WO2023160553A1 - 语音合成方法、装置、计算机可读介质及电子设备 - Google Patents

语音合成方法、装置、计算机可读介质及电子设备 Download PDF

Info

Publication number
WO2023160553A1
WO2023160553A1 PCT/CN2023/077478 CN2023077478W WO2023160553A1 WO 2023160553 A1 WO2023160553 A1 WO 2023160553A1 CN 2023077478 W CN2023077478 W CN 2023077478W WO 2023160553 A1 WO2023160553 A1 WO 2023160553A1
Authority
WO
WIPO (PCT)
Prior art keywords
text
synthesized
sequence
prosodic
phoneme
Prior art date
Application number
PCT/CN2023/077478
Other languages
English (en)
French (fr)
Inventor
林浩鹏
马泽君
Original Assignee
北京有竹居网络技术有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 北京有竹居网络技术有限公司 filed Critical 北京有竹居网络技术有限公司
Publication of WO2023160553A1 publication Critical patent/WO2023160553A1/zh

Links

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L13/00—Speech synthesis; Text to speech systems
    • G10L13/08—Text analysis or generation of parameters for speech synthesis out of text, e.g. grapheme to phoneme translation, prosody generation or stress or intonation determination
    • G10L13/10—Prosody rules derived from text; Stress or intonation
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks

