JPS59152496A - Voice analysis synthesization system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] Technical Field of the Invention The present invention relates to a speech analysis and synthesis system, particularly a speech analysis and synthesis system that analyzes input speech and extracts parameters using a linear prediction method. When extracting the pitch period based on the above, multiple components with larger autocorrelation coefficients are extracted as candidates, and the pitch period for the more reliable period is used as the reference.

This invention relates to a speech analysis and synthesis system that sets pitch periods for periods with very low reliability.

(■3) Technical background and problems Conventionally, in speech analysis and synthesis systems using linear prediction methods (including the -@ΦPARCOR method), the prediction residual as a result of linear prediction is examined, and the self-analysis of the prediction residual is performed. The pitch period is extracted by focusing on the fact that the change in the correlation coefficient sequence with respect to time delay takes a peak at the time delay position corresponding to the pitch period. That is, in extracting the pitch period, (1) find the time delay that gives the maximum value of the autocorrelation coefficient sequence of the prediction residual, examine the fluctuations from the pitch period for frames before and after l fll +, and (ii
If the il fluctuation is large, check whether there is a peak at a position corresponding to a half period or double period where the fluctuation becomes small.

(1v), a method is adopted in which the time delay giving the peak is taken as the pitch period.

However, in the case of this method, if there are fluctuations in the big period of the original speech or if the pitch structure collapses at phoneme boundaries, the autocorrelation coefficient of the prediction residual will not have a sharp peak, and the pitch This includes the problem that the period cannot be discovered correctly.

(C1 Objective and structure of the invention The present invention aims to solve the above points,
Reliably extracting the pitch period K.9
Eliminate manual error correction work.

The aim is to fully automate speech analysis and synthesis. Then, the speech analysis and synthesis system of the present invention analyzes the input speech using a linear prediction method and converts it into parameters, and also extracts the pitch period using the autocorrelation coefficient of the pre-i1+lI residual.

In the speech analysis and synthesis system that performs speech synthesis based on these results, the autocorrelation coefficient sequence of the above prediction residuals or the weighted moving average of the autocorrelation coefficients of the above prediction residuals is calculated. A pitch period determination candidate extraction unit that examines a relational number sequence and extracts a plurality of candidates with larger correlation coefficients, a time delay component extraction unit that extracts a time delay corresponding to each pitch period determination candidate, and the pitch period determination candidate. a logical judgment processing section that determines the pitch period based on the pitch period determination candidate extracted by the extraction section and the corresponding time delay component, and the logical judgment processing section determines the pitch period based on the value of the pitch period determination candidate. Check the reliability with which the pitch period can be set by the corresponding time delay component, and use the time delay component corresponding to the pitch period determination candidate with confidence M greater than the threshold to determine the period with reliability greater than the threshold. The present invention is characterized in that the pitch period is determined, and the big period for a period having reliability below a threshold value is determined so as to maintain continuity with the previously determined big period. The following describes the invention with reference to the drawings. A block diagram of an embodiment of the speech analysis and synthesis system, Figures 3 to 7, respectively.
An explanatory diagram illustrating the processing mode in the illustrated logical judgment processing section, FIG. 8 is a flowchart of an embodiment of the processing related to the first bus in the #i logical judgment processing section, and FIG. 9 is a second path in the logical judgment processing section. Flowchart 1 of an embodiment regarding processing related to FIG.
FIG. 12 shows a flowchart of an embodiment of the processing related to the fourth bus in the logic decision processing section.

In extracting various parameters for speech synthesis, as shown in Figure 1, the input speech is cut out at a predetermined time length (frame length), and the 7 adjacent frames are cut out so that they partially overlap. A frame is created, a frame number is assigned to each frame, and parameters and tea extraction are performed for each frame.

The input audio extracted in this way is (X(n) −) −(x(01+X(1) +
``'+x(%-+)), the input and processing as shown in FIG. 2 is performed.

In Fig. 2, 1 is an incoming audio correlation processing section, 2 is a linear prediction processing section j, a 3-digit residual power extraction section, 4 is a residual correlation processing section, 5 is a moving average processing section, and 6 is a maximum value. Extraction processing unit, 7i
-t pitch period determination candidate holding section; 8 is a time delay component holding section; 9 is a logical judgment processing section.

10 voiced/unvoiced determination units, 11 a voiced section 1 drive sound source generation section, 121, unvoiced (including silent) section drive sound υ12 signal generation generation 15 a switching processing section, and 14 a linear prediction synthesis section. ing.

