CN105023574B - A kind of method and system for realizing synthesis speech enhan-cement - Google Patents
A kind of method and system for realizing synthesis speech enhan-cement Download PDFInfo
- Publication number
- CN105023574B CN105023574B CN201410182886.6A CN201410182886A CN105023574B CN 105023574 B CN105023574 B CN 105023574B CN 201410182886 A CN201410182886 A CN 201410182886A CN 105023574 B CN105023574 B CN 105023574B
- Authority
- CN
- China
- Prior art keywords
- model
- parameter
- enhancing
- synthesis
- speech
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Abstract
The present invention relates to speech synthesis technique fields, disclose a kind of method and system for realizing synthesis speech enhan-cement, and this method includes：Initial speech synthetic model is built based on training data, the training data includes text data and voice data corresponding with the text data；Enhancing model is established, the enhancing model is used to simulate the synthesis speech parameter and the mapping relations of natural-sounding parameter of the initial speech synthetic model generation；After text to be synthesized is received, according to the initial speech synthetic model and the synthesis speech parameter of the corresponding text to be synthesized of enhancing model generation；Utilize the synthesis speech parameter generation continuous speech signal.Using the present invention, the enhancing effect of synthesis voice can be effectively improved.
Description
Technical field
The present invention relates to speech synthesis technique fields, and in particular to a kind of method and system for realizing synthesis speech enhan-cement.
Background technology
Realize it is man-machine between hommization, intelligentized effective interaction, build man-machine communication's environment of efficient natural, into
The active demand applied and developed for current information technology.As an important technology practical in voice technology, phonetic synthesis
Text information is converted into natural voice signal by technology or literary periodicals technology (Text-To-Speech, TTS), is realized
The real-time conversion of text, changes the troublesome operation that tradition is lifted up one's voice by playback realization machine of recording, and the system of saving is deposited
Space is stored up, in increasing current of information exchange, particularly needs the dynamic queries often changed application neck in the information content
Domain has played increasingly important role.
Based on the speech synthesis system of parameter synthesis due to having preferable robustness and generalization to obtain widely should
With, however this method has stronger smoothing effect, the voice of synthesis is flat and sound quality is easily damaged, in terms of naturalness is synthesized
Show not ideal enough, there are certain rooms for promotion in practical application.How to improve the naturalness of synthesis voice is synthesis system
Practical important leverage.
For this purpose, the naturalness of synthesis voice, master are mainly improved using the method for synthesis speech enhan-cement in the prior art
Want technology that can be summarized as：The Heuristicses such as the sense of hearing characteristic based on people carry out post filtering to generation frequency spectrum parameter or synthesis voice
Processing, such as the formant of pairing into voice carry out enhancing processing, the dynamic characteristic for strengthening generation frequency spectrum parameter, so as to improve conjunction
Into the sound quality of voice.
In fact, the acoustic characteristic of different speakers is sending out difference there are detail differences, and for same speaker
During sound, there is also detail differences for acoustic characteristic.And the synthesis sound enhancement method based on Heuristics, it can only cause enhanced
Synthesis voice meets the sense of hearing of people on the whole, and enhancing effect is unsatisfactory.
Invention content
The embodiment of the present invention provides a kind of method and system for realizing synthesis speech enhan-cement, to improve the enhancing of synthesis voice
Effect.
For this purpose, the embodiment of the present invention provides following technical solution：
A kind of method for realizing synthesis speech enhan-cement, including：
Based on training data build initial speech synthetic model, the training data include text data and with the text
The corresponding voice data of data；The initial speech synthetic model includes：Duration modeling, spectral model, fundamental frequency model；
Enhancing model is established, the enhancing model is used to simulate the synthesis voice ginseng of the initial speech synthetic model generation
Number and the mapping relations of natural-sounding parameter；
After text to be synthesized is received, according to the initial speech synthetic model and the corresponding institute of enhancing model generation
The synthesis speech parameter of text to be synthesized is stated, including：According to the enhancing model to the frequency in the initial speech synthetic model
Spectrum model and/or fundamental frequency model carry out enhancing processing, the spectral model and/or fundamental frequency model enhanced；Utilize the enhancing
Spectral model and/or the corresponding text to be synthesized of fundamental frequency model generation frequency spectrum parameter and/or base frequency parameters；Using described
Other voices in addition to spectral model and/or fundamental frequency model of the corresponding text to be synthesized of initial speech synthetic model generation
Parameter；
Utilize the synthesis speech parameter generation continuous speech signal.
Preferably, the enhancing model of establishing includes：
The synthesis speech parameter of all training datas is generated according to the initial speech synthetic model；
Extract the natural-sounding parameter of all training datas；
Determine the topological structure of enhancing model；
Using the data of the synthesis speech parameter of the correspondence training data and natural-sounding parameter to gathering as training, root
Parameter training is carried out according to the topological structure, obtains enhancing model.
Preferably, the enhancing model is：The mapping model or GMM model of linear function or DNN models.
Preferably, the synthesis speech parameter and the mapping relations of natural-sounding parameter of the initial speech synthetic model generation
It is distributed for the synthesis speech parameter of initial speech synthetic model generation and the condition of natural-sounding parameter.