Definitions

  • the present disclosure relates to the technical field of speech synthesis, and in particular, to a speech synthesis method, device, computer readable medium and electronic equipment.
  • prosody refers to the composition of dependent segments (vowels and consonants) in speech, that is, the properties of syllables or larger units. These properties form language functions such as intonation, intonation, stress and rhythm. Prosody can reflect many features of a speaker or utterance: the speaker's emotional state, the form of the utterance (statement, question, or command), the presence or absence of emphasis, contrast, focus, and other language that cannot be characterized by grammar and lexical expressions Elements, different expressions of the same prosodic event can convey rich semantics and emotional changes. In tasks such as speech synthesis, how to combine the prosodic features of the text to make the synthesized audio more natural and smooth has become the focus of research.
  • the present disclosure provides a speech synthesis method, including:
  • a tone and break index (Tones and Break Indices, TOBI) characterizing sequence and prosodic acoustic features of the phoneme level corresponding to the text to be synthesized are generated, and according to the TOBI characterizing sequence and The prosodic acoustic feature generates acoustic feature information corresponding to the text to be synthesized;
  • first audio information corresponding to the text to be synthesized is generated.
  • the present disclosure provides a speech synthesis device, including:
  • An acquisition module configured to acquire a phoneme sequence corresponding to the text to be synthesized
  • the first generation module is configured to generate the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized according to the phoneme sequence and the text to be synthesized obtained by the acquisition module, and according to the TOBI characterizing the sequence and the prosodic acoustic features, and generating acoustic feature information corresponding to the text to be synthesized;
  • the second generating module is configured to generate first audio information corresponding to the text to be synthesized according to the acoustic feature information generated by the first generating module.
  • the present disclosure provides a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processing device, the steps of the method provided in the first aspect of the present disclosure are implemented.
  • an electronic device including:
  • At least one processing device configured to execute the at least one computer program in the storage device to implement the steps of the method provided in the first aspect of the present disclosure.
  • the present disclosure provides a computer program, which implements the steps of the method provided in the first aspect of the present disclosure when the computer program is executed by a processing device.
  • the present disclosure provides a computer program product, where the computer program product includes a computer program, and when the computer program is executed by a processing device, the steps of the method provided in the first aspect of the present disclosure are implemented.
  • the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized are generated, and according to the TOBI representation sequence and Prosodic acoustic features, generating acoustic feature information corresponding to the text to be synthesized; finally, generating first audio information corresponding to the text to be synthesized according to the acoustic feature information.
  • the corresponding TOBI representation sequence and prosodic acoustic features of the text to be synthesized are referred to at the same time, that is, not only the prosodic features of the language level of the text to be synthesized are referred to, but also the prosodic features of the acoustic level of the text to be synthesized are taken into account. performance in different dimensions.
  • different sentences can be endowed with appropriate rhythm, emphasis and intonation characteristics, and the corresponding prosodic acoustic features can explicitly reflect the specific acoustic manifestation of the corresponding prosodic event, thereby improving the prosodic naturalness of the synthesized audio while controlling the audio
  • different intensities can be assigned to multiple stress positions to achieve different emphasis of semantic expression, or the intonation changes of interrogative sentences can be realized through intensity adjustment to convey different semantics (emotions). Therefore, under the same prosodic language expression, different prosodic acoustic features can reflect different semantic changes, thereby making the synthesized audio more natural, more cadenced, and more in line with the semantics expressed by the speaker.
  • Fig. 1 is a flowchart of a speech synthesis method according to an exemplary embodiment.
  • Fig. 2 is a schematic structural diagram of a speech synthesis model according to an exemplary embodiment.
  • Fig. 3 is a schematic structural diagram of a prosodic language feature prediction module according to an exemplary embodiment.
  • Fig. 4 is a flowchart showing a method for training a speech synthesis model according to an exemplary embodiment.
  • Fig. 5 is a flowchart of a speech synthesis method according to another exemplary embodiment.
  • Fig. 6 is a block diagram of a speech synthesis device according to an exemplary embodiment.
  • Fig. 7 is a block diagram of an electronic device according to an exemplary embodiment.
  • the current speech synthesis method mainly uses the prosodic features of the language level, that is, the artificially labeled TOBI (Tones and Break Indices) data to realize the rhythm control of the synthesized audio, so as to improve the naturalness of the speech synthesis. , but the intensity of the synthesized audio is not controllable.
  • TOBI Tones and Break Indices
  • the present disclosure provides a speech synthesis method, device, computer-readable medium and electronic equipment.
  • the term “comprise” and its variations are open-ended, ie “including but not limited to”.
  • the term “based on” is “based at least in part on”.
  • the term “one embodiment” means “at least one embodiment”; the term “another embodiment” means “at least one further embodiment”; the term “some embodiments” means “at least some embodiments.” Relevant definitions of other terms will be given in the description below.
  • Fig. 1 is a flowchart of a speech synthesis method according to an exemplary embodiment. As shown in FIG. 1, the method includes S101-S103.
  • the above-mentioned text to be synthesized may be Chinese, English, Japanese and other language texts.
  • a phoneme sequence corresponding to the text to be synthesized may be obtained through a grapheme-to-phoneme (G2P) model.
  • the G2P model can use a recurrent neural network (Recurrent Neural Network, RNN) and a long-short-term memory network (Long Short-Term Memory, LSTM) to realize the transformation from grapheme to phoneme.
  • RNN Recurrent Neural Network
  • LSTM Long Short-Term Memory
  • a phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized are generated, and according to the TOBI representation sequence and prosodic acoustic features, acoustic feature information corresponding to the text to be synthesized is generated.
  • the TOBI representation sequence is used to reflect the prosodic features of the language level of the text to be synthesized, that is, the prosodic language feature, which refers to the prosodic language phenomenon defined by the ToBI system in primitive linguistics, which belongs to discrete features, and can specifically include tone , intonation, pitch stress, and prosodic boundaries.
  • tone refers to the change of the pitch of the voice.
  • tone refers to the change of the pitch of the voice.
  • Intonation that is, the tone of speech, is the configuration and change of speed, speed, and severity in a sentence. a sentence except words There are lexical meaning and intonation meaning. The meaning of intonation is the attitude or tone expressed by the speaker's intonation. The complete meaning of a sentence is the lexical meaning plus the intonation meaning. The same sentence, with different intonations, has different meanings, sometimes even thousands of miles away.
  • Pitch accent used to describe the pitch change of stressed syllables, can control the rhythm of emphasized information and accented rhythmic language, and its scope is on the main stressed syllable, or the main stress and the main stress of the same word on one syllable.
  • the pitch and stress control is only performed on the main stressed syllable, and other redundant information such as secondary stress and zero stress are ignored, so as to achieve the effect of information simplification.
  • the pitch-accent information is used to indicate the syllable position in the text to be synthesized where the specified accent phenomenon exists, wherein the specified accent phenomenon may include high-accent, low-accent, rising-accent, low-rising-accent and high-falling-accent.
  • the pitch target is high, the fundamental frequency curve (f0) is high and flat, and the sense of hearing is Chinese Yinping;
  • the pitch target is low, the fundamental frequency curve is low and flat, and the sense of hearing is Chinese The first half of the upper voice;
  • the rising stress, the pitch target is high, the fundamental frequency curve shows a rising trend, and the sense of hearing is Chinese Yangping;
  • the low rising stress, the pitch target is low, and if it acts on a single syllable, the fundamental frequency curve shows a downward trend , there is a slight rise at the end.
  • the fundamental frequency curve will show a downward trend in the main stress, and a rising trend in the syllable after the main stress, and the sense of hearing is Chinese upper tone;
  • the fundamental frequency curve shows a downward trend, and the sense of hearing is Chinese.
  • Prosodic boundaries are used to indicate where pauses should be made when text is to be synthesized. For example, the prosodic boundary is divided into four pause levels of "#1", “#2", “#3” and "#4", and the pause levels increase sequentially. Among them, English and Japanese have no obvious prosodic level, so this disposition is empty.
  • the prosodic acoustic feature (that is, the prosodic feature at the acoustic level) is a broad definition of the measured physical quantity that represents the acoustic characteristics of speech, such as timbre, formant, fundamental frequency, or formant intensity.
  • the acoustic features that are more closely related to the prosodic events defined by the ToBI system in linguistics duration, fundamental frequency, and energy.
  • the high rising tone of the prosodic language feature "sentence tone" can be specifically expressed as the corresponding fundamental frequency in a speech segment Continue to climb to the fundamental frequency highs in a sentence. Therefore, the prosodic acoustic feature in the present disclosure includes at least one of the fundamental frequency, energy and pronunciation duration of the phoneme level corresponding to the text to be synthesized, which is a continuous feature.
  • Acoustic feature information may be, for example, a mel spectrum, a spectral envelope, or the like.
  • first audio information corresponding to the text to be synthesized is generated.
  • the first audio information corresponding to the text to be synthesized can be obtained by inputting the acoustic feature information into the vocoder, wherein the vocoder can be, for example, a Wavenet vocoder, a Griffin-Lim vocoder, etc. .
  • the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized are generated, and according to the TOBI representation sequence and Prosodic acoustic features, generating acoustic feature information corresponding to the text to be synthesized; finally, generating first audio information corresponding to the text to be synthesized according to the acoustic feature information.
  • the corresponding TOBI representation sequence and prosodic acoustic features of the text to be synthesized are referred to at the same time, that is, not only the prosodic features of the language level of the text to be synthesized are referred to, but also the prosodic features of the acoustic level of the text to be synthesized are taken into account. performance in different dimensions.
  • different sentences can be endowed with appropriate rhythm, emphasis and intonation characteristics, and the corresponding prosodic acoustic features can explicitly reflect the specific acoustic manifestation of the corresponding prosodic event, thereby improving the prosodic naturalness of the synthesized audio while controlling the audio
  • different intensities can be assigned to multiple stress positions to achieve different emphasis of semantic expression, or the intonation changes of interrogative sentences can be realized through intensity adjustment to convey different semantics (emotions). Therefore, under the same prosodic language expression, different prosodic acoustic features can reflect different semantic changes, thereby making the synthesized audio more natural, more cadenced, and more in line with the semantics expressed by the speaker.
  • the phoneme sequence and the text to be synthesized can be input into the pre-trained speech synthesis model, so that the speech synthesis model can generate the phoneme-level TOBI representation sequence and prosody corresponding to the text to be synthesized according to the phoneme sequence and the text to be synthesized Acoustic features, and according to the TOBI characterization sequence and prosodic acoustic features, generate acoustic feature information corresponding to the text to be synthesized.
  • the above-mentioned speech synthesis model includes an encoding network, an attention network, a decoding network, a prosodic language feature prediction module, a prosodic acoustic feature prediction module, an embedding layer, a first splicing module, a second splicing module and a third splicing module , wherein, the prosodic language feature prediction module, the first splicing module, the encoding network, the second splicing module, the prosodic acoustic feature prediction module, the third splicing module, the attention network, and the decoding network are sequentially connected, and the first splicing module is also connected with The embedding layer is connected, the second stitching module is also connected with the prosodic acoustic feature prediction module, and the third stitching module is also connected with the encoding network.
  • the prosodic language feature prediction module is used to generate a phoneme-level TOBI representation sequence corresponding to the text to be synthesized according to the text to be synthesized.
  • the embedding layer is used to generate a phoneme representation sequence corresponding to the text to be synthesized according to the phoneme sequence, wherein the phoneme representation sequence is formed by sorting the word vectors corresponding to each phoneme in the text to be synthesized according to the sequence of the corresponding phonemes in the text to be synthesized,
  • the word vectors corresponding to each phoneme in the synthesized text may be determined according to the pre-established correspondence between phonemes and word vectors.
  • the first splicing module is configured to splice the phoneme-level TOBI representation sequence and the phoneme representation sequence to obtain a first splicing sequence.
  • An encoding network is used to encode the first spliced sequence to generate an encoded sequence.
  • the second splicing module is configured to splice the coding sequence and the phoneme-level TOBI representation sequence to obtain a second spliced sequence.
  • the prosodic acoustic feature prediction module is configured to generate prosodic acoustic features corresponding to the text to be synthesized according to the second splicing sequence.
  • the rhythmic acoustic feature prediction module may be a shallow network composed of a convolutional layer + a bidirectional LSTM layer + a fully connected layer.
  • the third splicing module is used for splicing the coding sequence and the prosodic acoustic features to obtain the third splicing sequence;
  • the attention network is configured to generate a semantic representation corresponding to the text to be synthesized according to the third splicing sequence.
  • the attention network may be a location-sensitive attention (Locative Sensitive Attention) attention network, or an attention network based on a Gaussian Mixture Model (GMM), that is, GMM attention.
  • GMM Gaussian Mixture Model
  • the decoding network is used to generate acoustic feature information corresponding to the text to be synthesized according to the semantic representation.
  • the above-mentioned prosodic language feature prediction module includes a first sub-embedding layer, a prosodic language feature prediction network, a second sub-embedding layer and an extension layer connected in sequence.