The extracted input sound (X(n)) as described in Section 1 is subjected to autocorrelation (rl) with respect to the input sound in the processing section I. Then, the parameters related to the vocal tract are extracted in the linear prediction processing unit 2, and the prediction residual (e(n)) is sent to the residual power extraction unit, the residual correlation processing unit 4, and the voiced/unvoiced determination unit 1.
It can be supplied to 0.

Illustrated units 3, 4, 5, 6, 7, and 8.9 correspond to nine pitch period extraction sections when the present invention is applied. The pitch period extracted by the pitch period extracting section is supplied to the voiced section drive sound source generating section 11 at the time of 9 synthesis. Then, a vibration component corresponding to the pitch period and the electric power from the extractor 3 is generated in the form of a pulse or a triangular wave.

Further, the unvoiced section driving sound source generating section 12 generates power corresponding to the unvoiced section in the form of white noise.

The voiced/unvoiced determining section 10 determines whether a voiced section is a voiced section or an unvoiced section, and performs a switching process in a switching processing section 13 at the time of 9 synthesis. Then, the linear prediction synthesis unit 14 includes the switching processing unit 13
Parameters related to the vocal tract are added to the output from the voice tract, and the result is output as 2-synthesized speech (y(-)).

An important feature of the invention is that the illustrated unit 3°4.5 in FIG.
, 6, 7, and 8.9, and will be specifically explained below.

When the prediction residual (e (b)) is extracted as shown in FIG. 2, the residual power extraction section 3 extracts the power of the residual (e (wa)). On the other hand, the residual correlation processing section 4 calculates autocorrelation for the residuals. That is, let the prediction residual be the autocorrelation coefficient for the time delay i of the prediction residual of the m-th frame.

(However, i=0.1...mark...N-1>-+11
A series of i autocorrelation coefficients ρ is extracted when

When each coefficient ρl(f) is viewed in time series, there is a possibility that extreme unevenness exists undesirably, and the moving average processing unit 5 has a value of 9, for example.

ρ: (e) - ku (ρ1 - mountain + ρ1 (lI +) ten ρ,.□
(→) However, I e ['l' min, 2
T, max)---------------(21
Take a moving average in the form of: Note that the above Tm
1n and 2Tmax will be described later. Second
In the configuration shown in the figure, the autocorrelation coefficient ρ'+h) obtained by taking the above-mentioned moving average is used, but in the following explanation, for the sake of simplicity, we will draw the autocorrelation coefficient ρ', (→ with rp as the coefficient ρ
1 (described as →.

The maximum i1^ extraction processing unit 6 shown in FIG. In addition, the time delay 1 corresponding to each selected coefficient JC is stored in the time delay component holding unit 8. That is, the time delay 1 corresponding to each selected coefficient JC is stored in the time delay component holding unit 8. That is, the time delay 1 that is considered to have a pitch period in an unpleasant manner is stored. The range of i is [T surface n.

Tmax), in order to deal with the case where the pitch in the original voice is dropped, the search range of the time delay i in the above equation 11) is set as [Tm1n, 2 Tmax
), but within the range of this time delay i, five candidates are selected in descending order of the value of the autocorrelation coefficient of the prediction residual. Then, a time delay corresponding to this is selected. In general, this selection can be expressed as a formula.

ρ (Ll (c) = max ρ, (c)
□(3)! 6 (Tm1n, 2Tmax)iV
τ(k)(→, 1k<j-1 ward J-5 τ”(m) −nrg Ill; IX
ρ+ (@ −−−−−−−−−−f4)iE
('J”min, 2'l”+nax 'J1'\τ
(J → i %, k (j −11 × j × 5). Note that this is a function that takes the time delay 1 that gives the above arg ρ (1) as a value.

As above.

(ρ(')(,,+,ρ(')(,0,...)
ρ3゛) (E)) (J”) (→ , (2) (→,...
...Js)(→) is selected, and each of the above ρ(I)(c)
is regarded as the pitch period with the corresponding time delay Jl)(e), and it is said that it corresponds to the reliability of the case.

The logic judgment processing unit 9 shown in FIG.
Based on 〈I)(→ and τ(I)(→), each Freno,
Determine the pitch family Jul for each pitch group. The processing in the processing section 9 is as shown in FIG.
7. First pass, second pass as schematically shown in [7].

It consists of each process of the third pass and the fourth bus. This will be explained below.

Assume that the current input audio is as shown in FIG. 3 fA1 when the frame number is plotted on the horizontal axis and the above-mentioned first reliability value is plotted on the vertical axis.

Assume that the input audio is arranged as a silent section, a voiceless section, a voiced section, a silent section, a voiced section, and a silent section as shown in FIG. 5 CB+.

In the logic judgment processing section 9, in the first pass, the reliability ρ''- is equal to the threshold θ as shown in FIG. 3 FC+.