Preferably, it is described according to it is described enhancing model to the spectral model and/or base in the initial speech synthetic model
Frequency model carries out enhancing processing, and the spectral model and/or fundamental frequency model enhanced includes：
Spectral model and/or the model parameter of fundamental frequency model are obtained from the initial speech synthetic model；
Enhancing processing is carried out to the model parameter using the enhancing model, obtains enhanced model parameter；
Enhanced model parameter is substituted into corresponding spectral model and/or the model parameter of fundamental frequency model, is enhanced
Spectral model and/or fundamental frequency model.
A kind of system for realizing synthesis speech enhan-cement, including：
Initial model establishes module, for being based on training data structure initial speech synthetic model, the training data packet
Include text data and voice data corresponding with the text data；The initial speech synthetic model includes：Duration modeling, frequency
Spectrum model, fundamental frequency model；
Enhance model building module, enhance model for establishing, the enhancing model closes for simulating the initial speech
The synthesis speech parameter and the mapping relations of natural-sounding parameter generated into model；
Receiving module, for receiving text to be synthesized；
Parameter generation module, for being treated according to the initial speech synthetic model and the enhancing model generation are corresponding
The synthesis speech parameter of synthesis text；The parameter generation module includes：Model enhancement unit, for according to the enhancing model
Enhancing processing is carried out to the spectral model in the initial speech synthetic model and/or fundamental frequency model, the frequency spectrum mould enhanced
Type and/or fundamental frequency model；Enhance speech parameter generation unit, for utilizing the spectral model and/or fundamental frequency model of the enhancing
The frequency spectrum parameter and/or base frequency parameters of the corresponding text to be synthesized of generation；Initial speech parameter generating unit, for utilizing
State other languages in addition to spectral model and/or fundamental frequency model of the corresponding text to be synthesized of initial speech synthetic model generation
Sound parameter；
Synthesis module, for utilizing the synthesis speech parameter generation continuous speech signal.
Preferably, the enhancing model building module includes：
Speech parameter generation unit is synthesized, for generating the conjunction of all training datas according to the initial speech synthetic model
Into speech parameter；
Natural-sounding parameter extraction unit, for extracting the natural-sounding parameter of all training datas；
Topological structure determination unit, for determining the topological structure of enhancing model；
Training unit, for the data of the synthesis speech parameter of the training data and natural-sounding parameter will to be corresponded to making
Gather for training, parameter training is carried out according to the topological structure, obtain enhancing model.
Preferably, the model enhancement unit includes：
Model parameter acquiring unit, for obtaining spectral model and/or fundamental frequency mould from the initial speech synthetic model
The model parameter of type；
Model parameter enhancement unit for carrying out enhancing processing to the model parameter using the enhancing model, obtains
Enhanced model parameter；
Enhance model generation unit, for enhanced model parameter to be substituted corresponding spectral model and/or fundamental frequency mould
The model parameter of type, the spectral model enhanced and/or fundamental frequency model.
The method and system provided in an embodiment of the present invention for realizing synthesis speech enhan-cement, Statistics-Based Method structure are used for
The synthesis speech parameter of traditional voice synthetic model generation and the enhancing model of the mapping relations of natural-sounding parameter are simulated, then
Using the enhancing model and the synthesis speech parameter of the corresponding text to be synthesized of traditional voice synthetic model generation, and then described in utilization
Synthesize speech parameter generation continuous speech signal.The scheme of the embodiment of the present invention, therefore can using natural parameters,acoustic as guidance
The minutia of parameters,acoustic has stronger assurance during with to different speakers and same pronunciation human hair difference sound, can grab
The firmly characteristic of specific speaker, the effect for making synthesis speech enhan-cement are more preferable.
Description of the drawings
It in order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, below will be to institute in embodiment
Attached drawing to be used is needed to be briefly described, it should be apparent that, the accompanying drawings in the following description is only one described in the present invention
A little embodiments for those of ordinary skill in the art, can also be obtained according to these attached drawings other attached drawings.
Fig. 1 is the flow chart for the method that the embodiment of the present invention realizes synthesis speech enhan-cement；
Fig. 2 is according to a kind of flow of initial speech synthetic model and enhancing model synthesis voice in the embodiment of the present invention
Figure；
Fig. 3 is according to another flow of initial speech synthetic model and enhancing model synthesis voice in the embodiment of the present invention
Figure；
Fig. 4 is the structure diagram for the system that the embodiment of the present invention realizes synthesis speech enhan-cement；
Fig. 5 is a kind of specific implementation structure diagram of parameter generation module in the embodiment of the present invention；
Fig. 6 is another specific implementation structure diagram of parameter generation module in the embodiment of the present invention.
Specific embodiment
In order to which those skilled in the art is made to more fully understand the scheme of the embodiment of the present invention, below in conjunction with the accompanying drawings and implement
Mode is described in further detail the embodiment of the present invention.
Since the acoustic characteristic of different speakers is there are detail differences, and for same speaker, sending out not unisonance
When, there is also detail differences for acoustic characteristic.And the Heuristicses such as the existing synthesis sense of hearing characteristic of sound enhancement method based on people
Post filtering processing is carried out to generation frequency spectrum parameter or synthesis voice, is not concerned with the details characteristic of speaker parameters,acoustic, it can only
So that enhanced synthesis voice meets the sense of hearing of people on the whole, it is impossible to obtain ideal enhancing effect.For this purpose, the present invention is real
It applies example in view of the problems of the existing technology, a kind of method and system for realizing synthesis speech enhan-cement, the side based on statistics is provided
Method builds to simulate the synthesis speech parameter of traditional voice synthetic model generation and the increasing of the mapping relations of natural-sounding parameter
Then strong model corresponds to the synthesis speech parameter of text to be synthesized using the enhancing model and the generation of traditional voice synthetic model,
And then utilize the synthesis speech parameter generation continuous speech signal.