  • the first sub-embedding layer is used to extract a word-level deep representation corresponding to the text to be synthesized.
  • the first sub-embedding layer may be a TinyBert model based on distillation learning.
  • the prosodic language feature prediction network is used to generate word-level TOBI tags based on deep representations.
  • TOBI tags may include intonation, intonation, pitch stress and prosodic boundary.
  • the prosodic language feature prediction network may be a shallow network composed of a convolutional layer + a bidirectional LSTM layer + a fully connected layer.
  • the second sub-embedding layer is used to generate a word-level TOBI representation sequence corresponding to the text to be synthesized according to the TOBI tag.
  • the extension layer is used to extend the word-level TOBI representation sequence to obtain the phoneme-level TOBI representation sequence corresponding to the text to be synthesized.
  • the word-level TOBI representation corresponding to the word can be copied L-1 times to obtain the phoneme-level TOBI representation corresponding to the word, where L is the word contained in the word number of phonemes.
  • the text to be synthesized includes word A and word B connected in sequence, wherein, word A includes three phonemes, word B includes 4 phonemes, the TOBI representation of the word level corresponding to word A is M, and the word level corresponding to word B
  • the TOBI representation of the word A is N
  • the TOBI representation of the phoneme level corresponding to the word A is MMM
  • the TOBI representation of the word B is NNNN
  • the TOBI representation sequence of the phoneme level corresponding to the text to be synthesized is MMMNNNN.
  • the above-mentioned speech synthesis model can be obtained through training through S401-S403 shown in FIG. 4 .
  • a training phoneme sequence corresponding to the training text, a word-level training TOBI label, a training prosodic acoustic feature, and training acoustic feature information are determined.
  • the training text can be the text extracted from the real speech, and the labeler can first mark the word-level TOBI corresponding to the training text by listening to the speech corresponding to the training text (that is, the word-level training TOBI Label).
  • the training phoneme sequence corresponding to the training text may be acquired in the same manner as the phoneme sequence corresponding to the text to be synthesized in S101 above.
  • the training prosodic acoustic features corresponding to the training text can be determined in the following way: based on open source tools (such as librosa or straight, etc.), the frame-level fundamental frequency and energy features can be extracted from the real speech corresponding to the training text, and then, can be For each phoneme in the training text, the average value of the fundamental frequencies of the multiple frames corresponding to the phoneme is used as the fundamental frequency of the phoneme, and the average value of the energy of the multiple frames corresponding to the phoneme is used as the energy of the phoneme, that is Obtain the phoneme-level fundamental frequency and phoneme-level energy; at the same time, obtain the pronunciation duration of each phoneme in the training text based on the forced alignment tool.
  • open source tools such as librosa or straight, etc.
  • training acoustic feature information corresponding to the training text can be obtained by inputting the training text into a speech synthesis model (eg, Tacotron model, Deepvoice 3 model, Tacotron 2 model, Wavenet model, etc.).
  • a speech synthesis model eg, Tacotron model, Deepvoice 3 model, Tacotron 2 model, Wavenet model, etc.
  • the output of the first sub-embedding layer is used as the input of the prosodic language feature prediction network, and the word-level training TOBI label is used as the target output of the prosodic language feature prediction network
  • the output of the prosodic language feature prediction network is used as the input of the second sub-embedding layer
  • the output of the second sub-embedding layer is used as the input of the expansion layer
  • the training phoneme sequence is used as the input of the embedding layer
  • the output of the expansion layer and the embedding layer The output of the first stitching module is used as the input of the first stitching module
  • the output of the first stitching module is used as the input of the encoding network
  • the output of the encoding network and the output of the expansion layer are used as the input of the second stitching module
  • the output of the second stitching module is used as the prosody
  • the loss function during speech synthesis model training is the sum of acoustic feature information loss and prosody feature loss.
  • the acoustic feature information loss is the mean square error between the acoustic feature information predicted by the decoding network and the training acoustic feature information
  • the prosodic feature loss includes the prediction loss of prosodic language features and the prediction loss of prosodic acoustic features, where the prediction of prosodic language features
  • the loss is the cross-entropy loss between the word-level TOBI predicted by the prosodic language feature prediction network and the word-level training TOBI label
  • the prediction loss of the prosodic acoustic feature is the difference between the prosodic acoustic feature predicted by the prosodic acoustic feature prediction module and the training prosodic acoustic feature. mean square error between.
  • the above method may further include the following S104.
  • the first audio information is synthesized with the target background music to obtain second audio information.
  • the above-mentioned target background music may be preset music, that is, any music set by the user, or default music.
  • the usage scene information corresponding to the text to be synthesized may be determined according to the text information of the text to be synthesized, wherein the usage scene information includes But not limited to news broadcasts, military introductions, fairy tales, campus broadcasts, etc.; then, according to the usage scenario information, determine the target background music that matches the usage scenario information.
  • the above-mentioned text information may be keywords.
  • automatic keyword identification of the text to be synthesized may be performed to intelligently predict usage scene information of the text to be synthesized according to keywords.
  • the target background music matching the use scene information may be determined by using the pre-stored correspondence between the use scene information and the background music according to the use scene information. For example, if the scene information is military introduction, the corresponding background music may be exciting music; if the scene information is a fairy tale, the corresponding background music may be brisk and lively music.
  • Fig. 6 is a block diagram of a speech synthesis device according to an exemplary embodiment. As shown in Figure 6, the device 600 includes:
  • Obtaining module 601 for obtaining the phoneme sequence corresponding to the text to be synthesized
  • the first generation module 602 is configured to generate the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized according to the phoneme sequence and the text to be synthesized obtained by the acquisition module 601, and according to the The TOBI representation sequence and the prosodic acoustic features are used to generate acoustic feature information corresponding to the text to be synthesized;
  • the second generation module 603 is configured to generate first audio information corresponding to the text to be synthesized according to the acoustic feature information generated by the first generation module 602 .
  • the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized are generated, and according to the TOBI representation sequence and Prosodic acoustic features, generating acoustic feature information corresponding to the text to be synthesized; finally, generating first audio information corresponding to the text to be synthesized according to the acoustic feature information.
  • the corresponding TOBI representation sequence and prosodic acoustic features of the text to be synthesized are referred to at the same time, that is, not only the prosodic features of the language level of the text to be synthesized are referred to, but also the prosodic features of the acoustic level of the text to be synthesized are taken into account. performance in different dimensions.
  • different sentences can be endowed with appropriate rhythm, emphasis and intonation characteristics, and the corresponding prosodic acoustic features can explicitly reflect the specific acoustic manifestation of the corresponding prosodic event, thereby improving the prosodic naturalness of the synthesized audio while controlling the audio
  • different intensities can be assigned to multiple stress positions to achieve different emphasis of semantic expression, or the intonation changes of interrogative sentences can be realized through intensity adjustment to convey different semantics (emotions). Therefore, under the same prosodic language expression, different prosodic acoustic features can reflect different semantic changes, thereby making the synthesized audio more natural, more cadenced, and more in line with the semantics expressed by the speaker.
  • the first generating module 602 is configured to input the phoneme sequence and the text to be synthesized into a pre-trained speech synthesis model, so that the phoneme sequence and the For the text to be synthesized, the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized are generated, and according to the TOBI representation sequence and the prosodic acoustic features to generate acoustic feature information corresponding to the text to be synthesized.
  • the speech synthesis model includes an encoding network, an attention network, a decoding network, a prosodic language feature prediction module, a prosodic acoustic feature prediction module, an embedding layer, a first splicing module, a second splicing module, and a third splicing module;
  • the prosodic language feature prediction module is used to generate a phoneme-level TOBI representation sequence corresponding to the text to be synthesized according to the text to be synthesized;
  • the embedding layer is configured to generate a phoneme representation sequence corresponding to the text to be synthesized according to the phoneme sequence;
  • the first splicing module is configured to splice the phoneme-level TOBI representation sequence and the phoneme representation sequence to obtain a first splicing sequence
  • the encoding network is configured to encode the first spliced sequence to generate a coding sequence
  • the second splicing module is configured to splice the coding sequence and the phoneme-level TOBI representation sequence to obtain a second splicing sequence
  • the prosodic acoustic feature prediction module is configured to generate prosodic acoustic features corresponding to the text to be synthesized according to the second splicing sequence;
  • the third splicing module is configured to splice the coding sequence and the prosodic acoustic features to obtain a third splicing sequence
  • the attention network is configured to generate a semantic representation corresponding to the text to be synthesized according to the third splicing sequence
  • the decoding network is configured to generate acoustic feature information corresponding to the text to be synthesized according to the semantic representation.
  • the prosodic language feature prediction module includes a first sub-embedding layer, a prosodic language feature prediction network, a second sub-embedding layer and an extension layer connected in sequence;
  • the first sub-embedding layer is used to extract the deep representation of the word level corresponding to the text to be synthesized
  • the prosodic language feature prediction network is used to generate TOBI tags at the word level according to the deep representation
  • the second sub-embedding layer is used to generate the TOBI characterization sequence corresponding to the word level of the text to be synthesized according to the TOBI tag;
  • the extension layer is configured to expand the word-level TOBI representation sequence to obtain a phoneme-level TOBI representation sequence corresponding to the text to be synthesized.
  • the speech synthesis model is obtained by training a model training device, wherein the model training device includes:
  • training text acquisition module used to obtain training text
  • Determining module for determining the corresponding training phoneme sequence of described training text, the training TOBI label of word level, training prosodic acoustic feature and training acoustic feature information
  • the training module is used to use the training text as the input of the first sub-embedding layer, and the output of the first sub-embedding layer as the input of the prosodic language feature prediction network, and the word-level training
  • the TOBI label is used as the target output of the prosodic language feature prediction network
  • the output of the prosodic language feature prediction network is used as the input of the second sub-embedding layer
  • the output of the second sub-embedding layer is used as the expansion layer
  • the input of the training phoneme sequence is used as the input of the embedding layer
  • the output of the expansion layer and the output of the embedding layer are used as the input of the first splicing module
  • the output of the first splicing module is As the input of the coding network
  • the output of the coding network and the output of the expansion layer are used as the input of the second concatenation module
  • the output of the second concatenation module is used as the prosodic acoustic feature prediction module.
  • Input using the training prosodic acoustic feature as the target output of the prosodic acoustic feature prediction module, using the output of the prosodic acoustic feature prediction module and the output of the encoding network as the input of the third splicing module, and using the The output of the third splicing module is used as the input of the attention network, and the output of the attention network is used as the decoding
  • the input of the network is to perform model training by using the training acoustic feature information as the target output of the decoding network to obtain the speech synthesis model.
  • the prosodic acoustic feature includes at least one of fundamental frequency, energy and pronunciation duration of the phoneme level corresponding to the text to be synthesized.
  • the device 600 also includes:
  • a synthesis module configured to synthesize the first audio information and target background music to obtain second audio information.
  • model training device may be integrated in the above-mentioned speech synthesis device 600, or may be independent of the above-mentioned speech synthesis device 600, which is not specifically limited in the present disclosure.
  • the present disclosure also provides a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processing device, the steps of the above-mentioned speech synthesis method provided by the present disclosure are realized.
  • FIG. 7 it shows a schematic structural diagram of an electronic device (terminal device or server) 700 used to implement an embodiment of the present disclosure.
  • the terminal equipment in the embodiment of the present disclosure may include but not limited to such as mobile phone, notebook computer, digital broadcast receiver, personal digital assistant (Personal Digital Assistant, PDA), tablet computer (PAD), portable multimedia player (Portable multimedia player) , PMP), mobile terminals such as vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and fixed terminals such as digital TVs, desktop computers, and the like.
  • PDA Personal Digital Assistant
  • PMP portable multimedia player
  • mobile terminals such as vehicle-mounted terminals (such as vehicle-mounted navigation terminals)
  • fixed terminals such as digital TVs, desktop computers, and the like.
  • the electronic device shown in FIG. 7 is only an example, and should not limit the functions and application scope of the embodiments of the present disclosure.
  • an electronic device 700 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 708 is loaded into the program in the random access memory (Random Access Memory, RAM) 703 to execute various appropriate actions and processes.
  • a processing device such as a central processing unit, a graphics processing unit, etc.
  • RAM Random Access Memory
  • various programs and data necessary for the operation of the electronic device 700 are also stored.