For each frame in an interval exceeding , i.e.

The pitch family M is tentatively determined for each frame that satisfies the following. Such decisions are made in chronological order as indicated by the arrows shown in Figure 9, and then also in the opposite direction of the arrows. FIG. 6 (C1 represents processing corresponding to the first pass).

In other words, considering the processing that corresponds to the order of the nine time series, the pitch period corresponding to the first frame is τ (c) = , (+) (→□−−□−−, −−−−−
+6), and for subsequent frames, the condition is that the pitch period deviation is within a predetermined 7r width θ (referred to as j+1-d) l ''(m) T'(m-+) l : 5-θ,
−−−−−−−−−−−−−−−−−−−−−−−t
7), set the pitch period to (T”'(→,..., JI)(→,...
J5) Select from (→) in descending order of J. If there is no frame that satisfies Equation (7), the pitch period is not determined for the frame and the process proceeds to the ninth frame. The next frame determines the pitch period according to equation (6). To process in the opposite direction to the time series, consider τ8(→ instead of τF(steep).

This is done by converting the time as m' = M - m + 1. And τr (ha) thorn (
For frames in which both ``f'' and ``f'' are determined, the pitch period τ(c) is used.

τ(→w min (τ′(kaτ8(c))−1−−−
−, −, −−−−−−−−−+81, and for frames for which only one is determined, the pitch period for which only one is determined is defined as pitch period τ(→).

In other cases, the pitch period τ(→ is not determined for the Freno).

In the second pass, the pitch period of each frame determined in the first pass is checked, and if there is a pitch period undetermined portion in a frame of 3 frames or less as shown in FIG. The pitch period of point Q ((rn+n+1)th frame) is 1r (nrl-n+1) with respect to the pitch period of P (F mth frame)
r(ml l≦−v'(n −1−1−) ”θr
-------+91 is selected from among the above five candidates. If there is no corresponding item for the five items, (+)(→/2 f7j
(corresponding to half of the pitch period of the candidate whose txt was previously omitted) will also be investigated.

” 1. Processing also applies to 1. time series and reverse direction.

Convert and process the time as m' = M - m + 1 to find the smaller value as the pitch period of the frame at point P and point Q. 5 and 5 are explanatory diagrams showing the processing for the first pass and the second pass, respectively, in the form of a schematic flowchart. In the flowchart shown in FIG. 4 corresponding to the first pass, the frame is a voiced frame and has a high reliability.

Let the pitch period of the first frame be the first candidate.

For subsequent frames, candidates that maintain continuity are used as the pitch period of the frame. In the flowchart shown in FIG. 5 corresponding to the second pass, frames that are reliable and have a distance of 3 frames or less are processed so as to maintain continuity.

As shown in FIG. 3 (1), the third bus is used to respond to frames that are voice frames and whose pitch period has not yet been determined, using a frame that has been determined to be false as a core, or a frame 9 hours earlier or Later, the pitch period for each frame is determined by selecting from among the five candidates selected with low reliability so that the pitch period is continuous as shown by the arrows in the figure.

FIG. 6 #'i is an explanatory diagram showing the processing for the third pass in the form of a general flowchart. Find the beginning and end of a group of frames with low reliability.

The pitch period is determined for each frame within the frame group while maintaining continuity. 11) V
The second candy is also investigated as a candidate.

The fourth bus checks pitch continuity for the entire voiced frame, and if it is discontinuous.

For example, linear interpolation is performed from the pitch period of the previous and next frames. FIG. 7 shows a schematic 70-chart in that case.

Needless to say, in the above process, the constant Tm1n.

Tmax, θ1. θ1 etc. are determined in preliminary experiments.

Figure 8 shows details of the processing of the first panel shown in Figure 4. Reference numeral 15 in the figure corresponds to processing related to voiced frames, 16 corresponds to processing related to frames with high reliability, 17 corresponds to processing when the previous frame is silent/unvoiced, and 18 corresponds to processing related to the pitch period. 19 corresponds to processing to maintain continuity of K'iJ, and 19 corresponds to processing when the previous frame has low reliability.

FIG. 9 shows details of the second pass processing ff1K shown in FIG. The reference numeral 20 in the figure corresponds to the process when the separation is within 6 frames, and corresponds to the process for checking extended continuity as shown in the 21st digit equation (9), and 22 corresponds to the process for checking the extended continuity as shown in the 21st digit equation (9). It also supports processing to correct the pitch period to maintain continuity.