As shown in Figure 1, being the flow chart for the method that the embodiment of the present invention realizes synthesis speech enhan-cement, include the following steps：
Step 101, based on training data build initial speech synthetic model, the training data include text data and with
The corresponding voice data of the text data.
The initial speech synthetic model can be built using traditional parameter synthesis method, including：It is each basic
The corresponding binary decision tree of synthesis unit, spectral model, fundamental frequency model, duration modeling etc..For example, it may be used based on HMM's
For spectral model, leaf is simulated using GMM (Gaussian Mixture Mode, gauss hybrid models) for parameter synthesis method
The spectrum distribution of node, Gaussage usually can be determined as positive integer with reference to training data scale, for example, select Gaussage for
1。
Step 102, enhancing model is established, the enhancing model is used to simulate the conjunction of the initial speech synthetic model generation
Into speech parameter and the mapping relations of natural-sounding parameter.
Since the setting of enhancing model has important influence into voice enhancing effect with optimization pairing, in this hair
In bright embodiment, using the enhancing model set-up mode based on data-driven, using natural parameters,acoustic as guidance, real embodiment
The minutia of parameters,acoustic when different speakers and same pronunciation human hair difference sound, and then improve synthesis speech enhan-cement
Effect.
The building process for enhancing model is as follows：
(1) the synthesis speech parameter of all training datas is generated according to initial speech synthetic model；
(2) the natural-sounding parameter of all training datas is extracted；
(3) topological structure of enhancing model is determined；
(4) using the data of the synthesis speech parameter of the correspondence training data and natural-sounding parameter to as training set
It closes, parameter training is carried out according to the topological structure, obtain enhancing model.
It should be noted that in practical applications, the enhancing for spectral characteristic and/or fundamental frequency characteristic can be built respectively
Model.For example, for the enhancing model of spectral characteristic, specific building process is as follows：
(1) spectral model in initial speech synthetic model generates the synthesis frequency spectrum parameter of all training datas.
According to the spectral model and force alignment result, it may be determined that the corresponding spectral model sequence of training data.Tool
Body, for single basic voice unit, according to alignment duration information is forced repeatedly to be copied selected spectral model, obtain
Take the spectrum signature series model of the basic voice unit.
The likelihood score summation of the corresponding spectral model sequence of training data is counted, is calculated as follows：
Wherein W be calculate dynamic parameter window function matrix, CsFor frequency spectrum parameter to be generated, MsAnd UsRespectively frequency spectrum
The mean value and covariance matrix of model.The likelihood score summation of obvious spectral model is the function of target spectrum characteristic vector.
(2) the natural frequency spectrum parameter of all training datas is extracted.
(3) topological structure of spectral enhancement model is determined.
Spectral enhancement model is used to simulate the frequency spectrum parameter of traditional voice synthetic model generation and reflecting for natural frequency spectrum parameter
Relationship is penetrated, in embodiments of the present invention, the mapping model of linear function may be used, GMM model or DNN models can also be used
Etc. mathematical statistical models.In general, the model the fine then better in its sufficient simulation effect of data.
(4) parameter training is carried out to spectral enhancement model according to the topological structure, obtains the spectral enhancement model of optimization,
Establish synthesis frequency spectrum parameter xtWith natural frequency spectrum parameter ytCondition distribution p (yt|xt)。
It is similar to the above for the building process of the enhancing model of fundamental frequency characteristic, it is not described in detail herein.
Step 103, after text to be synthesized is received, according to the initial speech synthetic model and the enhancing model life
Into the synthesis speech parameter of the correspondence text to be synthesized.
Based on above-mentioned enhancing model, in practical applications, various ways may be used to initial speech synthetic model or conjunction
Enhanced into speech parameter, can obtain good enhancing effect, specific implementation process will be described in detail later.
Step 104, the synthesis speech parameter generation continuous speech signal is utilized.
The method provided in an embodiment of the present invention for realizing synthesis speech enhan-cement, Statistics-Based Method structure pass for simulating
The synthesis speech parameter and the enhancing model of the mapping relations of natural-sounding parameter of phonetic synthesis of uniting model generation, then utilizing should
Enhance the synthesis speech parameter of model and the corresponding text to be synthesized of traditional voice synthetic model generation, and then utilize the synthesis language
Sound parameter generates continuous speech signal.Due to it is described enhancing model be using natural parameters,acoustic as instruct, can be to not
The minutia of parameters,acoustic has stronger assurance when same speaker and same pronunciation human hair difference sound, can catch specific
The characteristic of speaker, the effect for making synthesis speech enhan-cement are more preferable.Moreover, the scheme of the embodiment of the present invention is in practical synthesis task
Operand will not be increased, be conducive to the real time implementation of product.