  • the processing device 701, ROM 702, and RAM 703 are connected to each other through a bus 704.
  • An input/output (I/O) interface 705 is also connected to the bus 704 .
  • an input device 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; including, for example, a liquid crystal display (Liquid Crystal Display, LCD) , an output device 707 such as a speaker, a vibrator, etc.; a storage device 708 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 709.
  • the communication means 709 may allow the electronic device 700 to communicate with other devices wirelessly or by wire to exchange data. While FIG. 7 shows electronic device 700 having various means, it should be understood that implementing or having all of the means shown is not a requirement. More or fewer means may alternatively be implemented or provided.
  • embodiments of the present disclosure include a computer program product, which includes a computer program carried on a non-transitory computer readable medium, where the computer program includes program code for executing the method shown in the flowchart.
  • the computer program may be downloaded and installed from a network via communication means 709, or from storage means 708, or from ROM 702.
  • the processing device 701 the above-mentioned functions defined in the methods of the embodiments of the present disclosure are executed.
  • the computer-readable medium mentioned above in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two.
  • a computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination thereof.
  • Computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable Read Only Memory (Erasable Programmable Read Only Memory, EPROM or flash memory), optical fiber, portable compact disk read-only Memory (Compact Disk Read Only Memory, CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave carrying computer-readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device .
  • Program code embodied on a computer readable medium may be transmitted by any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.
  • the client and the server can communicate using any currently known or future network protocols such as Hypertext Transfer Protocol (HyperText Transfer Protocol, HTTP), and can communicate with digital data in any form or medium Communications (eg, communication networks) are interconnected.
  • Examples of communication networks include local area networks ("Local Area Network, LAN”), wide area networks ("Wide Area Network, WAN”), Internet networks (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), and any currently known or future developed networks.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device, or may exist independently without being incorporated into the electronic device.
  • the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by the electronic device, the electronic device: acquires a phoneme sequence corresponding to the text to be synthesized; according to the phoneme sequence and the to-be-synthesized Synthesizing the text, generating TOBI representation sequences and prosodic acoustic features corresponding to the phoneme level of the text to be synthesized, and generating acoustic feature information corresponding to the text to be synthesized according to the TOBI representation sequence and the prosodic acoustic features; The acoustic feature information is used to generate first audio information corresponding to the text to be synthesized.
  • Computer program code for carrying out operations of the present disclosure may be written in one or more programming languages, or combinations thereof, including but not limited to object-oriented programming languages—such as Java, Smalltalk, C++, and Includes conventional procedural programming languages - such as "C" or similar programming languages.
  • the program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
  • the remote computer may be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (for example, using an Internet service provider to connected via the Internet).
  • LAN local area network
  • WAN wide area network
  • Internet service provider for example, using an Internet service provider to connected via the Internet.
  • each block in a flowchart or block diagram may represent a module, program segment, or portion of code that contains one or more logical functions for implementing specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented by a dedicated hardware-based system that performs the specified functions or operations , or may be implemented by a combination of dedicated hardware and computer instructions.
  • the modules involved in the embodiments described in the present disclosure may be implemented by software or by hardware.
  • the name of the module does not constitute the qualification of the module itself in some cases, for example, the acquisition module It can also be described as "a module for obtaining the phoneme sequence corresponding to the text to be synthesized".
  • exemplary types of hardware logic components include: Field Programmable Gate Arrays (Field Programmable Gate Arrays, FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (Application Specific Standard Parts, ASSP), System On Chip (System On Chip, SOC), Complex Programmable Logic Device (Complex Programmable Logic Device, CPLD) and so on.
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.
  • machine-readable storage media would include one or more wire-based electrical connections, portable computer discs, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.
  • RAM random access memory
  • ROM read only memory
  • EPROM or flash memory erasable programmable read only memory
  • CD-ROM compact disk read only memory
  • magnetic storage or any suitable combination of the foregoing.
  • Example 1 provides a speech synthesis method, including: obtaining a phoneme sequence corresponding to a text to be synthesized; generating the text to be synthesized according to the phoneme sequence and the text to be synthesized The corresponding phoneme-level TOBI representation sequence and prosodic acoustic feature, and according to the TOBI representation sequence and the prosodic acoustic feature, generate the acoustic feature information corresponding to the text to be synthesized; according to the acoustic feature information, generate the to-be-synthesized text First audio information corresponding to the synthesized text.
  • Example 2 provides the method of Example 1, generating the phoneme-level TOBI representation sequence and prosody corresponding to the text to be synthesized according to the phoneme sequence and the text to be synthesized Acoustic features, and according to the TOBI characterization sequence and the prosodic acoustic features, generate the acoustic feature information corresponding to the text to be synthesized, including: input the phoneme sequence and the text to be synthesized into the pre-trained speech synthesis
  • the TOBI representation sequence and prosodic acoustic features of the phoneme level corresponding to the text to be synthesized are generated by the speech synthesis model, and according to the TOBI representation sequence and the text to be synthesized
  • the prosodic acoustic features are used to generate acoustic feature information corresponding to the text to be synthesized.
  • Example 3 provides the method of Example 2, and the speech synthesis model includes an encoding network, an attention network, a decoding network, a prosodic language feature prediction module, a prosodic acoustic feature prediction module, and an embedding layer , a first splicing module, a second splicing module, and a third splicing module; wherein, the prosodic language feature prediction module is used to generate a phoneme-level TOBI representation sequence corresponding to the text to be synthesized according to the text to be synthesized; The embedding layer is used to generate a phoneme representation sequence corresponding to the text to be synthesized according to the phoneme sequence; the first splicing module is used to perform the TOBI representation sequence of the phoneme level with the phoneme representation sequence Splicing to obtain the first splicing sequence; the encoding network is used to encode the first splicing sequence to generate a coding sequence; the second splicing module is used to combine
  • Example 4 provides the method of Example 3, the prosodic language feature prediction
  • the module includes a first sub-embedding layer, a prosodic language feature prediction network, a second sub-embedding layer and an extension layer connected in sequence; wherein, the first sub-embedding layer is used to extract the deep layer of the word level corresponding to the text to be synthesized Representation; the prosodic language feature prediction network is used to generate TOBI tags at the word level according to the deep representation; the second sub-embedding layer is used to generate words corresponding to the text to be synthesized according to the TOBI tags Level TOBI representation sequence; the extension layer is used to expand the word-level TOBI representation sequence to obtain the phoneme-level TOBI representation sequence corresponding to the text to be synthesized.
  • Example 5 provides the method of Example 4, and the speech synthesis model is obtained by training in the following manner: obtaining training text; determining the training phoneme sequence corresponding to the training text, and training at the word level TOBI label, training prosodic acoustic features and training acoustic feature information; by using the training text as the input of the first sub-embedding layer, the output of the first sub-embedding layer is used as the input of the prosodic language feature prediction network , use the word-level training TOBI label as the target output of the prosodic language feature prediction network, use the output of the prosodic language feature prediction network as the input of the second sub-embedding layer, and use the second sub-embedding
  • the output of the layer is used as the input of the expansion layer
  • the training phoneme sequence is used as the input of the embedding layer
  • the output of the expansion layer and the output of the embedding layer are used as the input of the first splicing module
  • the output of the layer is used as the input
  • Example 6 provides the method of any one of Examples 1-5, the prosodic acoustic features include the fundamental frequency, energy and pronunciation duration of the phoneme level corresponding to the text to be synthesized at least one of .
  • Example 7 provides the method of any one of Examples 1-5, the method further comprising: synthesizing the first audio information with the target background music to obtain the second audio information.
  • Example 8 provides a speech synthesis device, including: an acquisition module, configured to acquire a phoneme sequence corresponding to text to be synthesized; a first generation module, configured to acquire According to the phoneme sequence and the text to be synthesized, generate the phoneme-level TOBI representation sequence and prosodic acoustic features corresponding to the text to be synthesized, and generate the to-be-synthesized text according to the TOBI representation sequence and the prosodic acoustic features. Acoustic feature information corresponding to the synthesized text; a second generating module configured to generate first audio information corresponding to the text to be synthesized according to the acoustic feature information generated by the first generating module.
  • Example 9 provides the device of Example 8, the first generation module is used to input the phoneme sequence and the text to be synthesized into a pre-trained speech synthesis model, Generate a phoneme-level TOBI characterization sequence and prosodic acoustic features corresponding to the text to be synthesized based on the phoneme sequence and the text to be synthesized through the speech synthesis model, and generate feature, generating acoustic feature information corresponding to the text to be synthesized.
  • Example 10 provides the device of Example 9, the speech synthesis model includes an encoding network, an attention network, a decoding network, a prosodic language feature prediction module, a prosodic acoustic feature prediction module, an embedding layer , a first splicing module, a second splicing module, and a third splicing module; wherein, the prosodic language feature prediction module is configured to generate a phoneme-level TOBI representation sequence corresponding to the text to be synthesized according to the text to be synthesized; The embedding layer, configured to generate a phoneme representation sequence corresponding to the text to be synthesized according to the phoneme sequence; the first splicing module is configured to splice the phoneme-level TOBI representation sequence with the phoneme representation sequence to obtain The first splicing sequence; the coding network is used to encode the first splicing sequence to generate a coding sequence; the second splicing module is used to perform the coding sequence
  • Example 11 provides the apparatus of Example 10, the prosodic language feature prediction module includes a first sub-embedding layer, a prosodic language feature prediction network, a second sub-embedding layer, and an extended layer; wherein, the first sub-embedding layer is used to extract the deep representation of the word level corresponding to the text to be synthesized; the prosodic language feature prediction network is used to generate the TOBI label of the word level according to the deep representation ; The second sub-embedding layer is used to generate the TOBI characterization sequence corresponding to the word level of the text to be synthesized according to the TOBI tag; the extension layer is used to expand the TOBI characterization sequence of the word level , to obtain the phoneme-level TOBI representation sequence corresponding to the text to be synthesized.
  • the first sub-embedding layer is used to extract the deep representation of the word level corresponding to the text to be synthesized
  • the prosodic language feature prediction network is used to generate the TOBI label of the word level according to the deep
  • Example 12 provides the device of Example 11, wherein the speech synthesis model is obtained through training with a model training device, wherein the model training device includes: a training text acquisition module, configured to acquire training text Determination module, for determining the training phoneme sequence corresponding to the training text, the training TOBI label of word level, training prosodic acoustic features and training acoustic feature information; training module, for using the training text as the first The input of the sub-embedding layer, the output of the first sub-embedding layer is used as the input of the prosodic language feature prediction network, the training TOBI label of the word level is used as the target output of the prosodic language feature prediction network, and the The output of the prosodic language feature prediction network is used as the input of the second sub-embedding layer, the output of the second sub-embedding layer is used as the input of the expansion layer, and the training phoneme sequence is used as the input of the embedding layer , using the output
  • Example 13 provides the device in any one of Examples 8-12, the prosodic acoustic features include the fundamental frequency, energy and pronunciation duration of the phoneme level corresponding to the text to be synthesized at least one.
  • Example 14 provides the device in any one of Examples 8-12, the device further comprising: a synthesis module, configured to synthesize the first audio information with target background music, Get the second audio information.
  • a synthesis module configured to synthesize the first audio information with target background music, Get the second audio information.
  • Example 15 provides a computer-readable medium on which a computer program is stored, and when the computer program is executed by a processing device, the method described in any one of Examples 1-7 is implemented A step of.
  • Example 16 provides an electronic device, including: a storage device storing at least one computer program thereon; at least one processing device configured to execute the program in the storage device At least one computer program to implement the steps of any one of the methods in Examples 1-7.
  • Example 17 provides a computer program.
  • the computer program is executed by a processing device, the steps of any one of the methods described in Examples 1-7 are implemented.
  • Example 18 provides a computer program product, the computer program product includes a computer program, and when the computer program is executed by a processing device, any one of Examples 1-7 is implemented. method steps.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Signal Processing (AREA)
  • Machine Translation (AREA)

