frame, speech sections determined to be of the same type (voiced, unvoiced or silent) are transferred from the input buffer circuit 4 to the processing buffer circuit 5 and stored.

(ST5) It is determined whether or not the voice section stored in the processing buffer circuit 5 is silent, unvoiced, or voiced. In the case of a silent section, the procedure proceeds to (ST6), in the case of a silent section, the procedure proceeds to (ST11), and in the case of a voice section, (ST9).
Move to

(ST6) It is determined whether or not the silent section is a pause section. In the case of the pause section, the process proceeds to (ST6-1), and in the case of not the pause section, the process proceeds to (ST8). However,
At the time of activation of the real-time speech speed conversion device in FIG.

(ST6-1) An initial value of 1 is substituted for a variable k representing the number of a voiced section appearing after the pause section.

(ST7) The section length of the pause is examined, and an algorithm preset according to the section length (Ryu Ikezawa et al., "A Method for Absorbing Time Expansion Associated with Speech Speed Conversion," 1992 Speech Research Group, 1992. P.49-P.56)
The sound is shortened to a degree that does not cause discomfort in hearing.

In the present embodiment, a silent section having a section length exceeding 862 ms is uniformly reduced to the value of 862 ms (Ryu Ikezawa et al., Spring Meeting of the Acoustical Society of Japan in 1992, "Time associated with speech speed conversion"). One Method for Absorbing Stretch "2-6-2 P.331-P.332 (1992-
3)) If a silent section continues after the silent section of 862 ms has elapsed, the remaining silent data is discarded and the start point of the next phrase is waited for.

(ST8) After the processed silence section signal in the processing buffer circuit 5 is transferred to the file circuit 6 and stored therein, the processing buffer circuit 5 is cleared. Next, the process proceeds to (ST12).

(ST12) It is determined whether the audio signal has been processed to the end. If the determination is affirmative, the process of this step (ST9) is terminated. If the determination is negative, (ST0-2).
Return to

(ST9) The section determined as a voiced section in (ST5) is subjected to a voiced section process shown in FIG. 6 described later. The origin of the time axis in the processing of this section is defined as V_st. In addition, the start time t _k of the k-th voiced in the phrase, the section length referred to as w _k.

(ST9-1) The variable k described above is incremented to k + 1.

(ST10) The speech speed converted voice data in the processing buffer circuit 5 is stored in the memory of the file circuit 6, and the processing buffer circuit 5 is cleared.
Thereafter, the process proceeds to (ST12).

(ST11) If it is determined in (ST5) that the section to be processed is unvoiced, the voice signal in this unvoiced section is transferred from the processing buffer circuit 5 to the file circuit 6 and stored therein. Clear Thereafter, the process proceeds to (ST12).

Next, the details of the ST9 voiced section processing routine of FIG. 6 will be described.

(ST14) First, the pitch of a voiced section is extracted.

(ST15) Next, it is determined whether or not the variable k is k = 1. In the case of k = 1, that is, in the case of the first voiced section appearing after the pause section, the process proceeds to (ST15-1).
If not, the process proceeds to (ST15-2).

[0069] (ST15-1) substitutes the time t ₁ to the variable V_st indicating the origin of the time axis in the process of the voiced segment. Next, the process proceeds to (ST16).

(ST15-2) Whether the variable k is 3 or less,
That is, it is determined whether or not k is 2 or 3. If k is 2 or 3, the process proceeds to (ST16), and if k is 4 or more, the process proceeds to (ST17).

(ST16) The maximum pitch frequency in the k-th voiced section is defined as P _k . P _{k for} k = 1, 2, 3
Save the value of.

(ST16-1) It is determined whether or not the variable k is k = 3. If k = 3, the process proceeds to the next (ST16-2). Otherwise, that is, if k = 1 and 2, (ST1-2).
Jump to 7).

(ST16-2) Three voiced sections P ₁ , P
The maximum value of _2, P ₃ and Pitch_max. Next, the process proceeds to (ST17).

(ST17) t _k is the interval [V_st, V
_St + T] is determined. If it is included, the operation moves to (ST17-1), otherwise (S17-1).
Move to T12). (In this embodiment, as described above, T = 20
00 ms. ) (ST17-1) V_st> determines whether the t _1.

When V_st> t ₁ , since the semantic importance is often low near the end point of the utterance, g (w) is not particularly applied in this embodiment, and (ST17-1) is directly applied to (ST17-1). ST19). Otherwise, it moves to the next (ST18).

[0076] (ST18) a time length required to effect the "slow-sensitive" on audibility caused by the conversion at the beginning of the phrase and T _1. T ₁ is the experimental results from (Imai, Atsushi other 1993 Acoustical Society of Japan Autumn Research Workshop "story-speed real-time absorption method of time stretching associated with the conversion." 1
9-10P. 361-P. 362 (1993-3)).
About 1/4 of T is desirable, and in this embodiment, T ₁ = 450.
ms.

In this processing block, it is determined whether or not the end time t _k + w _{k of} the k-th voiced section is included in the section [V_st, V_st + T ₁ ]. If it is included, the next (S
The process moves to T18-1), otherwise (ST19).

(ST18-1) k-th voiced section length w _k
Then, it is determined whether or not the preset section length W ₁ satisfies w _k ≦ W ₁ . If the determination is affirmative (ST2
0), and to the negative judgment (ST19).

[0079] In the speech speed conversion by the expansion of the voiced interval, but the conversion effect as those interval length is short becomes smaller, W ₁ is,
When the input voice is converted into speech speed at a uniform magnification of about 1.3 times, the critical voiced section length at which the speech speed conversion effect on the auditory perception is not felt much is derived by an experiment. ₁
= 150 ms.

(ST19) Magnification function f set in advance
(T) is applied to extend the voiced section. The f (t) is a monotonically decreasing function, in this example using a cosine function as the following equation (1) was changed magnification from r _s to r _e.

(See the graph of the curve in FIG. 7)

[0082]

[Number 1] _{f (t) = r e +0.5} (r s -r e) {cosπ (t-V_st) /T+1.0} (1) where, in V_st ≦ t ≦ V_st + T present embodiment, 1. _{0 ≦ r s ≦ 1.6,0.7 ≦ r} e <
The value was arbitrarily determined in the range of 1.0. Thereafter, the process returns to the main routine of FIG.

(ST20) Regarding the section length w _k of the voiced section, g (w
_The voiced section is expanded by applying the magnification determined in _k ).

The magnification function g (w) used in this embodiment is a linear function represented by the following equation (2), and the magnification is changed from g (0) to g (w).
₁ ) was changed. Thereafter, the process returns to the main routine of FIG.

(See the graph of the right-hand straight line in FIG. 7)

[0086]

[Number 2] g (w) = (- ( r s 2 -f (W 1)) w / W 1) + r s 2 (2) However, in _{V_st = 0 g (W 1)} = f (W 1) And

(ST21) If the maximum pitch frequency P _k of the voiced section to be processed satisfies the condition of the following equation (3), go to (ST22); otherwise, go to (ST22).
Move to 23).

[0088]

P _k > Pitch_max × Th2 (3) In this embodiment, Th2 = 0.7.

[0089] substituting the time t _k to (ST22) variable V_st.

(ST22-1) The variable r _{s is set} to (r _s −Th
Substitute 3).

[0091] As a result, f (t) is (r _s -Th
3) to r _e to change the magnification. In this embodiment, T
h3 was set to 0.1. Then, the above (ST17)
Return to

[0092] The extension magnification (ST23) voiced segment extending in the r _e. That is, the speech speed is kept at the highest speed.
Thereafter, the voiced section processing routine of (ST9) ends,
It returns to the main routine of FIG.

[0093]

As described above, according to the present invention,
In the method of separating the unvoiced section, unvoiced section, and voiced section of the input voice and slowly converting the utterance speed (speech speed) by expanding the voiced section, when all voiced sections are converted at a fixed magnification In order to eliminate the variation in speech speed conversion effect on auditory sensation caused by the difference in the length of each voiced section, a speech speed conversion effect on auditory sensation corresponding to a desired magnification is obtained for voiced sections shorter than a certain value. As described above, since the data is expanded at a higher magnification in accordance with the section length, a natural and stable speech speed conversion effect can be obtained for any uttered voice. That is, according to the present invention, it is possible to stably and naturally convert the speech speed to the one desired by the listener.

[Brief description of the drawings]

FIG. 1 is a graph showing a magnification function in a conventional method.

FIG. 2 is a timing chart showing a distribution of a voiced section length in one phrase on a time axis when a conventional method is applied.

FIG. 3 is a graph showing a magnification function according to an embodiment of the present invention.

Is a block diagram showing a circuit configuration example of a real-time speech speed conversion apparatus of one embodiment of the present invention; FIG.

FIG. 5 is a main flowchart showing an operation example of the real-time speech speed conversion device shown in FIG. 4;

FIG. 6 is a flowchart showing details of a voiced section processing routine shown in FIG. 5;

FIG. 7 is a diagram showing a configuration of the real-time speech speed converter shown in FIG.
It is a timing chart which shows the example of an operation at the time of applying the function of (t) and g (w).

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 Audio input circuit 2 CPU circuit 3 PROM circuit 4 Input buffer circuit 5 Processing buffer circuit 6 File circuit 7 Audio output circuit 8 Bus f (t) Magnification function g (w) used to absorb time expansion accompanying speech speed conversion the voiced segment less than a predetermined interval length W _1, determined in accordance with the section length w, f (t) the largest magnification r _e previously magnification function r _s predefined for stretching at higher standards magnification ratio determined by The minimum magnification T The predicted phrase length (the average time length of the phrase) Ph_st The starting point of the phrase (the section between the pause and the next pause) V_st The origin of the time axis in the processing of the voiced section _Pk The maximum of the k-th voiced section Pitch frequency pitch_max First three voiced sections P ₁ , P
_2, the maximum segment length value W ₁ is previously set w k _k-th voiced length i frame number k voiced numbers t _k start time of the k-voiced of P ₃

──────────────────────────────────────────────────続 き Continuing on the front page (72) Nobumasa Kiyoyama 1-10-11 Kinuta, Setagaya-ku, Tokyo Inside the Japan Broadcasting Corporation Broadcasting Research Institute (72) Eiichi Miyasaka 1-10-11 Kinuta, Setagaya-ku, Tokyo Japan Broadcasting Corporation Broadcasting Research Institute (56) References JP-A-1-93795 (JP, A) JP-A-5-257490 (JP, A) JP-A-5-80796 (JP, A) JP-A-4- 367898 (JP, A) JP-A-63-234299 (JP, A) JP-A-6-337696 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 21/04

Claims (4)

(57) [Claims] 1. A conversion that separates a silent section, a unvoiced section, and a voiced section of an input voice and expands a voiced section to reduce the utterance speed (speech speed) while maintaining the voice pitch. When performing, by sequentially detecting the time length of each voiced section, by multiplying the time length of each voiced section by a standard value of a uniform value, or smoothly changes with elapsed time,
A speech speed conversion device for obtaining an audible effect corresponding to the magnification, wherein a time length of a voiced section to be converted is equal to or less than a predetermined length.
Determination means for determining, by a determination result of said determining means, when said duration of the conversion subject to voiced interval Ru exceeds a predetermined length, multiplied by the reference magnification at the appearance time of the voiced but the time length of the voiced segments to be the conversion target, for said predetermined length following voiced interval, calculating means for multiplying a higher stretch ratio than that of the reference magnification according to the time length of the voiced To
A speech speed conversion device comprising: 2. The calculating means provides a higher expansion ratio for a short voiced section of 150 ms or less corresponding to the predetermined length, irrespective of the appearance time of the voiced section, compared to the reference magnification. 2. The method according to claim 1, further comprising: multiplying the expansion rate corresponding to the time length of the voiced section along a magnification function, and, in the case of the voiced section exceeding 150 ms, multiply the time length of the voiced section by the reference magnification. The speech speed conversion device according to 1 . 3. The method changes smoothly with the elapsed time.
When applying a magnification function characterized by setting a slow speech speed at the start point of this section in units of a section that occurs in a breath as a reference magnification, and gradually increasing the speech rate toward the end In the meantime , the calculating means calculates a short voiced section of 150 ms or less corresponding to the predetermined length that appears within a time 450 ms from the start time of this section,
Irrespective of the appearance time of the voiced section, multiplying by the expansion rate corresponding to the time length of the voiced section along a magnification function providing a higher expansion rate as compared with the reference magnification;
Voiced section exceeding 50 ms and elapsed time 450 ms
2. The speech speed conversion according to claim 1, wherein when the number exceeds the threshold , the time length of the voiced section is multiplied by the reference magnification.
Equipment . 4. The predetermined length refers to a maximum time length of a voiced section in which a “slow feeling” of the converted voice cannot be perceived audibly when a practical value is set as a reference magnification. For the following voiced sections, a new scaling function g (w) using the time length w as a variable is introduced,
The expansion magnification is given according to the magnification function, and the magnification given by the magnification function is a value higher than the reference magnification, and in particular, the shorter the voiced section, the higher the magnification. The maximum amplification value and the minimum amplification value of the magnification by the magnification function are not fixed, but each can be proportionally changed by the value of the reference magnification function f (t) of the reference magnification. The speech speed conversion device according to 1 .