It should be noted that in practical applications, according to initial speech synthetic model and enhancing model generation synthesis voice
There are many modes of parameter.For example, can utilize corresponding enhancing model to the spectral model in initial speech synthetic model and/
Or fundamental frequency model carries out enhancing processing, utilizes enhancing treated spectral model and/or the corresponding text to be synthesized of fundamental frequency model generation
This frequency spectrum parameter and/or base frequency parameters generates other phonetic synthesis parameters by initial speech synthetic model, then utilizes these
Phonetic synthesis parameter generates continuous speech signal.For another example, correspondence can also be generated first with initial speech synthetic model to wait to close
Into the phonetic synthesis parameter (including duration parameters, frequency spectrum parameter, base frequency parameters) of text, corresponding enhancing model is then recycled
Enhancing processing is carried out (including frequency spectrum parameter and/or base frequency parameters) to some of which phonetic synthesis parameter, finally utilizes these
Phonetic synthesis parameter (mainly duration parameters) generation that enhanced phonetic synthesis parameter and other do not reinforce processing connects
Continuous voice signal.
Separately below according to initial speech synthetic model and enhancing model generation in citing the present invention will be described in detail embodiment
Synthesize the process of speech parameter.
As shown in Fig. 2, it is according to initial speech synthetic model and enhancing model generation synthesis voice in the embodiment of the present invention
A kind of flow chart of parameter, includes the following steps：
Step 201, the duration parameters and base frequency parameters of the corresponding text to be synthesized of initial speech synthetic model generation are utilized.
Step 202, enhancing processing is carried out to the spectral model in initial speech synthetic model according to enhancing model, is increased
Strong spectral model.
First, model parameter is obtained from initial spectral model, such as the spectral model based on GMM, it is denoted as xt；So
Afterwards using advance trained enhancing model, to model parameter xtEnhancing processing is carried out, i.e., according to p (yt|xt), after acquiring enhancing
Model parameter yt；Finally with enhanced model parameter ytThe model parameter of spectral model is replaced, obtains new spectral model,
This model is enhanced spectral model.
Step 203, the frequency spectrum parameter of the corresponding text to be synthesized of the spectral model of enhancing generation is utilized.
Step 204, it is generated and connected using the duration parameters, base frequency parameters and frequency spectrum parameter of the correspondence text to be synthesized
Continuous voice signal.
It should be noted that it in practical applications, can generate respectively for the enhancing model of spectral characteristic and for base
Therefore the enhancing model of frequency characteristic, can individually use the enhancing model for spectral characteristic in initial speech synthetic model
Spectral model carry out enhancing processing or individually using for fundamental frequency characteristic enhancing model in initial speech synthetic model
Fundamental frequency model carry out enhancing processing, can also integrate using above two for different characteristics enhancing model respectively to initial
Spectral model and fundamental frequency model in phonetic synthesis model carry out enhancing processing.Correspondingly, using enhanced spectral model and/
Or fundamental frequency model obtains corresponding to the base frequency parameters and/or frequency spectrum parameter of text to be synthesized, using these phonetic synthesis parameters and leads to
Cross other phonetic synthesis parameters that initial speech synthetic model obtains, you can generation continuous speech signal.
Using above-mentioned each phonetic synthesis parameter generation continuous speech signal process similarly to the prior art, it is no longer superfluous herein
It states.
It can be seen that the method for the embodiment of the present invention carries out enhancing processing to traditional voice synthetic model, in subsequent conjunction
Into in task, only corresponding phonetic synthesis parameter need to be obtained using enhanced phonetic synthesis model, operation will not be increased
Amount, and good enhancing effect can be reached.
As shown in figure 3, it is that the another of voice is synthesized according to initial speech synthetic model and enhancing model in the embodiment of the present invention
A kind of flow chart.
Step 301, duration parameters, the frequency spectrum parameter of corresponding text to be synthesized are generated respectively using initial speech synthetic model
And base frequency parameters.
Step 302, enhancing processing is carried out to frequency spectrum parameter using enhancing model, obtains enhanced frequency spectrum parameter.
Specifically, by the frequency spectrum parameter C in preceding formula (1)sSubstitute into front enhancing processing model p (yt|xt) in xt,
It can obtain enhanced frequency spectrum parameter yt。
Step 303, joined using the duration parameters, base frequency parameters and enhanced frequency spectrum of the correspondence text to be synthesized
Number generation continuous speech signal.
Flow shown in Fig. 3 is unlike flow shown in Fig. 2 it can be seen from above-mentioned flow, in this embodiment, first by
Base frequency parameters, frequency spectrum parameter and the duration parameters of the corresponding text to be synthesized of initial speech synthetic model generation, then again by phase
The enhancing model answered carries out enhancing processing to frequency spectrum parameter therein, so as to which enhanced phonetic synthesis parameter be made preferably to embody
The detail differences in acoustic characteristic when different speakers and same pronunciation human hair difference sound.By these enhancings, treated
Phonetic synthesis parameter passes through synthesizer with being combined together by other phonetic synthesis parameters that traditional voice synthetic model obtains
Synthesize voice.
It should be noted that in practical applications, the flow similar with above-mentioned Fig. 3 equally may be used, using corresponding
Enhancing model carries out enhancing processing to base frequency parameters, obtains enhanced base frequency parameters.Then corresponding text to be synthesized is utilized
Duration parameters, frequency spectrum parameter and enhanced base frequency parameters generation continuous speech signal.Or using for spectral characteristic
Enhancing model carries out the frequency spectrum parameter that initial speech synthetic model generates enhancing processing, while utilizes the increasing for fundamental frequency characteristic
Strong model carries out enhancing processing to the base frequency parameters that initial speech synthetic model generates, and is then given birth to using initial speech synthetic model
Into duration parameters and enhanced base frequency parameters and frequency spectrum parameter generation continuous speech signal.