Abstract

一种语音合成方法、装置、计算机可读介质及电子设备。该方法包括：获取待合成文本对应的音素序列（S101）；根据音素序列和待合成文本，生成待合成文本对应的TOBI表征序列和韵律声学特征，根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息（S102）；根据声学特征信息，生成待合成文本对应的第一音频信息（S103）。该方法使得合成音频更加自然，具有抑扬顿挫的听感，符合说话者所表达的语意。

Description

语音合成方法、装置、计算机可读介质及电子设备
相关申请交叉引用
本申请要求于2022年02月25日提交中国专利局、申请号为202210179831.4、发明名称为“语音合成方法、装置、计算机可读介质及电子设备”的中国专利申请的优先权，其全部内容通过引用并入本文。
技术领域
本公开涉及语音合成技术领域，具体地，涉及一种语音合成方法、装置、计算机可读介质及电子设备。
背景技术
在语言学中，韵律指的是讲话的过程中非独立音段(元音和辅音)的成分，即音节或更大单位的性质。这些性质形成语调、声调、重读和节奏等语言功能。韵律可以反映出说话者或话语的多种特征：说话者的感情状态、话语的形式(陈述、疑问还是命令)、是否存在强调、对比、焦点，以及其他无法由语法和词汇表达来表征的语言元素，相同韵律事件的表现形式不同可传达丰富的语义及其情感变化。在语音合成等任务中，如何结合文本的韵律特征使得合成的音频更加自然顺畅，成为研究的重点。
发明内容
提供该部分内容以便以简要的形式介绍构思，这些构思将在后面的具体实施方式部分被详细描述。该部分内容并不旨在表示要求保护的技术方案的关键特征或必要特征，也不旨在用于限制所要求的保护的技术方案的范围。
第一方面，本公开提供一种语音合成方法，包括：
获取待合成文本对应的音素序列；
根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的音调和中断指数(Tones and Break Indices，TOBI)表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；
根据所述声学特征信息，生成所述待合成文本对应的第一音频信息。
第二方面，本公开提供一种语音合成装置，包括：
获取模块，用于获取待合成文本对应的音素序列；
第一生成模块，用于根据所述获取模块获取到的所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；
第二生成模块，用于根据所述第一生成模块生成的所述声学特征信息，生成所述待合成文本对应的第一音频信息。
第三方面，本公开提供一种计算机可读介质，其上存储有计算机程序，所述计算机程序被处理装置执行时实现本公开第一方面提供的所述方法的步骤。
第四方面，本公开提供一种电子设备，包括：
存储装置，其上存储有至少一个计算机程序；
至少一个处理装置，用于执行所述存储装置中的所述至少一个计算机程序，以实现本公开第一方面提供的所述方法的步骤。
第五方面，本公开提供一种计算机程序，所述计算机程序被处理装置执行时实现本公开第一方面提供的所述方法的步骤。
第六方面，本公开提供一种计算机程序产品，所述计算机程序产品包括计算机程序，所述计算机程序被处理装置执行时实现本公开第一方面提供的所述方法的步骤。
在上述技术方案中，在获取到待合成文本对应的音素序列后，根据该音素序列和待合成文本，生成待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息；最后，根据声学特征信息，生成待合成文本对应的第一音频信息。在语音合成时，同时参考了待合成文本对应的TOBI表征序列和韵律声学特征，即不但参考了待合成文本语言层次的韵律特征，还参考了待合成文本声学层次的韵律特征，考虑到了韵律在不同维度上的表现。其中，根据TOBI表征序列能够赋予不同语句合适的节奏、强调和语调特性，同时对应的韵律声学特征可显式体现对应韵律事件的具体声学体现，从而在提升合成音频的韵律自然度的同时控制音频的强度(即幅度)，比如在多个重读位置可分配不同的强度来实现语义表达的强调重点不同，或通过强度调节实现疑问句的语调变化从而传达不同的语义(情感)。由此，能够在相同的韵律语言表现下，使得不同的韵律声学特征体现不同的语义变化，进而使得合成音频更加自然，更具有抑扬顿挫的听感，更符合说话者所表达的语意。
本公开的其他特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
结合附图并参考以下具体实施方式，本公开各实施例的上述和其他特征、优点及方面将变得更加明显。贯穿附图中，相同或相似的附图标记表示相同或相似的元素。应当理解附图是示意性的，元件和元素不一定按照比例绘制。在附图中：
图1是根据一示例性实施例示出的一种语音合成方法的流程图。
图2是根据一示例性实施例示出的一种语音合成模型的结构示意图。
图3是根据一示例性实施例示出的一种韵律语言特征预测模块的结构示意图。
图4是根据一示例性实施例示出的一种语音合成模型的训练方法的流程图。
图5是根据另一示例性实施例示出的一种语音合成方法的流程图。
图6是根据一示例性实施例示出的一种语音合成装置的框图。
图7是根据一示例性实施例示出的一种电子设备的框图。
具体实施方式
正如背景技术中论述的那样，在语音合成等任务中，如何结合文本的韵律特征使得合成的音频更加自然顺畅，成为研究的重点。为了提升合成音频的自然度，现阶段的语音合成方法主要通过使用语言层次的韵律特征，即人工标注的TOBI(Tones and Break Indices)数据来实现合成音频的韵律控制，以提升语音合成的自然度，但合成音频的强度不可控。
鉴于此，本公开提供一种语音合成方法、装置、计算机可读介质及电子设备。
下面将参照附图更详细地描述本公开的实施例。虽然附图中显示了本公开的某些实施例，然而应当理解的是，本公开可以通过各种形式来实现，而且不应该被解释为限于这里阐述的实施例，相反提供这些实施例是为了更加透彻和完整地理解本公开。应当理解的是，本公开的附图及实施例仅用于示例性作用，并非用于限制本公开的保护范围。
应当理解，本公开的方法实施方式中记载的各个步骤可以按照不同的顺序执行，和/或并行执行。此外，方法实施方式可以包括附加的步骤和/或省略执行示出的步骤。本公开的范围在此方面不受限制。
本文使用的术语“包括”及其变形是开放性包括，即“包括但不限于”。术语“基于”是“至少部分地基于”。术语“一个实施例”表示“至少一个实施例”；术语“另一实施例”表示“至少一个另外的实施例”；术语“一些实施例”表示“至少一些实施例”。其他术语的相关定义将在下文描述中给出。
需要注意，本公开中提及的“第一”、“第二”等概念仅用于对不同的装置、模块或单元进行区分，并非用于限定这些装置、模块或单元所执行的功能的顺序或者相互依存关系。
需要注意，本公开中提及的“一个”、“多个”的修饰是示意性而非限制性的，本领域技术人员应当理解，除非在上下文另有明确指出，否则应该理解为“一个或多个”。
本公开实施方式中的多个装置之间所交互的消息或者信息的名称仅用于说明性的目的，而并不是用于对这些消息或信息的范围进行限制。
图1是根据一示例性实施例示出的一种语音合成方法的流程图。如图1所示，该方法包括S101～S103。
在S101中，获取待合成文本对应的音素序列。
在本公开中，上述待合成文本可以为中文、英文、日语等语种文本。另外，可以通过字素到音素(Grapheme-to-Phoneme，G2P)模型来获取待合成文本对应的音素序列。
示例地，G2P模型可以采用循环神经网络(Recurrent Neural Network，RNN)和长短期记忆网络(Long Short-Term Memory，LSTM)来实现从字素到音素的转化。
在S102中，根据音素序列和待合成文本，生成待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息。
在本公开中，TOBI表征序列用于体现待合成文本语言层次的韵律特征，即韵律语言特征，其是指在原始语言学上ToBI体系所定义的韵律语言现象，属于离散特征，具体可以包括声调、语调、音高重音以及韵律边界。
其中，声调是指声音的高低升降的变化。示例地，中文中有四个声调：阴平、阳平、上声和去声，英文包括重读、次重读和轻读，日文包括重读和轻读。
语调(intonation)，即说话的腔调，就是一句话里快慢轻重的配置和变化。一句话除了词 汇意义(lexical meaning)还有语调意义(intonation meaning)。语调意义就是说话人用语调所表示的态度或口气。一句话的词汇意义加上语调意义才算是完全的意义。同样的句子，语调不同，意思就会不同，有时甚至会相差千里。
音高重音(pitch accent)，用于描述重读音节的音高变化，能够控制被强调信息与重音节奏型语言的节奏，其作用域在主重音音节上，或同一词的主重音与主重音后一音节上。在本公开中，将仅对主重音音节进行音高重音控制，忽略其他如次重音、零重音上的冗余信息，以达到信息精简的效果。相应地，音高重音信息用于指示待合成文本中存在指定重音现象的音节位置，其中，指定重音现象可以包括高重音、低重音、升重音、低升重音和高降重音。
具体来说，高重音，音高目标在高，基频曲线(f0)呈高平状，听感为汉语阴平；低重音，音高目标在低，基频曲线呈低平状，听感为汉语上声前半部分；升重音，音高目标在高，基频曲线呈攀升趋势，听感为汉语阳平；低升重音，音高目标在低，若作用在单音节上，基频曲线呈下降趋势，末尾有略微抬升，若作用在双音节上，基频曲线在主重音呈下降趋势，主重音后一音节上呈攀升趋势，听感为汉语上声；高降重音，音高目标在高，基频曲线呈下降趋势，听感为汉语去声。
韵律边界用于指示在待合成文本时应该在哪些地方进行停顿。示例地，韵律边界分为“#1”、“#2”、“#3”和“#4”四个停顿等级，其停顿程度依次增大。其中，英文和日文没有明显的韵律层级，因此该处置为空。
而韵律声学特征(即声学层次的韵律特征)则是在广泛定义表示语音声学特性的衡量物理量，如音色、共振峰、基频或者共振峰强度等。其中，与语言学上ToBI体系所定义的韵律事件联系更为紧密的声学特征：时长、基频、能量，例如韵律语言特征“句调”的高升调可具体表现为一个语音片段中对应基频持续爬升到一个句子中的基频高点。因此，本公开中的韵律声学特征包括待合成文本对应的音素级别的基频、能量以及发音时长中的至少一者，其是连续性特征。
声学特征信息可以例如是梅尔频谱、谱包络等。
在S103中，根据声学特征信息，生成待合成文本对应的第一音频信息。
在本公开中，可以通过将声学特征信息输入至声码器中，以得到待合成文本对应的第一音频信息，其中，声码器可以例如是Wavenet声码器、Griffin-Lim声码器等。
在上述技术方案中，在获取到待合成文本对应的音素序列后，根据该音素序列和待合成文本，生成待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息；最后，根据声学特征信息，生成待合成文本对应的第一音频信息。在语音合成时，同时参考了待合成文本对应的TOBI表征序列和韵律声学特征，即不但参考了待合成文本语言层次的韵律特征，还参考了待合成文本声学层次的韵律特征，考虑到了韵律在不同维度上的表现。其中，根据TOBI表征序列能够赋予不同语句合适的节奏、强调和语调特性，同时对应的韵律声学特征可显式体现对应韵律事件的具体声学体现，从而在提升合成音频的韵律自然度的同时控制音频的强度(即幅度)，比如在多个重读位置可分配不同的强度来实现语义表达的强调重点不同，或通过强度调节实现疑问句的语调变化从而传达不同的语义(情感)。由此，能够在相同的韵律语言表现下，使得不同的韵律声学特征体现不同的语义变化，进而使得合成音频更加自然，更具有抑扬顿挫的听感，更符合说话者所表达的语意。
下面针对上述S102中的根据音素序列和待合成文本，生成待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息的具体实施方式进行详细说明。
具体来说，可以将音素序列和待合成文本输入到预先训练好的语音合成模型中，以通过语音合成模型根据音素序列和待合成文本，生成待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息。
如图2所示，上述语音合成模型包括编码网络、注意力网络、解码网络、韵律语言特征预测模块、韵律声学特征预测模块、嵌入层、第一拼接模块、第二拼接模块以及第三拼接模块，其中，韵律语言特征预测模块、第一拼接模块、编码网络、第二拼接模块、韵律声学特征预测模块、第三拼接模块、注意力网络、解码网络依次连接，并且，第一拼接模块还与嵌入层连接，第二拼接模块还与韵律声学特征预测模块连接，第三拼接模块还与编码网络连接。
具体来说，韵律语言特征预测模块，用于根据待合成文本，生成待合成文本对应的音素级别的TOBI表征序列。
嵌入层，用于根据音素序列，生成待合成文本对应的音素表征序列，其中，音素表征序列由待合成文本中各音素对应的词向量按照相应音素在待合成文本中的先后顺序排序而成，可以根据预先建立的音素与词向量的对应关系来确定该合成文本中各音素对应的词向量。
第一拼接模块，用于将音素级别的TOBI表征序列与音素表征序列进行拼接，得到第一拼接序列。
编码网络，用于对第一拼接序列进行编码，生成编码序列。
第二拼接模块，用于将编码序列与音素级别的TOBI表征序列进行拼接，得到第二拼接序列。
韵律声学特征预测模块，用于根据第二拼接序列，生成待合成文本对应的韵律声学特征。
示例地，韵律声学特征预测模块可以为卷积层+双向LSTM层+全连接层构成的浅层网络。