FIG. 10 (Al (B+ indicates details of the processing of the sixth pass shown in FIG. 6. Reference numeral 26 in the figure
corresponds to the process of detecting a switch from a silence/voice frame to a frame with low reliability, and 24 corresponds to the process of detecting a switch from a frame with high 8-axis degree to a frame with low reliability. Ori.

Both processes 23 and 24 detect the beginning of a frame with low reliability. 25 corresponds to a process of detecting a switch from a frame with low reliability to a silent/unvoiced frame. Further, 26 corresponds to the process corresponding to process A shown in FIG. 27 corresponds to the process corresponding to process B shown in FIG. Furthermore, the second day corresponds to a process for detecting a switch from a frame with low reliability to a frame with high reliability, and 29 corresponds to process C shown in FIG.

30 corresponds to the processing shown in FIG.

FIG. 11 shows the processing steps 26, 27, 29°3 shown in FIG. 10.
Details of the sub-pass that performs each process in 0JC are shown. 31 in the figure corresponds to the process of selecting from among the candidates, 32 corresponds to the process of extracting the above-mentioned τ(l and 2), 63 corresponds to the process of S@ whose previous frame was silent/voiceless, 64 corresponds to processing when the previous frame was voiced, 35 corresponds to processing to search for continuity,
36 corresponds to the process of checking whether τ(172 maintains continuity).

FIG. 12 shows a summary 1 of the processing of the fourth bus shown in FIG.
Corresponds to processing when there is no voice.

39 is silent/voiceless for the next 7 frames. 40 corresponds to processing when the next frame is voiced, 41 corresponds to processing when the previous frame is voiced, and 42 corresponds to processing when the next frame is silent/unvoiced. 43 corresponds to processing when the subsequent frame is also voiced (interpolation).

(El As described in detail, according to the present invention, when determining the pitch period, even if the pitch fluctuates or drops, the pitch period can be rationally selected from among a plurality of candidates. You can choose and decide.

It becomes possible to eliminate processes such as relying on humans to supplement pitch period determination, and it becomes possible to achieve full automation of voice analysis and synthesis.

Note that in the above 812 description, one average of the autocorrelation coefficient sequence of the prediction residual was considered, but the present invention is not limited to this, and other weighted equals may be considered.

4. Figure 1 of the drawings is a diagram illustrating the concept of the frame referred to in the tree specification ¥4, and Figure 2 is a block diagram of an embodiment of the speech analysis and synthesis system of the present invention. , FIGS. 3 to 7 are explanatory diagrams each explaining the processing mode that goes to the logical judgment processing unit shown in FIG. 2, and FIG. Example flowchart, FIG. 9 is a flowchart of an embodiment of processing related to the second pass in the logic judgment processing section, and FIG. FIG. 11 is a flowchart of an embodiment of processing related to the third path in the logic judgment processing section, and FIG. 12 is a flowchart of one embodiment of processing related to the fourth bus in the logic judgment processing section.

In the figure, 1 is an input audio correlation processing section, 2 is a linear prediction processing section,
3-residual power extraction unit; 4, residual correlation processing unit; 5, moving average processing unit; and 6, maximum value extraction processing unit.

7 is a pitch period determination candidate holding section, 8 is a time delay component holding section, 9 is a logic judgment processing section, 10 is a voiced/unvoiced judgment section, 1
1 is a voiced section drive sound source generation unit.

Reference numeral 12 denotes an unvoiced (including silence) section drive sound source generation unit.

Reference numeral 16 represents a switching processing section, and reference numeral 14 represents a rock-shaped predictive synthesis section.

Watch applicant: Fujitsu Limited

Claims (1)

[Claims] A voice in which an input voice is analyzed by a linear prediction method and converted into parameters, a pitch period is extracted using an autocorrelation coefficient of a prediction residual, and voice synthesis is performed based on these results. In analysis and synthesis, the autocorrelation coefficient sequence of the above prediction residuals or the autocorrelation coefficient sequence of the weighted moving average of the autocorrelation coefficients of the prediction residuals is examined to find multiple ones with larger correlation values. a pitch period determination candidate extraction unit that extracts candidates, a time delay component extraction unit that extracts a time delay corresponding to each pitch period determination candidate, and a pitch period determination candidate extracted by the pitch period determination candidate extraction unit described in 2. The logical judgment processing section determines the pitch period based on the corresponding time delay component, and the logical judgment processing section determines the pitch period based on the corresponding time delay component based on the value of the pitch period determination candidate. The method uses time delay components corresponding to pitch period determination candidates with reliability higher than a threshold value. Determine the pitch period for a period with reliability above the threshold, and determine the pitch period for a period with reliability below the threshold to maintain continuity with the previously determined pitch period. A speech analysis and synthesis system characterized by the following.