Correspondingly, the embodiment of the present invention also provides a kind of system for realizing synthesis speech enhan-cement, as shown in figure 4, being that this is
The structure diagram of system.
In this embodiment, the system comprises：
Initial model establishes module 401, for being based on training data structure initial speech synthetic model, the training data
Including text data and voice data corresponding with the text data；
Enhance model building module 402, enhance model for establishing, the enhancing model is used to simulate the initial speech
The synthesis speech parameter of synthetic model generation and the mapping relations of natural-sounding parameter；
Receiving module 403, for receiving text to be synthesized；
Parameter generation module 404, for according to the initial speech synthetic model and the corresponding institute of enhancing model generation
State the synthesis speech parameter of text to be synthesized；
Synthesis module 405, for utilizing the synthesis speech parameter generation continuous speech signal.
Above-mentioned initial model, which establishes module 401, can use traditional parameter synthesis method structure initial speech synthesis mould
Type, the initial speech synthetic model include：The corresponding binary decision tree of each basic synthesis unit, spectral model, fundamental frequency model,
Duration modeling etc..For example, the parameter synthesis method based on HMM may be used, for spectral model, leaf segment is simulated using GMM
The spectrum distribution of point, Gaussage can usually be determined as positive integer with reference to training data scale, for example select Gaussage as 1.
Since the setting of enhancing model has important influence into voice enhancing effect with optimization pairing, in this hair
In bright embodiment, enhancing model building module 402 uses the enhancing model set-up mode based on data-driven, is joined with natural acoustics
The minutia of parameters,acoustic when number is as guidance, real embodiment difference speaker and same pronunciation human hair difference sound, and then
Improve the effect of synthesis speech enhan-cement.
Above-mentioned enhancing model building module 402 can specifically include following each unit：
Speech parameter generation unit is synthesized, for generating the conjunction of all training datas according to the initial speech synthetic model
Into speech parameter；
Natural-sounding parameter extraction unit, for extracting the natural-sounding parameter of all training datas；
Topological structure determination unit, for determining the topological structure of enhancing model；
Training unit, for the data of the synthesis speech parameter of the training data and natural-sounding parameter will to be corresponded to making
Gather for training, parameter training is carried out according to the topological structure, obtain enhancing model.
It should be noted that in practical applications, above-mentioned enhancing model building module 402 can be respectively built for frequency spectrum
The enhancing model of characteristic and/or fundamental frequency characteristic.Correspondingly, in enhancing model of the structure for spectral characteristic, the synthesis language
Sound parameter generating unit needs the skilled synthesis frequency spectrum parameter of spectral model generation institute in initial speech synthetic model；
Natural-sounding parameter extraction unit needs to extract the natural frequency spectrum parameter of all training datas.Similarly, in structure for fundamental frequency
During the enhancing model of characteristic, the synthesis speech parameter generation unit needs the fundamental frequency model in initial speech synthetic model
The skilled synthesis base frequency parameters of generation institute；Natural-sounding parameter extraction unit needs to extract the natural fundamental frequency of all training datas
Parameter.
The enhancing model that above-mentioned parameter generation module 404 is established based on enhancing model building module 402, may be used a variety of
Mode enhances initial speech synthetic model or synthesis speech parameter, can obtain good enhancing effect, correspondingly,
Parameter generation module 404 can will be described in detail later there are many specific implementation structure.
The system provided in an embodiment of the present invention for realizing synthesis speech enhan-cement, Statistics-Based Method structure pass for simulating
The synthesis speech parameter and the enhancing model of the mapping relations of natural-sounding parameter of phonetic synthesis of uniting model generation, then utilizing should
Enhance the synthesis speech parameter of model and the corresponding text to be synthesized of traditional voice synthetic model generation, and then utilize the synthesis language
Sound parameter generates continuous speech signal.Due to it is described enhancing model be using natural parameters,acoustic as instruct, can be to not
The minutia of parameters,acoustic has stronger assurance when same speaker and same pronunciation human hair difference sound, can catch specific
The characteristic of speaker, the effect for making synthesis speech enhan-cement are more preferable.
As shown in figure 5, it is a kind of specific implementation structure diagram of parameter generation module in the embodiment of the present invention.
In this embodiment, the parameter generation module includes：
Model enhancement unit 501, for according to it is described enhancing model to the frequency spectrum mould in the initial speech synthetic model
Type and/or fundamental frequency model carry out enhancing processing, the spectral model and/or fundamental frequency model enhanced；
Enhance speech parameter generation unit 502, for utilizing the spectral model of the enhancing and/or fundamental frequency model generation pair
Answer the frequency spectrum parameter and/or base frequency parameters of the text to be synthesized；
Initial speech parameter generating unit 503, for waiting to close using described in initial speech synthetic model generation correspondence
Into other speech parameters in addition to spectral model and/or fundamental frequency model of text.