第三拼接模块，用于将编码序列和韵律声学特征进行拼接，得到第三拼接序列；
注意力网络，用于根据第三拼接序列，生成待合成文本对应的语义表征。示例地，注意力网络可以为位置敏感注意力(Locative Sensitive Attention)的注意力网络，也可以为基于高斯混合模型(Gaussian Mixture Model，GMM)的注意力网络，即GMM attention。
解码网络，用于根据语义表征，生成待合成文本对应的声学特征信息。
如图3所示，上述韵律语言特征预测模块包括依次连接的第一子嵌入层、韵律语言特征预测网络、第二子嵌入层以及扩展层。
具体来说，第一子嵌入层，用于提取待合成文本对应的词级别的深层表征，示例地，第一子嵌入层可以为基于蒸馏学习的TinyBert模型。
韵律语言特征预测网络，用于根据深层表征，生成词级别的TOBI标签。其中，TOBI标签可以包括声调、语调、音高重音和韵律边界。
示例地，韵律语言特征预测网络可以为卷积层+双向LSTM层+全连接层构成的浅层网络。
第二子嵌入层，用于根据TOBI标签，生成待合成文本对应的词级别的TOBI表征序列。
扩展层，用于对词级别的TOBI表征序列进行扩展，得到待合成文本对应的音素级别的TOBI表征序列。
具体来说，可以针对待合成文本中的每一词，将该词对应的词级别的TOBI表征复制L-1次，得到该词对应的音素级别的TOBI表征，其中，L为该词包含的音素数量。
示例地，待合成文本包括依次连接的词A和词B，其中，词A包括三个音素，词B包括4个音素，词A对应的词级别的TOBI表征为M，词B对应的词级别的TOBI表征为N，则词A对应的音素级别的TOBI表征为MMM，词B对应的TOBI表征为NNNN，待合成文本对应的音素级别的TOBI表征序列为MMMNNNN。
另外，上述语音合成模型可以通过图4中所示的S401～S403来训练得到。
在S401中，获取训练文本。
在S402中，确定训练文本对应的训练音素序列、词级别的训练TOBI标签、训练韵律声学特征以及训练声学特征信息。
在本公开中，训练文本可以是从真实存在的语音中提取出的文本，标注人员可以首先通过听训练文本对应的语音的方式来标注训练文本对应的词级别的TOBI(即词级别的训练TOBI标签)。
可以通过与上述S101中获取待合成文本对应的音素序列相同的方式来获取训练文本对应的训练音素序列。
另外，可以通过以下方式来确定训练文本对应的训练韵律声学特征：可以基于开源工具(如librosa或straight等)等从训练文本对应的真实语音中提取帧级别的基频与能量特征，然后，可以针对训练文本中的每一音素，将该音素对应的多个帧的基频的平均值作为该音素的基频，将该音素对应的多个帧的能量的平均值作为该音素的能量，即得到音素级别的基频和音素级别的能量；同时，基于强制对齐工具获取训练文本中各音素的发音时长。
此外，可以通过将训练文本输入至语音合成模型(例如，Tacotron模型、Deepvoice 3模型、Tacotron 2模型、Wavenet模型等)中，得到训练文本对应的训练声学特征信息，例如，梅尔频谱特征信息。
在S403中，通过将训练文本作为第一子嵌入层的输入，将第一子嵌入层的输出作为韵律语言特征预测网络的输入，将词级别的训练TOBI标签作为韵律语言特征预测网络的目标输出，将韵律语言特征预测网络的输出作为第二子嵌入层的输入，将第二子嵌入层的输出作为扩展层的输入，将训练音素序列作为嵌入层的输入，将扩展层的输出和嵌入层的输出作为第一拼接模块的输入，将第一拼接模块的输出作为编码网络的输入，将编码网络的输出和扩展层的输出作为第二拼接模块的输入，将第二拼接模块的输出作为韵律声学特征预测模块的输入，将训练韵律声学特征作为韵律声学特征预测模块的目标输出，将韵律声学特征预测模块的输出和编码网络的输出作为第三拼接模块的输入，将第三拼接模块的输出作为注意力网络的输入，将注意力网络的输出作为解码网络的输入，将训练声学特征信息作为解码网络的目标输出的方式进行模型训练，以得到语音合成模型。
在本公开中，语音合成模型训练时的损失函数为声学特征信息损失和韵律特征损失之和。其中，声学特征信息损失是解码网络预测的声学特征信息与训练声学特征信息之间的均方差；韵律特征损失包括韵律语言特征的预测损失和韵律声学特征的预测损失，其中，韵律语言特征的预测损失是韵律语言特征预测网络预测的词级别的TOBI与词级别的训练TOBI标签之间的交叉熵损失；韵律声学特征的预测损失是韵律声学特征预测模块预测的韵律声学特征与训练韵律声学特征之间的均方差。
另外，为了提升用户体验，在上述步骤103获得与待合成文本对应的第一音频信息后，还可以为该第一音频信息添加背景音乐，这样，用户根据背景音乐和第一音频信息，更容易理解相应的文本内容。具体来说，如图5所示，上述方法还可以包括以下S104。
在S104中，将第一音频信息与目标背景音乐进行合成，得到第二音频信息。
在一种实施方式中，上述目标背景音乐可以为预设音乐，即可以是用户设定的任一音乐，也可以是默认的音乐。
在另一种实施方式中，在将第一音频信息与目标背景音乐进行合成之前，可以先根据待合成文本的文本信息，确定该待合成文本对应的使用场景信息，其中，该使用场景信息包括但不限于新闻播报、军武介绍、童话故事、校园广播等；然后，根据该使用场景信息，确定与该使用场景信息相匹配的目标背景音乐。
在本公开中，上述文本信息可以为关键词，此时，可以通过对待合成文本进行关键字自动识别，以根据关键词智能地预判该待合成文本的使用场景信息。
在确定出待合成文本对应的使用场景信息后，可以根据该使用场景信息，利用预先存储的使用场景信息与背景音乐的对应关系，确定与该使用场景信息匹配的目标背景音乐。例如，使用场景信息为军武介绍，其对应的背景音乐可以为激昂的音乐；使用场景信息为童话故事，则其对应的背景音乐可以为轻快活泼的音乐。
图6是根据一示例性实施例示出的一种语音合成装置的框图。如图6所示，该装置600包括：
获取模块601，用于获取待合成文本对应的音素序列；
第一生成模块602，用于根据所述获取模块601获取到的所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；
第二生成模块603，用于根据所述第一生成模块602生成的所述声学特征信息，生成所述待合成文本对应的第一音频信息。
在上述技术方案中，在获取到待合成文本对应的音素序列后，根据该音素序列和待合成文本，生成待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据TOBI表征序列和韵律声学特征，生成待合成文本对应的声学特征信息；最后，根据声学特征信息，生成待合成文本对应的第一音频信息。在语音合成时，同时参考了待合成文本对应的TOBI表征序列和韵律声学特征，即不但参考了待合成文本语言层次的韵律特征，还参考了待合成文本声学层次的韵律特征，考虑到了韵律在不同维度上的表现。其中，根据TOBI表征序列能够赋予不同语句合适的节奏、强调和语调特性，同时对应的韵律声学特征可显式体现对应韵律事件的具体声学体现，从而在提升合成音频的韵律自然度的同时控制音频的强度(即幅度)，比如在多个重读位置可分配不同的强度来实现语义表达的强调重点不同，或通过强度调节实现疑问句的语调变化从而传达不同的语义(情感)。由此，能够在相同的韵律语言表现下，使得不同的韵律声学特征体现不同的语义变化，进而使得合成音频更加自然，更具有抑扬顿挫的听感，更符合说话者所表达的语意。
可选地，所述第一生成模块602用于将所述音素序列和所述待合成文本输入到预先训练好的语音合成模型中，以通过所述语音合成模型根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序 列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息。
可选地，所述语音合成模型包括编码网络、注意力网络、解码网络、韵律语言特征预测模块、韵律声学特征预测模块、嵌入层、第一拼接模块、第二拼接模块以及第三拼接模块；
其中，所述韵律语言特征预测模块，用于根据所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列；
所述嵌入层，用于根据所述音素序列，生成所述待合成文本对应的音素表征序列；
所述第一拼接模块，用于将所述音素级别的TOBI表征序列与所述音素表征序列进行拼接，得到第一拼接序列；
所述编码网络，用于对所述第一拼接序列进行编码，生成编码序列；
所述第二拼接模块，用于将所述编码序列与所述音素级别的TOBI表征序列进行拼接，得到第二拼接序列；
所述韵律声学特征预测模块，用于根据所述第二拼接序列，生成所述待合成文本对应的韵律声学特征；
所述第三拼接模块，用于将所述编码序列和所述韵律声学特征进行拼接，得到第三拼接序列；
所述注意力网络，用于根据所述第三拼接序列，生成所述待合成文本对应的语义表征；
所述解码网络，用于根据所述语义表征，生成所述待合成文本对应的声学特征信息。
可选地，所述韵律语言特征预测模块包括依次连接的第一子嵌入层、韵律语言特征预测网络、第二子嵌入层以及扩展层；
其中，所述第一子嵌入层，用于提取所述待合成文本对应的词级别的深层表征；
所述韵律语言特征预测网络，用于根据所述深层表征，生成词级别的TOBI标签；
所述第二子嵌入层，用于根据所述TOBI标签，生成所述待合成文本对应的词级别的TOBI表征序列；
所述扩展层，用于对所述词级别的TOBI表征序列进行扩展，得到所述待合成文本对应的音素级别的TOBI表征序列。
可选地，所述语音合成模型通过模型训练装置训练得到，其中，该模型训练装置包括：
训练文本获取模块，用于获取训练文本；
确定模块，用于确定所述训练文本对应的训练音素序列、词级别的训练TOBI标签、训练韵律声学特征以及训练声学特征信息；
训练模块，用于通过将所述训练文本作为所述第一子嵌入层的输入，将所述第一子嵌入层的输出作为所述韵律语言特征预测网络的输入，将所述词级别的训练TOBI标签作为所述韵律语言特征预测网络的目标输出，将所述韵律语言特征预测网络的输出作为所述第二子嵌入层的输入，将所述第二子嵌入层的输出作为所述扩展层的输入，将所述训练音素序列作为所述嵌入层的输入，将所述扩展层的输出和所述嵌入层的输出作为所述第一拼接模块的输入，将所述第一拼接模块的输出作为所述编码网络的输入，将所述编码网络的输出和所述扩展层的输出作为所述第二拼接模块的输入，将所述第二拼接模块的输出作为所述韵律声学特征预测模块的输入，将所述训练韵律声学特征作为所述韵律声学特征预测模块的目标输出，将所述韵律声学特征预测模块的输出和所述编码网络的输出作为所述第三拼接模块的输入，将所述第三拼接模块的输出作为所述注意力网络的输入，将所述注意力网络的输出作为所述解码 网络的输入，将所述训练声学特征信息作为所述解码网络的目标输出的方式进行模型训练，以得到所述语音合成模型。
可选地，所述韵律声学特征包括所述待合成文本对应的音素级别的基频、能量以及发音时长中的至少一者。
可选地，所述装置600还包括：
合成模块，用于将所述第一音频信息与目标背景音乐进行合成，得到第二音频信息。
需要说明的是，上述模型训练装置可以集成于上述语音合成装置600中，也可以独立于上述语音合成装置600，本公开不作具体限定。
本公开还提供一种计算机可读介质，其上存储有计算机程序，所述计算机程序被处理装置执行时实现本公开提供的上述语音合成方法的步骤。
下面参考图7，其示出了用来实现本公开实施例的电子设备(终端设备或服务器)700的结构示意图。本公开实施例中的终端设备可以包括但不限于诸如移动电话、笔记本电脑、数字广播接收器、个人数字助理(Personal Digital Assistant，PDA)、平板电脑(PAD)、便携式多媒体播放器(Portable multimedia player，PMP)、车载终端(例如车载导航终端)等等的移动终端以及诸如数字TV、台式计算机等等的固定终端。图7示出的电子设备仅仅是一个示例，不应对本公开实施例的功能和使用范围带来任何限制。
如图7所示，电子设备700可以包括处理装置(例如中央处理器、图形处理器等)701，其可以根据存储在只读存储器(Read-Only Memory，ROM)702中的程序或者从存储装置708加载到随机访问存储器(Random Access Memory，RAM)703中的程序而执行各种适当的动作和处理。在RAM 703中，还存储有电子设备700操作所需的各种程序和数据。处理装置701、ROM 702以及RAM 703通过总线704彼此相连。输入/输出(I/O)接口705也连接至总线704。
通常，以下装置可以连接至I/O接口705：包括例如触摸屏、触摸板、键盘、鼠标、摄像头、麦克风、加速度计、陀螺仪等的输入装置706；包括例如液晶显示器(Liquid Crystal Display，LCD)、扬声器、振动器等的输出装置707；包括例如磁带、硬盘等的存储装置708；以及通信装置709。通信装置709可以允许电子设备700与其他设备进行无线或有线通信以交换数据。虽然图7示出了具有各种装置的电子设备700，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在非暂态计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信装置709从网络上被下载和安装，或者从存储装置708被安装，或者从ROM 702被安装。在该计算机程序被处理装置701执行时，执行本公开实施例的方法中限定的上述功能。
需要说明的是，本公开上述的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(Erasable Programmable Read Only Memory，EPROM或闪存)、光纤、便携式紧凑磁盘只读 存储器(Compact Disk Read Only Memory，CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(射频)等等，或者上述的任意合适的组合。
在一些实施方式中，客户端、服务器可以利用诸如超文本传输协议(HyperText Transfer Protocol，HTTP)之类的任何当前已知或未来研发的网络协议进行通信，并且可以与任意形式或介质的数字数据通信(例如，通信网络)互连。通信网络的示例包括局域网(“Local Area Network，LAN”)，广域网(“Wide Area Network，WAN”)，网际网(例如，互联网)以及端对端网络(例如，ad hoc端对端网络)，以及任何当前已知或未来研发的网络。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。
上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备：获取待合成文本对应的音素序列；根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；根据所述声学特征信息，生成所述待合成文本对应的第一音频信息。
可以以一种或多种程序设计语言或其组合来编写用于执行本公开的操作的计算机程序代码，上述程序设计语言包括但不限于面向对象的程序设计语言—诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言——诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)——连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，该模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本公开实施例中所涉及到的模块可以通过软件的方式实现，也可以通过硬件的方式来实现。其中，模块的名称在某种情况下并不构成对该模块本身的限定，例如，获取模块 还可以被描述为“获取待合成文本对应的音素序列的模块”。
本文中以上描述的功能可以至少部分地由一个或多个硬件逻辑部件来执行。例如，非限制性地，可以使用的示范类型的硬件逻辑部件包括：现场可编程门阵列(Field Programmable Gate Array，FPGA)、专用集成电路(Application Specific Integrated Circuit，ASIC)、专用标准产品(Application Specific Standard Parts，ASSP)、片上系统(System On Chip，SOC)、复杂可编程逻辑设备(Complex Programmable Logic Device，CPLD)等等。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
根据本公开的一个或多个实施例，示例1提供了一种语音合成方法，包括：获取待合成文本对应的音素序列；根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；根据所述声学特征信息，生成所述待合成文本对应的第一音频信息。
根据本公开的一个或多个实施例，示例2提供了示例1的方法，所述根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息，包括：将所述音素序列和所述待合成文本输入到预先训练好的语音合成模型中，以通过所述语音合成模型根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息。
根据本公开的一个或多个实施例，示例3提供了示例2的方法，所述语音合成模型包括编码网络、注意力网络、解码网络、韵律语言特征预测模块、韵律声学特征预测模块、嵌入层、第一拼接模块、第二拼接模块以及第三拼接模块；其中，所述韵律语言特征预测模块，用于根据所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列；所述嵌入层，用于根据所述音素序列，生成所述待合成文本对应的音素表征序列；所述第一拼接模块，用于将所述音素级别的TOBI表征序列与所述音素表征序列进行拼接，得到第一拼接序列；所述编码网络，用于对所述第一拼接序列进行编码，生成编码序列；所述第二拼接模块，用于将所述编码序列与所述音素级别的TOBI表征序列进行拼接，得到第二拼接序列；所述韵律声学特征预测模块，用于根据所述第二拼接序列，生成所述待合成文本对应的韵律声学特征；所述第三拼接模块，用于将所述编码序列和所述韵律声学特征进行拼接，得到第三拼接序列；所述注意力网络，用于根据所述第三拼接序列，生成所述待合成文本对应的语义表征；所述解码网络，用于根据所述语义表征，生成所述待合成文本对应的声学特征信息。
根据本公开的一个或多个实施例，示例4提供了示例3的方法，所述韵律语言特征预测 模块包括依次连接的第一子嵌入层、韵律语言特征预测网络、第二子嵌入层以及扩展层；其中，所述第一子嵌入层，用于提取所述待合成文本对应的词级别的深层表征；所述韵律语言特征预测网络，用于根据所述深层表征，生成词级别的TOBI标签；所述第二子嵌入层，用于根据所述TOBI标签，生成所述待合成文本对应的词级别的TOBI表征序列；所述扩展层，用于对所述词级别的TOBI表征序列进行扩展，得到所述待合成文本对应的音素级别的TOBI表征序列。
根据本公开的一个或多个实施例，示例5提供了示例4的方法，所述语音合成模型通过如下方式训练得到：获取训练文本；确定所述训练文本对应的训练音素序列、词级别的训练TOBI标签、训练韵律声学特征以及训练声学特征信息；通过将所述训练文本作为所述第一子嵌入层的输入，将所述第一子嵌入层的输出作为所述韵律语言特征预测网络的输入，将所述词级别的训练TOBI标签作为所述韵律语言特征预测网络的目标输出，将所述韵律语言特征预测网络的输出作为所述第二子嵌入层的输入，将所述第二子嵌入层的输出作为所述扩展层的输入，将所述训练音素序列作为所述嵌入层的输入，将所述扩展层的输出和所述嵌入层的输出作为所述第一拼接模块的输入，将所述第一拼接模块的输出作为所述编码网络的输入，将所述编码网络的输出和所述扩展层的输出作为所述第二拼接模块的输入，将所述第二拼接模块的输出作为所述韵律声学特征预测模块的输入，将所述训练韵律声学特征作为所述韵律声学特征预测模块的目标输出，将所述韵律声学特征预测模块的输出和所述编码网络的输出作为所述第三拼接模块的输入，将所述第三拼接模块的输出作为所述注意力网络的输入，将所述注意力网络的输出作为所述解码网络的输入，将所述训练声学特征信息作为所述解码网络的目标输出的方式进行模型训练，以得到所述语音合成模型。
根据本公开的一个或多个实施例，示例6提供了示例1-5中任一项的方法，所述韵律声学特征包括所述待合成文本对应的音素级别的基频、能量以及发音时长中的至少一者。
根据本公开的一个或多个实施例，示例7提供了示例1-5中任一项的方法，所述方法还包括：将所述第一音频信息与目标背景音乐进行合成，得到第二音频信息。
根据本公开的一个或多个实施例，示例8提供了一种语音合成装置，包括：获取模块，用于获取待合成文本对应的音素序列；第一生成模块，用于根据所述获取模块获取到的所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；第二生成模块，用于根据所述第一生成模块生成的所述声学特征信息，生成所述待合成文本对应的第一音频信息。
根据本公开的一个或多个实施例，示例9提供了示例8的装置，所述第一生成模块用于将所述音素序列和所述待合成文本输入到预先训练好的语音合成模型中，以通过所述语音合成模型根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息。
根据本公开的一个或多个实施例，示例10提供了示例9的装置，所述语音合成模型包括编码网络、注意力网络、解码网络、韵律语言特征预测模块、韵律声学特征预测模块、嵌入层、第一拼接模块、第二拼接模块以及第三拼接模块；其中，所述韵律语言特征预测模块，用于根据所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列；所述嵌入 层，用于根据所述音素序列，生成所述待合成文本对应的音素表征序列；所述第一拼接模块，用于将所述音素级别的TOBI表征序列与所述音素表征序列进行拼接，得到第一拼接序列；所述编码网络，用于对所述第一拼接序列进行编码，生成编码序列；所述第二拼接模块，用于将所述编码序列与所述音素级别的TOBI表征序列进行拼接，得到第二拼接序列；所述韵律声学特征预测模块，用于根据所述第二拼接序列，生成所述待合成文本对应的韵律声学特征；所述第三拼接模块，用于将所述编码序列和所述韵律声学特征进行拼接，得到第三拼接序列；所述注意力网络，用于根据所述第三拼接序列，生成所述待合成文本对应的语义表征；所述解码网络，用于根据所述语义表征，生成所述待合成文本对应的声学特征信息。
根据本公开的一个或多个实施例，示例11提供了示例10的装置，所述韵律语言特征预测模块包括依次连接的第一子嵌入层、韵律语言特征预测网络、第二子嵌入层以及扩展层；其中，所述第一子嵌入层，用于提取所述待合成文本对应的词级别的深层表征；所述韵律语言特征预测网络，用于根据所述深层表征，生成词级别的TOBI标签；所述第二子嵌入层，用于根据所述TOBI标签，生成所述待合成文本对应的词级别的TOBI表征序列；所述扩展层，用于对所述词级别的TOBI表征序列进行扩展，得到所述待合成文本对应的音素级别的TOBI表征序列。
根据本公开的一个或多个实施例，示例12提供了示例11的装置，所述语音合成模型通过模型训练装置训练得到，其中，该模型训练装置包括：训练文本获取模块，用于获取训练文本；确定模块，用于确定所述训练文本对应的训练音素序列、词级别的训练TOBI标签、训练韵律声学特征以及训练声学特征信息；训练模块，用于通过将所述训练文本作为所述第一子嵌入层的输入，将所述第一子嵌入层的输出作为所述韵律语言特征预测网络的输入，将所述词级别的训练TOBI标签作为所述韵律语言特征预测网络的目标输出，将所述韵律语言特征预测网络的输出作为所述第二子嵌入层的输入，将所述第二子嵌入层的输出作为所述扩展层的输入，将所述训练音素序列作为所述嵌入层的输入，将所述扩展层的输出和所述嵌入层的输出作为所述第一拼接模块的输入，将所述第一拼接模块的输出作为所述编码网络的输入，将所述编码网络的输出和所述扩展层的输出作为所述第二拼接模块的输入，将所述第二拼接模块的输出作为所述韵律声学特征预测模块的输入，将所述训练韵律声学特征作为所述韵律声学特征预测模块的目标输出，将所述韵律声学特征预测模块的输出和所述编码网络的输出作为所述第三拼接模块的输入，将所述第三拼接模块的输出作为所述注意力网络的输入，将所述注意力网络的输出作为所述解码网络的输入，将所述训练声学特征信息作为所述解码网络的目标输出的方式进行模型训练，以得到所述语音合成模型。
根据本公开的一个或多个实施例，示例13提供了示例8-12中任一的装置，所述韵律声学特征包括所述待合成文本对应的音素级别的基频、能量以及发音时长中的至少一者。
根据本公开的一个或多个实施例，示例14提供了示例8-12中任一的装置，所述装置还包括：合成模块，用于将所述第一音频信息与目标背景音乐进行合成，得到第二音频信息。
根据本公开的一个或多个实施例，示例15提供了一种计算机可读介质，其上存储有计算机程序，所述计算机程序被处理装置执行时实现示例1-7中任一项所述方法的步骤。
根据本公开的一个或多个实施例，示例16提供了一种电子设备，包括：存储装置，其上存储有至少一个计算机程序；至少一个处理装置，用于执行所述存储装置中的所述至少一个计算机程序，以实现示例1-7中任一项所述方法的步骤。
根据本公开的一个或多个实施例，示例17提供了一种计算机程序，所述计算机程序被处理装置执行时实现示例1-7中任一项所述方法的步骤。
根据本公开的一个或多个实施例，示例18提供了一种计算机程序产品，所述计算机程序产品包括计算机程序，所述计算机程序被处理装置执行时实现示例1-7中任一项所述方法的步骤。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开中所涉及的公开范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离上述公开构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。
此外，虽然采用特定次序描绘了各操作，但是这不应当理解为要求这些操作以所示出的特定次序或以顺序次序执行来执行。在一定环境下，多任务和并行处理可能是有利的。同样地，虽然在上面论述中包含了若干具体实现细节，但是这些不应当被解释为对本公开的范围的限制。在单独的实施例的上下文中描述的某些特征还可以组合地实现在单个实施例中。相反地，在单个实施例的上下文中描述的各种特征也可以单独地或以任何合适的子组合的方式实现在多个实施例中。
尽管已经采用特定于结构特征和/或方法逻辑动作的语言描述了本主题，但是应当理解所附权利要求书中所限定的主题未必局限于上面描述的特定特征或动作。相反，上面所描述的特定特征和动作仅仅是实现权利要求书的示例形式。关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。