It should be noted that in practical applications, aforementioned enhancing model building module 402 can be respectively generated for frequency spectrum
The enhancing model of characteristic and the enhancing model for fundamental frequency characteristic, therefore, in the embodiment shown in fig. 5, model enhancement unit 501
Can enhancing processing individually be carried out to the spectral model in initial speech synthetic model using the enhancing model for spectral characteristic,
Or enhancing processing is individually carried out to the fundamental frequency model in initial speech synthetic model using the enhancing model for fundamental frequency characteristic,
It can also integrate and the enhancing model of different characteristics is directed to respectively to the frequency spectrum mould in initial speech synthetic model using above two
Type and fundamental frequency model carry out enhancing processing.Correspondingly, enhancing speech parameter generation unit 502 can utilize enhanced frequency spectrum mould
Type and/or fundamental frequency model obtain corresponding to the base frequency parameters and/or frequency spectrum parameter of text to be synthesized, 405 profit of synthesis module in Fig. 4
The other phonetic synthesis parameters obtained with these phonetic synthesis parameters and initial speech parameter generating unit 503, you can generation connects
Continuous voice signal.
It can be seen that the system that the embodiment of the present invention realizes synthesis speech enhan-cement enhances traditional voice synthetic model
Processing in subsequent synthesis task, only need to obtain corresponding phonetic synthesis parameter i.e. using enhanced phonetic synthesis model
Can, operand will not be increased, and good enhancing effect can be reached.
Above-mentioned model enhancement unit 501 can by the enhancing to model mean value come the spectral model that is enhanced and/or
Fundamental frequency model, a kind of concrete structure of model enhancement unit 501 can include following each unit：
Model parameter acquiring unit, for obtaining spectral model and/or fundamental frequency mould from the initial speech synthetic model
The model parameter of type；
Model parameter enhancement unit for carrying out enhancing processing to the model parameter using the enhancing model, obtains
Enhanced model parameter；
Enhance model generation unit, for enhanced model parameter to be substituted corresponding spectral model and/or fundamental frequency mould
The model parameter of type, the spectral model enhanced and/or fundamental frequency model.
As shown in fig. 6, it is another specific implementation structure diagram of parameter generation module in the embodiment of the present invention.
In this embodiment, the parameter generation module includes：
Initial speech parameter generating unit 601 is waited to close for generating respectively using initial speech synthetic model described in correspondence
Into the duration parameters of text, frequency spectrum parameter and base frequency parameters；
Parameter enhancement unit 602, for being carried out at enhancing to the frequency spectrum parameter and/or base frequency parameters using enhancing model
Reason obtains enhanced frequency spectrum parameter and/or base frequency parameters, and the enhanced frequency spectrum parameter and/or base frequency parameters is made
To correspond to the frequency spectrum parameter and/or base frequency parameters of the text to be synthesized during synthesis voice.
Unlike structure of block diagram shown in Fig. 5, in this embodiment, first by 601 profit of initial speech parameter generating unit
With base frequency parameters, frequency spectrum parameter and the duration parameters of the corresponding text to be synthesized of initial speech synthetic model generation, then again by
Parameter enhancement unit 602 carries out enhancing processing using corresponding enhancing model to frequency spectrum parameter therein, enhanced so as to make
Phonetic synthesis parameter preferably embody different speakers and it is same pronunciation human hair difference sound when acoustic characteristic on details it is poor
It is different.Synthesis module 405 in Fig. 4 is by these enhancings treated phonetic synthesis parameter with being obtained by traditional voice synthetic model
Other phonetic synthesis parameters be combined together, voice is synthesized by synthesizer.
Using the embodiment of the present invention realize synthesis speech enhan-cement system, by statistical obtain different speakers, with
And during same pronunciation human hair difference sound parameters,acoustic minutia, and then increased using these minutia pairings into voice
Strength is managed, so as to obtain better enhancing effect.
Each embodiment in this specification is described by the way of progressive, identical similar portion between each embodiment
Point just to refer each other, and the highlights of each of the examples are difference from other examples.Especially for system reality
For applying example, since it is substantially similar to embodiment of the method, so describing fairly simple, related part is referring to embodiment of the method
Part explanation.System embodiment described above is only schematical, wherein described be used as separating component explanation
Unit may or may not be physically separate, the component shown as unit may or may not be
Physical unit, you can be located at a place or can also be distributed in multiple network element.It can be according to the actual needs
Some or all of module therein is selected to realize the purpose of this embodiment scheme.Those of ordinary skill in the art are not paying
In the case of creative work, you can to understand and implement.
The embodiment of the present invention is described in detail above, specific embodiment used herein carries out the present invention
It illustrates, the explanation of above example is only intended to help to understand the method and apparatus of the present invention；Meanwhile for the one of this field
As technical staff, thought according to the present invention, there will be changes in specific embodiments and applications, to sum up institute
It states, the content of the present specification should not be construed as limiting the invention.
Claims (8)
- A kind of 1. method for realizing synthesis speech enhan-cement, which is characterized in that including：Based on training data build initial speech synthetic model, the training data include text data and with the text data Corresponding voice data；The initial speech synthetic model includes：Duration modeling, spectral model, fundamental frequency model；Establish enhancing model, the enhancing model be used to simulating the synthesis speech parameter of initial speech synthetic model generation with The mapping relations of natural-sounding parameter；After text to be synthesized is received, treated according to the initial speech synthetic model and the enhancing model generation are corresponding The synthesis speech parameter of synthesis text, including：According to the enhancing model to the frequency spectrum mould in the initial speech synthetic model Type and/or fundamental frequency model carry out enhancing processing, the spectral model and/or fundamental frequency model enhanced；Utilize the frequency of the enhancing The frequency spectrum parameter and/or base frequency parameters of spectrum model and/or the corresponding text to be synthesized of fundamental frequency model generation；Using described initial The ginseng of other voices in addition to spectral model and/or fundamental frequency model of the corresponding text to be synthesized of phonetic synthesis model generation Number；Utilize the synthesis speech parameter generation continuous speech signal.
- 2. according to the method described in claim 1, it is characterized in that, the enhancing model of establishing includes：The synthesis speech parameter of all training datas is generated according to the initial speech synthetic model；Extract the natural-sounding parameter of all training datas；Determine the topological structure of enhancing model；Using the data of the synthesis speech parameter of the correspondence training data and natural-sounding parameter to gathering as training, according to institute It states topological structure and carries out parameter training, obtain enhancing model.
- 3. according to the method described in claim 2, it is characterized in that, the enhancing model is：The mapping model of linear function or Person's GMM model or DNN models.
- 4. the according to the method described in claim 1, it is characterized in that, synthesis voice ginseng of initial speech synthetic model generation The mapping relations of number and natural-sounding parameter are the synthesis speech parameter and natural-sounding of initial speech synthetic model generation The condition distribution of parameter.
- 5. according to the method described in claim 1, it is characterized in that, described close the initial speech according to the enhancing model Enhancing processing is carried out into the spectral model in model and/or fundamental frequency model, the spectral model and/or fundamental frequency model packet enhanced It includes：Spectral model and/or the model parameter of fundamental frequency model are obtained from the initial speech synthetic model；Enhancing processing is carried out to the model parameter using the enhancing model, obtains enhanced model parameter；Enhanced model parameter is substituted into corresponding spectral model and/or the model parameter of fundamental frequency model, the frequency enhanced Spectrum model and/or fundamental frequency model.
- 6. a kind of system for realizing synthesis speech enhan-cement, which is characterized in that including：Initial model establishes module, and for being based on training data structure initial speech synthetic model, the training data includes text Notebook data and voice data corresponding with the text data；The initial speech synthetic model includes：Duration modeling, frequency spectrum mould Type, fundamental frequency model；Enhance model building module, enhance model for establishing, the enhancing model is used to simulate the initial speech synthesis mould The synthesis speech parameter of type generation and the mapping relations of natural-sounding parameter；Receiving module, for receiving text to be synthesized；Parameter generation module, for described to be synthesized according to the initial speech synthetic model and the enhancing model generation correspondence The synthesis speech parameter of text；The parameter generation module includes：Model enhancement unit, for according to it is described enhancing model to institute The spectral model and/or fundamental frequency model stated in initial speech synthetic model carry out enhancing processing, the spectral model enhanced and/ Or fundamental frequency model；Enhance speech parameter generation unit, for utilizing the spectral model of the enhancing and/or fundamental frequency model generation pair Answer the frequency spectrum parameter and/or base frequency parameters of the text to be synthesized；Initial speech parameter generating unit, it is described initial for utilizing The ginseng of other voices in addition to spectral model and/or fundamental frequency model of the corresponding text to be synthesized of phonetic synthesis model generation Number；Synthesis module, for utilizing the synthesis speech parameter generation continuous speech signal.
- 7. system according to claim 6, which is characterized in that the enhancing model building module includes：Speech parameter generation unit is synthesized, for generating the synthesis language of all training datas according to the initial speech synthetic model Sound parameter；Natural-sounding parameter extraction unit, for extracting the natural-sounding parameter of all training datas；Topological structure determination unit, for determining the topological structure of enhancing model；Training unit, for the data of the synthesis speech parameter of the training data and natural-sounding parameter will to be corresponded to as instruction Practice set, parameter training is carried out according to the topological structure, obtain enhancing model.