Claims (12)

  1. 一种语音合成方法，包括：
    获取待合成文本对应的音素序列；
    根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的音调和中断指数TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；
    根据所述声学特征信息，生成所述待合成文本对应的第一音频信息。
  2. 根据权利要求1所述的方法，其中，所述根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息，包括：
    将所述音素序列和所述待合成文本输入到预先训练好的语音合成模型中，以通过所述语音合成模型根据所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息。
  3. 根据权利要求2所述的方法，其中，所述语音合成模型包括编码网络、注意力网络、解码网络、韵律语言特征预测模块、韵律声学特征预测模块、嵌入层、第一拼接模块、第二拼接模块以及第三拼接模块；
    其中，所述韵律语言特征预测模块，用于根据所述待合成文本，生成所述待合成文本对应的音素级别的TOBI表征序列；
    所述嵌入层，用于根据所述音素序列，生成所述待合成文本对应的音素表征序列；
    所述第一拼接模块，用于将所述音素级别的TOBI表征序列与所述音素表征序列进行拼接，得到第一拼接序列；
    所述编码网络，用于对所述第一拼接序列进行编码，生成编码序列；
    所述第二拼接模块，用于将所述编码序列与所述音素级别的TOBI表征序列进行拼接，得到第二拼接序列；
    所述韵律声学特征预测模块，用于根据所述第二拼接序列，生成所述待合成文本对应的韵律声学特征；
    所述第三拼接模块，用于将所述编码序列和所述韵律声学特征进行拼接，得到第三拼接序列；
    所述注意力网络，用于根据所述第三拼接序列，生成所述待合成文本对应的语义表征；
    所述解码网络，用于根据所述语义表征，生成所述待合成文本对应的声学特征信息。
  4. 根据权利要求3所述的方法，其中，所述韵律语言特征预测模块包括依次连接的第一子嵌入层、韵律语言特征预测网络、第二子嵌入层以及扩展层；
    其中，所述第一子嵌入层，用于提取所述待合成文本对应的词级别的深层表征；
    所述韵律语言特征预测网络，用于根据所述深层表征，生成词级别的TOBI标签；
    所述第二子嵌入层，用于根据所述TOBI标签，生成所述待合成文本对应的词级别的TOBI表征序列；
    所述扩展层，用于对所述词级别的TOBI表征序列进行扩展，得到所述待合成文本对 应的音素级别的TOBI表征序列。
  5. 根据权利要求4所述的方法，其中，所述语音合成模型通过如下方式训练得到：
    获取训练文本；
    确定所述训练文本对应的训练音素序列、词级别的训练TOBI标签、训练韵律声学特征以及训练声学特征信息；
    通过将所述训练文本作为所述第一子嵌入层的输入，将所述第一子嵌入层的输出作为所述韵律语言特征预测网络的输入，将所述词级别的训练TOBI标签作为所述韵律语言特征预测网络的目标输出，将所述韵律语言特征预测网络的输出作为所述第二子嵌入层的输入，将所述第二子嵌入层的输出作为所述扩展层的输入，将所述训练音素序列作为所述嵌入层的输入，将所述扩展层的输出和所述嵌入层的输出作为所述第一拼接模块的输入，将所述第一拼接模块的输出作为所述编码网络的输入，将所述编码网络的输出和所述扩展层的输出作为所述第二拼接模块的输入，将所述第二拼接模块的输出作为所述韵律声学特征预测模块的输入，将所述训练韵律声学特征作为所述韵律声学特征预测模块的目标输出，将所述韵律声学特征预测模块的输出和所述编码网络的输出作为所述第三拼接模块的输入，将所述第三拼接模块的输出作为所述注意力网络的输入，将所述注意力网络的输出作为所述解码网络的输入，将所述训练声学特征信息作为所述解码网络的目标输出的方式进行模型训练，以得到所述语音合成模型。
  6. 根据权利要求1-5中任一项所述的方法，其中，所述韵律声学特征包括所述待合成文本对应的音素级别的基频、能量以及发音时长中的至少一者。
  7. 根据权利要求1-6中任一项所述的方法，其中，所述方法还包括：
    将所述第一音频信息与目标背景音乐进行合成，得到第二音频信息。
  8. 一种语音合成装置，包括：
    获取模块，用于获取待合成文本对应的音素序列；
    第一生成模块，用于根据所述获取模块获取到的所述音素序列和所述待合成文本，生成所述待合成文本对应的音素级别的音调和中断指数TOBI表征序列和韵律声学特征，并根据所述TOBI表征序列和所述韵律声学特征，生成所述待合成文本对应的声学特征信息；
    第二生成模块，用于根据所述第一生成模块生成的所述声学特征信息，生成所述待合成文本对应的第一音频信息。
  9. 一种计算机可读介质，其上存储有计算机程序，所述计算机程序被处理装置执行时实现权利要求1-7中任一项所述方法的步骤。
  10. 一种电子设备，包括：
    存储装置，其上存储有至少一个计算机程序；
    至少一个处理装置，用于执行所述存储装置中的所述至少一个计算机程序，以实现权利要求1-7中任一项所述方法的步骤。
  11. 一种计算机程序，所述计算机程序被处理装置执行时实现权利要求1-7中任一项所述方法的步骤。
  12. 一种计算机程序产品，所述计算机程序产品包括计算机程序，所述计算机程序被处理装置执行时实现权利要求1-7中任一项所述方法的步骤。
PCT/CN2023/077478 2022-02-25 2023-02-21 语音合成方法、装置、计算机可读介质及电子设备 WO2023160553A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210179831.4 2022-02-25
CN202210179831.4A CN114495902A (zh) 2022-02-25 2022-02-25 语音合成方法、装置、计算机可读介质及电子设备

Publications (1)

Publication Number Publication Date
WO2023160553A1 true WO2023160553A1 (zh) 2023-08-31

Family

ID=81483936

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/077478 WO2023160553A1 (zh) 2022-02-25 2023-02-21 语音合成方法、装置、计算机可读介质及电子设备

Country Status (2)

Country Link
CN (1) CN114495902A (zh)
WO (1) WO2023160553A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114495902A (zh) * 2022-02-25 2022-05-13 北京有竹居网络技术有限公司 语音合成方法、装置、计算机可读介质及电子设备
CN118057520A (zh) * 2022-11-18 2024-05-21 脸萌有限公司 音频创作方法、装置及电子设备
CN116403562B (zh) * 2023-04-11 2023-12-05 广州九四智能科技有限公司 一种基于语义信息自动预测停顿的语音合成方法、系统

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110534089A (zh) * 2019-07-10 2019-12-03 西安交通大学 一种基于音素和韵律结构的中文语音合成方法
CN110782870A (zh) * 2019-09-06 2020-02-11 腾讯科技（深圳）有限公司 语音合成方法、装置、电子设备及存储介质
CN111754976A (zh) * 2020-07-21 2020-10-09 中国科学院声学研究所 一种韵律控制语音合成方法、系统及电子装置
CN112365880A (zh) * 2020-11-05 2021-02-12 北京百度网讯科技有限公司 语音合成方法、装置、电子设备及存储介质
CN113327580A (zh) * 2021-06-01 2021-08-31 北京有竹居网络技术有限公司 语音合成方法、装置、可读介质及电子设备
CN113421550A (zh) * 2021-06-25 2021-09-21 北京有竹居网络技术有限公司 语音合成方法、装置、可读介质及电子设备
US20210350795A1 (en) * 2020-05-05 2021-11-11 Google Llc Speech Synthesis Prosody Using A BERT Model
CN114495902A (zh) * 2022-02-25 2022-05-13 北京有竹居网络技术有限公司 语音合成方法、装置、计算机可读介质及电子设备

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110534089A (zh) * 2019-07-10 2019-12-03 西安交通大学 一种基于音素和韵律结构的中文语音合成方法
CN110782870A (zh) * 2019-09-06 2020-02-11 腾讯科技（深圳）有限公司 语音合成方法、装置、电子设备及存储介质
US20210350795A1 (en) * 2020-05-05 2021-11-11 Google Llc Speech Synthesis Prosody Using A BERT Model
CN111754976A (zh) * 2020-07-21 2020-10-09 中国科学院声学研究所 一种韵律控制语音合成方法、系统及电子装置
CN112365880A (zh) * 2020-11-05 2021-02-12 北京百度网讯科技有限公司 语音合成方法、装置、电子设备及存储介质
CN113327580A (zh) * 2021-06-01 2021-08-31 北京有竹居网络技术有限公司 语音合成方法、装置、可读介质及电子设备
CN113421550A (zh) * 2021-06-25 2021-09-21 北京有竹居网络技术有限公司 语音合成方法、装置、可读介质及电子设备
CN114495902A (zh) * 2022-02-25 2022-05-13 北京有竹居网络技术有限公司 语音合成方法、装置、计算机可读介质及电子设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZOU YUXIANG, LIU SHICHAO, YIN XIANG, LIN HAOPENG, WANG CHUNFENG, ZHANG HAOYU, MA ZEJUN: "Fine-Grained Prosody Modeling in Neural Speech Synthesis Using ToBI Representation", INTERSPEECH 2021, ISCA, ISCA, 1 January 2021 (2021-01-01), ISCA, pages 3146 - 3150, XP093086777, DOI: 10.21437/Interspeech.2021-883 *

Also Published As

Publication number Publication date
CN114495902A (zh) 2022-05-13

Similar Documents

Publication Publication Date Title
CN111899719B (zh) 用于生成音频的方法、装置、设备和介质
CN111369971B (zh) 语音合成方法、装置、存储介质和电子设备
CN112309366B (zh) 语音合成方法、装置、存储介质及电子设备
CN111292720B (zh) 语音合成方法、装置、计算机可读介质及电子设备
WO2021218324A1 (zh) 歌曲合成方法、装置、可读介质及电子设备
CN112786006B (zh) 语音合成方法、合成模型训练方法、装置、介质及设备
WO2023160553A1 (zh) 语音合成方法、装置、计算机可读介质及电子设备
CN107516511B (zh) 意图识别和情绪的文本到语音学习系统
CN112786011B (zh) 语音合成方法、合成模型训练方法、装置、介质及设备
CN111369967B (zh) 基于虚拟人物的语音合成方法、装置、介质及设备
CN112331176B (zh) 语音合成方法、装置、存储介质及电子设备
WO2021189984A1 (zh) 语音合成方法、装置、设备及计算机可读存储介质
CN112927674B (zh) 语音风格的迁移方法、装置、可读介质和电子设备
CN112309367B (zh) 语音合成方法、装置、存储介质及电子设备
CN112786013B (zh) 基于唱本的语音合成方法、装置、可读介质和电子设备
CN111292719A (zh) 语音合成方法、装置、计算机可读介质及电子设备
CN112802446B (zh) 音频合成方法及装置、电子设备和计算机可读存储介质
CN113327580A (zh) 语音合成方法、装置、可读介质及电子设备
CN113421550A (zh) 语音合成方法、装置、可读介质及电子设备
CN114255738A (zh) 语音合成方法、装置、介质及电子设备
CN116189652A (zh) 语音合成方法、装置、可读介质及电子设备
KR20080049813A (ko) 스피치 다이얼로그 방법 및 장치
CN114242035A (zh) 语音合成方法、装置、介质以及电子设备
CN114464164A (zh) 语音合成方法、装置、可读介质及电子设备
CN114155829A (zh) 语音合成方法、装置、可读存储介质及电子设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23759178

Country of ref document: EP

Kind code of ref document: A1