- 8. system according to claim 6, which is characterized in that the model enhancement unit includes：Model parameter acquiring unit, for obtaining spectral model and/or fundamental frequency model from the initial speech synthetic model Model parameter；Model parameter enhancement unit for carrying out enhancing processing to the model parameter using the enhancing model, is enhanced Model parameter afterwards；Enhance model generation unit, for enhanced model parameter to be substituted corresponding spectral model and/or fundamental frequency model Model parameter, the spectral model enhanced and/or fundamental frequency model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410182886.6A CN105023574B (en) | 2014-04-30 | 2014-04-30 | A kind of method and system for realizing synthesis speech enhan-cement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410182886.6A CN105023574B (en) | 2014-04-30 | 2014-04-30 | A kind of method and system for realizing synthesis speech enhan-cement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105023574A CN105023574A (en) | 2015-11-04 |
| CN105023574B true CN105023574B (en) | 2018-06-15 |
Family
ID=54413492
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410182886.6A Active CN105023574B (en) | 2014-04-30 | 2014-04-30 | A kind of method and system for realizing synthesis speech enhan-cement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105023574B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109427340A (en) * | 2017-08-22 | 2019-03-05 | 杭州海康威视数字技术股份有限公司 | A kind of sound enhancement method, device and electronic equipment |
| CN108231086A (en) * | 2017-12-24 | 2018-06-29 | 航天恒星科技有限公司 | A kind of deep learning voice enhancer and method based on FPGA |
| CN109346058A (en) * | 2018-11-29 | 2019-02-15 | 西安交通大学 | A kind of speech acoustics feature expansion system |
| CN111613224A (en) * | 2020-04-10 | 2020-09-01 | 云知声智能科技股份有限公司 | Personalized voice synthesis method and device |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6038533A (en) * | 1995-07-07 | 2000-03-14 | Lucent Technologies Inc. | System and method for selecting training text |
| CN1835075A (en) * | 2006-04-07 | 2006-09-20 | 安徽中科大讯飞信息科技有限公司 | Speech synthetizing method combined natural sample selection and acaustic parameter to build mould |
| CN1835074A (en) * | 2006-04-07 | 2006-09-20 | 安徽中科大讯飞信息科技有限公司 | Speaking person conversion method combined high layer discription information and model self adaption |
| CN102568476A (en) * | 2012-02-21 | 2012-07-11 | 南京邮电大学 | Voice conversion method based on self-organizing feature map network cluster and radial basis network |
| CN102982809A (en) * | 2012-12-11 | 2013-03-20 | 中国科学技术大学 | Conversion method for sound of speaker |
| CN103065619A (en) * | 2012-12-26 | 2013-04-24 | 安徽科大讯飞信息科技股份有限公司 | Speech synthesis method and speech synthesis system |
-
2014
- 2014-04-30 CN CN201410182886.6A patent/CN105023574B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6038533A (en) * | 1995-07-07 | 2000-03-14 | Lucent Technologies Inc. | System and method for selecting training text |
| CN1835075A (en) * | 2006-04-07 | 2006-09-20 | 安徽中科大讯飞信息科技有限公司 | Speech synthetizing method combined natural sample selection and acaustic parameter to build mould |
| CN1835074A (en) * | 2006-04-07 | 2006-09-20 | 安徽中科大讯飞信息科技有限公司 | Speaking person conversion method combined high layer discription information and model self adaption |
| CN102568476A (en) * | 2012-02-21 | 2012-07-11 | 南京邮电大学 | Voice conversion method based on self-organizing feature map network cluster and radial basis network |
| CN102982809A (en) * | 2012-12-11 | 2013-03-20 | 中国科学技术大学 | Conversion method for sound of speaker |
| CN103065619A (en) * | 2012-12-26 | 2013-04-24 | 安徽科大讯飞信息科技股份有限公司 | Speech synthesis method and speech synthesis system |
Non-Patent Citations (3)
| Title |
|---|
| 《Estimation of Window Coefficients for Dynamic Feature Extraction for HMM-Based Speech Synthesis》;Ling-Hui Chen ET AL.;《INTERSPEECH 2011 12th Annual Conference of the International Speech Communication Association》;20110831;第1801-1804页 * |
| 《基于声学统计建模的语音合成技术研究》;胡郁等;《中文信息学报》;20111130;第25卷(第6期);第127-136页 * |
| 《基于话者无关模型的说话人转换方法》;陈凌辉等;《模式识别与人工智能》;20130331;第26卷(第3期);第254-259页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105023574A (en) | 2015-11-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105118498B (en) | The training method and device of phonetic synthesis model | |
| Wu et al. | Deep neural networks employing multi-task learning and stacked bottleneck features for speech synthesis | |
| CN104036774B (en) | Tibetan dialect recognition methods and system | |
| CN103928023B (en) | A kind of speech assessment method and system | |
| CN104766603B (en) | Construct the method and device of personalized singing style Spectrum synthesizing model | |
| CN110223705A (en) | Phonetics transfer method, device, equipment and readable storage medium storing program for executing | |
| CN101064104B (en) | Emotion voice creating method based on voice conversion | |
| CN101490740B (en) | Audio combining device | |
| CN1835074B (en) | Speaking person conversion method combined high layer discription information and model self adaption | |
| CN104272382A (en) | Method and system for template-based personalized singing synthesis | |
| CN105023574B (en) | A kind of method and system for realizing synthesis speech enhan-cement | |
| CN105590625A (en) | Acoustic model self-adaptive method and system | |
| CN105023570B (en) | A kind of method and system for realizing sound conversion | |
| CN106875942A (en) | Acoustic model adaptive approach based on accent bottleneck characteristic | |
| CN105551071A (en) | Method and system of face animation generation driven by text voice | |
| CN107452379A (en) | The identification technology and virtual reality teaching method and system of a kind of dialect language | |
| CN101578659A (en) | Voice tone converting device and voice tone converting method | |
| CN109523616A (en) | A kind of FA Facial Animation generation method, device, equipment and readable storage medium storing program for executing | |
| CN103456295B (en) | Sing synthetic middle base frequency parameters and generate method and system | |
| CN109952609A (en) | Speech synthesizing method | |
| CN110491393A (en) | The training method and relevant apparatus of vocal print characterization model | |
| CN105206257A (en) | Voice conversion method and device | |
| CN108053814A (en) | A kind of speech synthesis system and method for analog subscriber song | |
| CN108877835A (en) | Evaluate the method and system of voice signal | |
| CN108417207A (en) | A kind of depth mixing generation network self-adapting method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information |
Address after: Wangjiang Road high tech Development Zone Hefei city Anhui province 230088 No. 666 Applicant after: Iflytek Co., Ltd. Address before: Wangjiang Road high tech Development Zone Hefei city Anhui province 230088 No. 666 Applicant before: Anhui USTC iFLYTEK Co., Ltd. |
|
| COR | Change of bibliographic data | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |