Voice-based age prediction method and related device

Abstract

The invention discloses a voice-based age prediction method and a related device, and relates to the technical field of voice processing, and the method comprises the steps: obtaining the voice data of a to-be-predicted speaker, recognizing the voice data, obtaining a recognition text, determining the age prediction reference features of the to-be-predicted speaker according to the voice data and the recognition text, and obtaining the age prediction reference features of the to-be-predicted speaker. The age prediction reference features comprise acoustic features and character event relation features; and finally, based on the age prediction reference features of the to-be-predicted speaker, determining an age prediction result of the to-be-predicted speaker. In the scheme, the character event relation characteristics are used for representing the age range and the credibility of the to-be-predicted speaker determined on the basis of the character event related to the age of the to-be-predicted speaker and have relatively high reference, so that the character event relation characteristics are considered when age prediction is carried out on the basis of voice, and the accuracy of age prediction is improved. The accuracy of an age prediction result can be improved.

Description

Technical Field

[0001] The present application relates to the field of speech processing technology, and in particular to a speech-based age prediction method and related devices. Background Art

[0002] In some scenarios, it is necessary to predict the speaker's age based on the voice. For example, in a speech scenario, it is necessary to predict the speaker's age. For another example, with the development of artificial intelligence technology, intelligent voice interaction devices, such as speakers and storytelling machines, are becoming more and more popular. Users of intelligent voice interaction devices such as speakers or storytelling machines may be users of different ages (children, adults, the elderly, etc.). If the user's age can be distinguished based on the voice input by the user, and then the service content corresponding to the user's age can be returned in a targeted manner, it will undoubtedly improve the intelligence and fun of intelligent voice interaction devices.

[0003] Traditional speech-based age prediction methods only consider the common acoustic features of the speaker, such as the Mel-frequency cepstral coefficients. The common acoustic features of the speaker alone are not sufficient to distinguish the vocal characteristics and habits of speakers of different ages, which leads to inaccurate age prediction results.

[0004] Therefore, how to provide a speech-based age prediction method to improve the accuracy of age prediction results has become a technical problem that needs to be urgently solved by those skilled in the art. Summary of the invention

[0005] In view of the above problems, the present application provides a method and related device for age prediction based on speech, so as to achieve the purpose of improving the accuracy of age prediction results. The specific scheme is as follows: The first aspect of the present application provides a method for age prediction based on speech, comprising: Obtaining speech data of the speaker to be predicted; Recognizing the voice data to obtain recognized text; Determine, based on the speech data and the recognized text, age prediction reference features of the speaker to be predicted, the age prediction reference features comprising acoustic features and character event relationship features; the character event relationship features are used to characterize the age range and credibility of the speaker to be predicted determined based on character events related to the age of the speaker to be predicted; Based on the age prediction reference features of the speaker to be predicted, an age prediction result of the speaker to be predicted is determined.

[0006] In a possible implementation, the age prediction reference feature of the speaker to be predicted further includes: At least one of a speaking habit feature, a speaking topic feature, a title feature, a name feature, and a gender-biased vocabulary feature, wherein the speaking habit feature is used to characterize at least one of the catchphrases, modal particles, speech speed, and speaking intervals of the speaker to be predicted, the speaking topic feature is used to characterize a topic related to the speech content of the speaker to be predicted, the title feature is used to characterize the title of the speaker to be predicted by others, the name feature is used to characterize the name of the speaker to be predicted, and the gender-biased vocabulary feature is used to characterize vocabulary that tends to distinguish between genders.

[0007] In a possible implementation, the acoustic features include: At least one of a Mel-frequency cepstral coefficient feature, a voice intonation feature, and an audio strength feature.

[0008] In a possible implementation, the speaking habit features include: At least one of a catchphrase feature, an interjection feature, a speech speed feature, and a speaking interval feature.

[0009] In a possible implementation, the method of determining the speech topic feature includes: determining subject-related content of the identified text; The features of the subject-related content of the recognized text are determined as the speech subject features.

[0010] In a possible implementation, the gender-biased vocabulary feature is determined in the following manner: Based on a pre-built male-oriented vocabulary resource library and a pre-built female-oriented vocabulary resource library, determining each male-oriented vocabulary and each female-oriented vocabulary in the recognition text; Identify masculine and feminine lexical features; The male-biased lexical features and female-biased lexical features are concatenated to obtain the speaker's gender-biased lexical features.

[0011] In a possible implementation, the method for determining the character-event relationship feature includes: Get a pre-built library of person-event relationships; Performing named entity recognition and relationship extraction on the text to obtain entities in the recognized text; Matching the entity in the recognized text with the person-event relationship resource library to determine the target entity in the recognized text and the time corresponding to the target entity; Determine at least one age range and its credibility by using the target entity in the identified text and the time corresponding to the target entity; The at least one age range and its credibility are fused to obtain a final age range and a final credibility as the person-event relationship feature.

[0012] In a possible implementation, determining the age prediction result of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted includes: Determining the gender of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted; The age prediction reference features of the speaker to be predicted are input into an age prediction model corresponding to the gender of the speaker to be predicted to obtain an age prediction result of the speaker to be predicted.

[0013] A second aspect of the present application provides a speech-based age prediction device, comprising: An acquisition unit, used for acquiring speech data of a speaker to be predicted; A speech recognition unit, used to recognize the speech data to obtain a recognized text; an age prediction reference feature determination unit, configured to determine age prediction reference features of the speaker to be predicted based on the speech data and the recognition text, wherein the age prediction reference features include acoustic features and character event relationship features; the character event relationship features are used to characterize the age range and credibility of the speaker to be predicted determined based on character events related to the age of the speaker to be predicted; The age prediction result determining unit is used to determine the age prediction result of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted.

[0014] In a possible implementation, the age prediction reference feature of the speaker to be predicted further includes: At least one of a speaking habit feature, a speaking topic feature, a title feature, a name feature, and a gender-biased vocabulary feature, wherein the speaking habit feature is used to characterize at least one of the catchphrases, modal particles, speech speed, and speaking intervals of the speaker to be predicted, the speaking topic feature is used to characterize a topic related to the speech content of the speaker to be predicted, the title feature is used to characterize the title of the speaker to be predicted by others, the name feature is used to characterize the name of the speaker to be predicted, and the gender-biased vocabulary feature is used to characterize vocabulary that tends to distinguish between genders.

[0015] In a possible implementation, the acoustic features include: At least one of a Mel-frequency cepstral coefficient feature, a voice intonation feature, and an audio strength feature.

[0016] In a possible implementation, the speaking habit features include: At least one of a catchphrase feature, an interjection feature, a speech speed feature, and a speaking interval feature.

[0017] In a possible implementation, the age prediction reference feature determination unit is specifically used to: determining subject-related content of the identified text; The features of the subject-related content of the recognized text are determined as the speech subject features.

[0018] In a possible implementation, the age prediction reference feature determination unit is specifically used to: Based on a pre-built male-oriented vocabulary resource library and a pre-built female-oriented vocabulary resource library, determining each male-oriented vocabulary and each female-oriented vocabulary in the recognition text; Identify masculine and feminine lexical features; The male-biased lexical features and female-biased lexical features are concatenated to obtain the speaker's gender-biased lexical features.

[0019] In a possible implementation, the age prediction reference feature determination unit is specifically used to: Get a pre-built library of person-event relationships; Performing named entity recognition and relationship extraction on the text to obtain entities in the recognized text; Matching the entity in the recognized text with the person-event relationship resource library to determine the target entity in the recognized text and the time corresponding to the target entity; Determine at least one age range and its credibility by using the target entity in the identified text and the time corresponding to the target entity; The at least one age range and its credibility are fused to obtain a final age range and a final credibility as the character-event relationship feature.

[0020] In a possible implementation, the age prediction result determination unit is specifically used to: Determining the gender of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted; The age prediction reference feature of the speaker to be predicted is input into an age prediction model corresponding to the gender of the speaker to be predicted to obtain an age prediction result of the speaker to be predicted.

[0021] A third aspect of the present application provides a computer program product, comprising computer-readable instructions, which, when executed on an electronic device, enables the electronic device to implement the voice-based age prediction method of the first aspect or any implementation of the first aspect.

[0022] A fourth aspect of the present application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the voice-based age prediction method of the above-mentioned first aspect or any implementation manner of the first aspect.

[0023] A fifth aspect of the present application provides a computer-readable storage medium, which carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement the speech-based age prediction method of the above-mentioned first aspect or any implementation of the first aspect.

[0024] By means of the above technical scheme, the present application provides a speech-based age prediction method and related devices. After acquiring the speech data of the speaker to be predicted, the speech data is first recognized to obtain the recognition text, and then the age prediction reference features of the speaker to be predicted are determined based on the speech data and the recognition text. The age prediction reference features include acoustic features and character-event relationship features; finally, based on the age prediction reference features of the speaker to be predicted, the age prediction result of the speaker to be predicted is determined. In this scheme, since the character-event relationship features are used to characterize the age range and its credibility of the speaker to be predicted determined based on the character events related to the age of the speaker to be predicted, and have a high reference value, therefore, when performing age prediction based on speech, the character-event relationship features are taken into account, which can improve the accuracy of the age prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the originals and elements are not necessarily drawn to scale.

[0026] Figure 1 A flowchart of a method for predicting age based on speech provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a speech-based age prediction device provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method section of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0028] The embodiments of the present application are described below in conjunction with the accompanying drawings. Those skilled in the art will appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0029] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, which is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0030] Reference Figure 1 , Figure 1 A flowchart of a method for predicting age based on speech is provided in an embodiment of the present application, such as Figure 1 As shown, a speech-based age prediction method provided in an embodiment of the present application may include the following steps, which are described in detail below.

[0031] S101: Acquire speech data of a speaker to be predicted; In the present application, the speaker to be predicted can be an ordinary user of the intelligent voice interaction device, or a specific user in certain scenarios, such as a speaker, and the present application does not impose any limitation on this. The voice data of the speaker to be predicted can be voice data related to the speaker to be predicted collected in real time, or voice data related to the speaker to be predicted collected and stored in advance, and the present application does not impose any limitation on this. In addition, the voice data of the speaker to be predicted can also be extracted from a video of the speaker to be predicted. The voice data of the speaker to be predicted can include its own voice data, and can also include voice data of its interaction with a third person, and the present application does not impose any limitation on this.

[0032] S102: Recognize the voice data to obtain recognized text; In the present application, speech recognition technology can be used to perform speech recognition on the speech data of the predicted speaker to obtain the recognized text. Specifically, firstly, endpoint detection is performed on the speech data to obtain the starting point and the ending point of the valid speech segment, and the valid speech segment can be detected by existing methods, such as using the VAD (Voice Activity Detection) method; then, feature extraction is performed on the valid speech segment to obtain the acoustic features of the valid speech segment; then, the acoustic features of the valid speech segment and the pre-trained acoustic model are used to calculate the acoustic score, and then the posterior probability output by the acoustic model is used as the input of the decoding module to obtain the corresponding recognized text.

[0033] It should be noted that in the present application, the acoustic features of the effective speech segment can be MFCC (Mel Frequency Cepstrum Coefficient), and the acoustic model can be DNN (Deep Neural Networks), LSTM (Long Short-Term Memory), BiLSTM (Bidirectional Long Short-Term Memory), UBiLSTM (Universal BiLSTM), CNN (Convolutional Neural Networks) and other deep learning models. The above-mentioned acoustic model to decoding module belongs to speech recognition technology, which is relatively mature and will not be described in detail in this application.

[0034] Considering that most of the texts recognized by speech do not have punctuation marks, and the whole text only has one paragraph or one sentence, it is not easy to process. Therefore, in this application, relevant technologies can be used to add punctuation to the text recognized by speech to obtain the final recognized text. Among them, adding punctuation mainly adds sentence-level punctuation: "?", ".", "!", ";", which are used for subsequent sentence segmentation. Adding punctuation can be done in a model-based way. This technology is relatively mature. For example, iFlytek's hearing system has the function of adding punctuation, which will not be described in detail here.

[0035] S103: determining age prediction reference features of the speaker to be predicted based on the speech data and the recognized text, wherein the age prediction reference features include acoustic features and character event relationship features; the character event relationship features are used to characterize the age range and credibility of the speaker to be predicted determined based on character events related to the age of the speaker to be predicted; In the present application, the acoustic features of the speaker to be predicted include: at least one of MFCC (Mel Frequency Cepstrum Coefficient) features, intonation features, and audio strength features.

[0036] In the present application, the person-event relationship feature is obtained by performing semantic mining on the identified text to calculate the speaker's age range and credibility. The specific determination method will be described in detail in subsequent embodiments.

[0037] S104: Determine an age prediction result of the speaker to be predicted based on the age prediction reference features of the speaker to be predicted.

[0038] The present embodiment provides a method for age prediction based on speech. After acquiring the speech data of the speaker to be predicted, the speech data is first recognized to obtain the recognition text, and then the age prediction reference features of the speaker to be predicted are determined based on the speech data and the recognition text. The age prediction reference features include acoustic features and character event relationship features; finally, based on the age prediction reference features of the speaker to be predicted, the age prediction result of the speaker to be predicted is determined. In this solution, since the character event relationship features are used to characterize the age range and credibility of the speaker to be predicted determined based on character events related to the age of the speaker to be predicted, and have a high reference value, therefore, when performing age prediction based on speech, considering the character event relationship features can improve the accuracy of the age prediction results.

[0039] In the above embodiment, it is pointed out that the acoustic features include at least one of: Mel-frequency cepstral coefficient features, intonation features, and audio strength features. For ease of understanding, the present application provides a more detailed description, as follows: MFCC (Mel Frequency Cepstrum Coefficient) feature is one of the most commonly used acoustic features in speech recognition. It is proposed based on the research on the human hearing mechanism. The extraction of MFCC features mainly includes pre-filtering, A / D conversion, pre-emphasis, framing, windowing, fast Fourier transform, triangular window filtering, logarithm calculation and short-time energy. The specific process is not described in detail here.

[0040] Intonation features mainly refer to the changes in the speaker's intonation when speaking. Without considering the speaker's personal speaking characteristics, considering that the intonation of older speakers should be flat, the intonation can reflect some age characteristics to a certain extent. The intonation feature can be determined by adding windows to the speech data (the window length can be set based on the scene requirements), and for each speech data segment of the window length, the intonation feature corresponding to the speech data segment is determined respectively, and the intonation features corresponding to the continuous n speech data segments are spliced ​​to obtain the intonation feature of the speaker to be predicted. For each speech data segment of the window length, the speech data segment is framed (for example, one frame can be every 10ms), and for each frame, the fundamental frequency of the speech in the frame is extracted (the partial tone with the largest amplitude and the lowest frequency in a natural complex tone is the first harmonic, which is generally called the fundamental frequency, and the frequency of the fundamental frequency is called the fundamental frequency), and the variance of the fundamental frequency in the speech data segment is used as the intonation feature of the speech data segment.

[0041] The audio strength feature mainly refers to the volume of the speaker's voice when speaking, that is, the audio amplitude is also called the sound pressure level. The audio strength feature can be determined by windowing the voice data (the window length can be set based on the scene requirements), and for each voice data segment of the window length, the audio strength feature corresponding to the voice data segment is determined respectively, and the audio strength features corresponding to the consecutive n voice data segments are spliced ​​to obtain the audio strength feature of the speaker to be predicted. For each voice data segment of the window length, the average sound pressure level of the voice data segment can be measured as the audio strength feature corresponding to the voice data segment. The measurement method is an existing mature technology and is not described in detail in this application.

[0042] In order to further improve the accuracy of age prediction results, in a preferred implementation, the age prediction reference features of the speaker to be predicted also include at least one of speaking habit features, speaking topic features, title features, name features and gender-biased vocabulary features.

[0043] Taking into account that a catchphrase is a habitual expression of a person, its formation is often related to the social hot spots at the time, etc., and can make certain distinctions among age groups to a certain extent. People of different age groups will also have differences in modal particles, and people of different age groups will also have differences in speaking speed (for example, older people may speak slower, etc.). People of different age groups will also have different lengths of rest intervals during speaking. Therefore, the speaking habit features are used to characterize at least one of the catchphrases, modal particles, speaking speed and speaking intervals of the speaker to be predicted. In the present application, the speaking habit features of the speaker to be predicted include: at least one of catchphrase features, modal particle features, speaking speed features and speaking interval features.

[0044] Among them, in the present application, the catchphrase feature is used to characterize the catchphrases commonly used by the speaker to be predicted; the modal particle feature is used to characterize the modal particles commonly used by the speaker to be predicted; the speaking speed feature is used to characterize the total number of words spoken by the speaker to be predicted per unit time; the speaking interval feature is used to characterize the ratio of the rest time of the speaker to be predicted to the duration of the voice data, and the ratio of the number of rest times to the duration of the voice data.

[0045] Considering that people of different ages and genders may have different research directions, or different research fields have different age requirements, for example, some relatively new fields, such as "online games" may tend to be young, while fields such as "calligraphy" and "international relations" may tend to be old. Therefore, people of different ages will also have different speech content. Therefore, in this application, it is necessary to determine the speech topic feature. The speech topic feature is used to characterize the topic related to the speech content of the speaker to be predicted, Considering that some titles can reflect approximate age information, such as "professor", "master", etc., who are generally over 30 years old, the speaker's age can be judged based on the title of the speaker. In this application, the speaker's title feature can also be determined. The title feature is used to characterize the title of the speaker to be predicted by others. In a possible implementation, the title of the speaker can be obtained from the voice data of the speaker's interaction with a third person, and the vector of the speaker's title can be calculated as the speaker's title feature. If the speaker's title cannot be obtained, the speaker's title features can be set to zero.

[0046] Considering that there is a certain correlation between name and gender, the name feature of the speaker can be considered in this application, and the name feature is used to characterize the name of the speaker to be predicted. In a possible implementation, the name of the speaker can be obtained, and the vector of his name can be calculated as the name feature of the speaker. For example, a general speaker will have a self-introduction before starting a speech. The information extraction technology can be used to obtain the speaker's name, segment it, and then accumulate the corresponding word vector as the speaker's name feature. If the speaker's name cannot be obtained, the speaker's name features can be set to zero.

[0047] Considering that some words tend to be used by men, such as "brother", and some words tend to be used by women, such as honorifics, some words can be used to distinguish gender. Therefore, in this application, the gender-biased vocabulary features of the speaker can be obtained. The gender-biased vocabulary features are used to characterize words that tend to distinguish gender.

[0048] In the above embodiment, it is pointed out that the speaking habit feature includes at least one of: a catchphrase feature, an interjection feature, a speech speed feature, and a speaking interval feature. For ease of understanding, the determination method of the above-mentioned various speaking habit features is described in detail in the embodiment of the present application, as follows: In this application, the method for determining the catchphrase features of the speaker to be predicted includes: Determine the vector of each target phrase in the recognition text; and determine the catchphrase features of the speaker to be predicted based on the similarity between the vector of each target phrase and the resources in the pre-built catchphrase static resource library.

[0049] In the present application, the target phrase can be a phrase that appears frequently in the recognized text, for example, a phrase that appears more frequently than a certain threshold. In one possible implementation, information extraction technology can be used to obtain phrases that appear frequently in the recognized text as target phrases. The vector of each target phrase can be determined using a word segmentation tool (such as the NLPIR-ICTCLAS Chinese word segmentation system) and a vector representation tool (such as word2vec and other tools). Specifically, the target phrase is segmented and the word vectors of each word are accumulated to obtain its sentence vector as the vector of the target phrase. The weight here can be calculated by a keyword algorithm to obtain the importance of each word, such as TFIDF and other algorithms, which will not be described in detail here.

[0050] In this application, a static resource library of catchphrases can be pre-built. Specifically, some commonly used colloquialisms and Internet buzzwords can be obtained by crawlers from the Internet and offline manual collection. Then, word segmentation tools (such as NLPIR-ICTCLAS Chinese word segmentation system) and vector representation tools (such as word2vec and other tools) are used to determine the vectors of these colloquialisms and Internet buzzwords, and a static resource library of catchphrases is constructed. For example, the popular Internet phrase "My brother is awesome" has become a catchphrase for many young people. The sentence "My brother is awesome" can be segmented first, and the word vector corresponding to each word can be obtained through tools such as word2vec. The sentence vector of the sentence is weighted and summed. The weight here can be obtained by keyword algorithms to obtain the importance of each word, such as TFIDF and other algorithms, which will not be described in detail here. The sentence vector obtained is a resource in the static resource library of catchphrases.

[0051] Based on the similarity between the vector of each target phrase and the resources in the pre-built static resource library of catchphrases, the catchphrase features of the speaker to be predicted are determined. Specifically, the vector of each target phrase can be calculated with the resources in the pre-built static resource library of catchphrases for similarity, such as cosine distance, Euclidean distance or other vector distances, and the vector of the target phrase with the highest similarity is selected as the catchphrase feature of the speaker to be predicted.

[0052] In the present application, the method for determining the modal particle features of the speaker to be predicted includes: determining the vectors of each modal particle in the recognized text; and fusing the vectors of each modal particle to obtain the modal particle features of the speaker to be predicted.

[0053] In the present application, information extraction technology can be used to obtain each modal particle that appears in the identified text. The vector of each modal particle can be determined by using word segmentation and vector representation tools (such as word2vec and other tools). Specifically, the modal particle is segmented and the word vectors of each word are accumulated to obtain its sentence vector as the vector of the modal particle. The weight here can be obtained by keyword algorithms to obtain the importance of each word, such as TFIDF and other algorithms, which will not be described in detail here.

[0054] In the present application, the vectors of the various modal particles may be accumulated to obtain a modal particle center vector as the modal particle feature of the speaker to be predicted.

[0055] In the present application, the method for determining the speaking rate characteristics of the speaker to be predicted includes: windowing the speech data (the window length can be set based on the scenario requirements), determining the speaking rate characteristics corresponding to each speech data segment of each window length, and splicing the speaking rate characteristics corresponding to n consecutive speech data segments to obtain the speaking rate characteristics of the speaker to be predicted.

[0056] In the present application, the method for determining the speaking interval characteristics of the speaker to be predicted includes: windowing the speech data (the window length can be set based on the scenario requirements), determining the speaking interval characteristics corresponding to each speech data segment of each window length, and splicing the speaking interval characteristics corresponding to n consecutive speech data segments to obtain the speaking interval characteristics of the speaker to be predicted.

[0057] In another embodiment of the present application, a method for determining the speech topic feature is described, and the method may include the following steps: S201: Determine the subject-related content of the recognized text; In one possible implementation, the subject-related content of the recognized text may be extractive content, that is, a number of original sentences with high subject degrees in the text may be extracted as subject-related content, and the subject-related content of the recognized text may be extracted using text automatic summarization technology. In another possible implementation, the subject-related content of the recognized text may also be generative subject-related content, and a summary generation model (such as a large language model) may be used to generate the subject-related content of the recognized text. The text automatic summarization technology and summary generation model are existing mature technologies and are not described in detail in this application.

[0058] S202: Determine features of the subject-related content of the recognized text as the speech subject features.

[0059] In the present application, word segmentation technology can be used to obtain each word in each sentence in the topic-related content of the recognized text, accumulate the word vectors of each sentence to obtain a sentence vector, and accumulate the sentence vectors to obtain the features of the topic-related content of the recognized text as the speech topic features.

[0060] In another embodiment of the present application, a method for determining the gender-biased vocabulary feature is described, and the method may include the following steps: S301: determining each male-oriented word and each female-oriented word in the recognition text based on a pre-built male-oriented word resource library and a pre-built female-oriented word resource library; In the present application, a male-oriented vocabulary resource library and a female-oriented vocabulary resource library can be pre-constructed, wherein the male-oriented vocabulary resource library contains word vectors corresponding to each manually collected male-oriented vocabulary, and the female-oriented vocabulary resource library contains word vectors corresponding to each manually collected female-oriented vocabulary. The recognition text is segmented, and the vectors of each segmented word are determined. By calculating the similarity between the vectors of each segmented word and each vector in the male-oriented vocabulary resource library and the female-oriented vocabulary resource library, such as cosine distance, Euclidean distance or other vector distance, if the similarity corresponding to a segmented word is higher than a preset threshold value, the segmented word is considered to be a male-oriented word or a female-oriented word.

[0061] S302: Determine male-oriented vocabulary features and female-oriented vocabulary features; In the present application, the male-oriented vocabulary feature is used to indicate the weighted sum of the frequencies of each male-oriented vocabulary, and the female-oriented vocabulary feature is used to indicate the weighted sum of the frequencies of female-oriented vocabulary. The frequency of each male-oriented vocabulary is the ratio of the number of each male-oriented vocabulary in the full text of the recognition text to the total number of times, and the weight of each male-oriented vocabulary is the ratio of the male-oriented vocabulary to the total number of male-oriented vocabulary in the full text of the recognition text. The frequency of each female-oriented vocabulary is the ratio of the number of each female-oriented vocabulary in the full text of the recognition text to the total number of times, and the weight of each female-oriented vocabulary is the ratio of the female-oriented vocabulary to the total number of female-oriented vocabulary in the full text of the recognition text.

[0062] For ease of understanding, taking honorific words as an example, in this application, a respectful word resource library can be pre-constructed, and the respectful word resource library contains word vectors corresponding to each honorific word collected manually. The ratio of the number of all honorific words in the full text to the total number of words is counted as the frequency of the honorific words, and at the same time, the word vectors of all the honorific words are weighted and accumulated according to the ratio of each honorific word in the total honorific words. For example, the total number of words is 100, and the number of honorific words is judged to be 20, then the ratio of honorific words to the total number of words is 0.2. If there are four honorific words among the 20 honorific words, that is, each honorific word has 5 repetitions, then the proportion of each honorific word in the honorific word set is 0.25, then the word vectors of these four honorific words are weighted and accumulated, and the weights are all 0.25. The extracted honorific word feature is then (vector, 0.2).

[0063] S303: Concatenate the male-biased lexical features and the female-biased lexical features to obtain the speaker's gender-biased lexical features.

[0064] In another embodiment of the present application, a method for determining the character-event relationship feature is described, and the method may include the following steps: S401: Acquire a pre-built character-event relationship resource library; In this application, a person-event relationship resource library can be pre-built. The person-event relationship resource library mainly includes entities and their corresponding time. In this application, a large amount of resources can be obtained from the Internet through crawlers, and information can be extracted manually to extract entities and their corresponding time.

[0065] Entities include but are not limited to: "world celebrities", such as "Ma ×"; "famous events at home and abroad", such as "×× Olympic Games"; "relatively representative objects", such as "yo-yo"; "relationship entities", such as "classmate relationship", "blood relationship", "colleague relationship", "teacher-student relationship", etc. The time corresponding to the relationship entity should be a relative time relationship, for example, the time corresponding to "classmate relationship" should be "similar age", the time corresponding to "blood relationship" should be the age information of common blood relationships, for example, the time corresponding to "father-son relationship" should be "+-20"; the time corresponding to "teacher-student relationship" should be "the teacher's age is higher than the student's age"; the time corresponding to "colleague relationship" should be "similar age".

[0066] The above entities are selected because they can indirectly obtain the age information of the speaker to be predicted. For example, for the named entity "world celebrities", if the speaker said "Actually, Ma × and I were classmates before.", we can judge from this sentence that the speaker and Ma × are classmates. We believe that classmates should be of similar age, that is, through Ma ×'s age and "classmate relationship" we can get the approximate age of the speaker. Another example is the named entity "more famous events at home and abroad", if the speaker said: "I was still in high school during the ×× Olympics.", through this sentence we believe that the speaker was about 18 years old at the ×× Olympics, that is, in 2008, and we can infer that the speaker's date of birth is about 1990.

[0067] In this application, the format of the person-event relationship resource library can be in the form of map<vector, time>, where the vector is mainly obtained by weighted accumulation of the corresponding word vector after entity segmentation, and the weight varies according to its word attributes. For example, the weight of a noun may be higher, such as "×× Olympic Games Opening", which is segmented into "××", "Olympic Games", and "Opening". The weight of "Olympic Games" can be 0.5, and the weight of "××" as a place name can be 0.3, "Opening" 0.2, etc. The specific weight distribution can be adjusted. The word segmentation and word attribute annotation used in it belong to the existing technology, such as the domestic "NLPIR-ICTCLAS Chinese Word Segmentation System", etc., and the word attributes can also be marked while segmenting, which will not be described in detail in this application.

[0068] S402: performing named entity recognition and relationship extraction on the text to obtain entities in the recognized text; In the present application, the recognized text can be divided into sentences and each sentence can be subjected to named entity recognition to obtain the named entities therein, and the named entities mainly include: "person", "thing", and "event"; for the named entity "person", its relationship entities need to be further determined, and the relationship entities mainly include age-related relationships, including but not limited to "classmate relationship", "blood relationship", "teacher-student relationship", "colleague relationship", etc.

[0069] S403: Matching the entity in the recognized text with the person-event relationship resource library to determine the target entity in the recognized text and the time corresponding to the target entity; In the present application, the target entity in the recognized text is a named entity in the recognized text that matches the person-event relationship resource library. In the present application, when matching the entity in the recognized text with the person-event relationship resource library, the vector of the entity in the recognized text can be obtained first, and then the vector of the entity in the recognized text is matched with the person-event relationship resource library. The matching rule is that if the nearest distance between the vector to be searched and a vector in the person-event relationship resource library is higher than a preset threshold value, it is considered to be matched, and the corresponding entity is determined to be the target entity. If the nearest distance between the vector to be searched and each vector in the person-event relationship resource library is lower than a preset threshold value, it is considered to be unmatched, and the corresponding entity is determined to be a non-target entity. Among them, the vector to be searched is the vector of the entity in the recognized text.

[0070] S404: Determine at least one age range and its credibility by using the target entity in the recognized text and the time corresponding to the target entity; In this application, if the target entity is "person" and its relationship entity, it is necessary to jointly determine an age range based on the time corresponding to the named entity "person" and the time corresponding to its relationship entity. The credibility of the age range can be determined by the word vector distance in the matching process, the credibility of the relationship entity, and the credibility of the age determined by the relationship entity.

[0071] In this application, if the target entity is an "object", an age range can be determined based on the time corresponding to the "object", and the credibility of the age range can be determined by the word vector distance in the matching process, such as 0.8. For ease of understanding, assume that the recognition text contains a sentence "I played yo-yo when I was a child", where "yo-yo" is the target entity and its corresponding time is "1995-2010". Combined with semantic analysis, it can be determined that the speaker was about 5-15 years old from 1995 to 2010, and thus the speaker's date of birth can be inferred to be about 1980-2005.

[0072] In this application, if the target entity is an "event", an age range can be determined based on the time corresponding to the "event", and the credibility of the age range can be determined by the word vector distance in the matching process, such as 0.7. For ease of understanding, assume that the recognition text contains a sentence "I was still in high school when the ×× event occurred", where "×× event" is the target entity and its corresponding time is "2001". Combined with semantic analysis, it can be determined that the speaker's age range in 2001 should be "15-20 years old", from which it can be inferred that the speaker's date of birth is approximately 1986-1991.

[0073] S405: Fusing the at least one age range and its credibility to obtain a final age range and a final credibility as the person-event relationship feature.

[0074] In the present application, the obtained age intervals and their credibility can be spliced ​​according to the interval size to obtain the final age interval and the final credibility. In a possible implementation, all intervals with credibility higher than the threshold can be found and solved as the maximum interval. In the case of an intersection, the credibility of each interval can be summed and averaged to obtain a new credibility. For example, the time range (1980-2005, 0.8) and (1986-1991, 0.7) obtained in the above steps can be finally represented as {(1980-1986, 0.75), (1986-1991, 0.7), (1991-2005, 0.8)}.

[0075] In another embodiment of the present application, a specific implementation method of determining the age prediction result of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted is described, and the method may include the following steps: S501: Determine the gender of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted; In the present application, the age prediction reference features of the speaker to be predicted can be input into the gender judgment model to obtain the gender of the speaker to be predicted. The structure of the gender judgment model can choose the traditional GMM modeling method or the CNN modeling method. The specific case is not specified. The output of the model is a regularized two-dimensional result, that is, the score for males and the score for females, such as (0.6, 0.2).

[0076] The training process is the same as the traditional training process, that is, the data is divided into labeled and test data. The labeling is done manually to classify the audio into male and female, and extract the relevant age prediction reference features as described above. I will not go into details here.

[0077] S502: Inputting the age prediction reference features of the speaker to be predicted into an age prediction model corresponding to the gender of the speaker to be predicted to obtain an age prediction result of the speaker to be predicted.

[0078] In the present application, the various features in the age prediction reference features of the speaker to be predicted can be spliced ​​and input into the age prediction model corresponding to the gender of the speaker to be predicted, and finally the interval with the largest posterior score is selected as the age prediction result of the speaker to be predicted according to the output result of the model.

[0079] In this embodiment, age prediction models of different genders need to be pre-trained. The age prediction model of each gender can be obtained by training a large amount of speech data of speakers of different age groups of the gender.

[0080] The specific training methods are as follows: First, collect a large amount of speech data from speakers of different genders and ages. For each gender, divide the age group into +-n years, with a total of m intervals, where n can be 5 and m can be 9. For example, take those under 20 as the first interval, 20-30 as the second interval, and so on. Divide all speakers into corresponding intervals manually, and use numbers to represent them, such as 1 for under 20, 2 for 20-30, 3 for 30-40...9 for 90-100, etc. The n and m values ​​here are adjustable.

[0081] Determine the age prediction reference features corresponding to each piece of speech data. The specific method can refer to the method for determining the age prediction reference features of the speaker to be predicted, which will not be described in detail here. It should be noted that the person-event relationship features can be manually annotated during training, that is, the speaker age range and its credibility can be obtained by artificial semantic judgment.

[0082] Finally, the age prediction reference features corresponding to the speech data are used as the input of the age prediction model. The model outputs m posterior values, which represent the probability that the speaker is in each age range. Finally, the age range with the largest posterior value is selected, and its posterior value is its credibility. It should be noted that when the age prediction reference features corresponding to the speech data are used as the input of the age prediction model, each feature is concatenated as the input.

[0083] In this application, the age prediction model can be a deep learning model commonly used in pattern recognition, such as GMM, CNN, encoder-decoder, BiLSTM, etc. Considering that LSTM can solve the problem of long-term memory, and BiLSTM can see both the history and the future compared to LSTM, therefore, as a preferred implementation method, this case can use BiLSTM, i.e., a bidirectional long-short-term neural network. The specific training process is the same as the traditional BiLSTM training process, and this case will not be described in detail.

[0084] Considering that the results of the above model may be more dependent on the rationality of the training data, the rationality of the model structure and other uncontrollable factors, which may lead to the model output results being not desirable, that is, when the maximum a posteriori values ​​are very low, such as lower than 0.5, and the interval credibility in the person-event relationship features is very high, such as 0.9, and the corresponding interval length is less than a certain threshold, such as 10 years, we can directly select the interval and credibility in the person-event relationship features as the final system output.

[0085] A method for predicting age based on speech is described above, and a device for executing the method for predicting age based on speech is described below.

[0086] See also Figure 2 , Figure 2 The following is a schematic diagram of the structure of a speech-based age prediction device provided in an embodiment of the present application. Figure 2 As shown, the speech-based age prediction device comprises: An acquisition unit 11, used to acquire speech data of a speaker to be predicted; A speech recognition unit 12, used to recognize the speech data to obtain a recognized text; The age prediction reference feature determination unit 13 is used to determine the age prediction reference features of the speaker to be predicted based on the speech data and the recognition text, wherein the age prediction reference features include acoustic features and character event relationship features; the character event relationship features are used to characterize the age range and credibility of the speaker to be predicted determined based on character events related to the age of the speaker to be predicted; The age prediction result determining unit 14 is configured to determine the age prediction result of the speaker to be predicted based on the age prediction reference features of the speaker to be predicted.

[0087] In a possible implementation, the age prediction reference feature of the speaker to be predicted further includes: At least one of a speaking habit feature, a speaking topic feature, a title feature, a name feature, and a gender-biased vocabulary feature, wherein the speaking habit feature is used to characterize at least one of the catchphrases, modal particles, speech speed, and speaking intervals of the speaker to be predicted, the speaking topic feature is used to characterize a topic related to the speech content of the speaker to be predicted, the title feature is used to characterize the title of the speaker to be predicted by others, the name feature is used to characterize the name of the speaker to be predicted, and the gender-biased vocabulary feature is used to characterize vocabulary that tends to distinguish between genders.

[0088] In a possible implementation, the acoustic features include: At least one of a Mel-frequency cepstral coefficient feature, a voice intonation feature, and an audio strength feature.

[0089] In a possible implementation, the speaking habit features include: At least one of a catchphrase feature, an interjection feature, a speech speed feature, and a speaking interval feature.

[0090] In a possible implementation, the age prediction reference feature determination unit is specifically used to: determining subject-related content of the identified text; The features of the subject-related content of the recognized text are determined as the speech subject features.

[0091] In a possible implementation, the age prediction reference feature determination unit is specifically used to: Based on a pre-built male-oriented vocabulary resource library and a pre-built female-oriented vocabulary resource library, determining each male-oriented vocabulary and each female-oriented vocabulary in the recognition text; Identify masculine and feminine lexical features; The male-biased lexical features and female-biased lexical features are concatenated to obtain the speaker's gender-biased lexical features.

[0092] In a possible implementation, the age prediction reference feature determination unit is specifically used to: Get a pre-built library of person-event relationships; Performing named entity recognition and relationship extraction on the text to obtain entities in the recognized text; Matching the entity in the recognized text with the person-event relationship resource library to determine the target entity in the recognized text and the time corresponding to the target entity; Determine at least one age range and its credibility by using the target entity in the identified text and the time corresponding to the target entity; The at least one age range and its credibility are fused to obtain a final age range and a final credibility as the person-event relationship feature.

[0093] In a possible implementation, the age prediction result determination unit is specifically used to: Determining the gender of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted; The age prediction reference features of the speaker to be predicted are input into an age prediction model corresponding to the gender of the speaker to be predicted to obtain an age prediction result of the speaker to be predicted.

[0094] The present application also provides an electronic device in an embodiment. Figure 3 As shown, it shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present application. The electronic device in the embodiment of the present application may include but is not limited to fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 3 The electronic device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0095] like Figure 3As shown, the electronic device may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 to a random access memory (RAM) 603. When the electronic device is powered on, various programs and data required for the operation of the electronic device are also stored in the RAM 603. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0096] Typically, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a memory card, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although Figure 3 An electronic device having various devices is shown, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead.

[0097] An embodiment of the present application also provides a computer program product, including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements any of the speech-based age prediction methods provided in the embodiments of the present application.

[0098] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the speech-based age prediction methods provided in the embodiments of the present application.

[0099] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0100] Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0101] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0102] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

Claims

1. A method for predicting age based on speech, characterized in that: include: Obtaining speech data of the speaker to be predicted; Recognizing the voice data to obtain recognized text; Determining age prediction reference features of the speaker to be predicted based on the speech data and the recognized text, wherein the age prediction reference features include acoustic features and character-event relationship features; The character-event relationship feature is used to characterize the age range and credibility of the speaker to be predicted determined based on the character events related to the age of the speaker to be predicted; Based on the age prediction reference features of the speaker to be predicted, an age prediction result of the speaker to be predicted is determined.

2. The method according to claim 1, characterized in that The age prediction reference features of the speaker to be predicted also include: At least one of a speaking habit feature, a speaking topic feature, a title feature, a name feature, and a gender-biased vocabulary feature, wherein the speaking habit feature is used to characterize at least one of the catchphrases, modal particles, speech speed, and speaking intervals of the speaker to be predicted, the speaking topic feature is used to characterize a topic related to the speech content of the speaker to be predicted, the title feature is used to characterize the title of the speaker to be predicted by others, the name feature is used to characterize the name of the speaker to be predicted, and the gender-biased vocabulary feature is used to characterize vocabulary that tends to distinguish between genders.

3. The method according to claim 1, characterized in that The acoustic characteristics include: At least one of a Mel-frequency cepstral coefficient feature, a voice intonation feature, and an audio strength feature.

4. The method according to claim 2, characterized in that: The speaking habit features include: At least one of a catchphrase feature, an interjection feature, a speech speed feature, and a speaking interval feature.

5. The method according to claim 2, characterized in that: The method for determining the speaking topic feature includes: determining subject-related content of the identified text; The features of the subject-related content of the recognized text are determined as the speech subject features.

6. The method according to claim 2, characterized in that The method for determining the gender-biased vocabulary features includes: Based on a pre-built male-oriented vocabulary resource library and a pre-built female-oriented vocabulary resource library, determining each male-oriented vocabulary and each female-oriented vocabulary in the recognition text; Identify masculine and feminine lexical features; The male-biased lexical features and female-biased lexical features are concatenated to obtain the speaker's gender-biased lexical features.

7. The method according to claim 1, characterized in that The method for determining the character-event relationship feature includes: Get a pre-built library of person-event relationships; Performing named entity recognition and relationship extraction on the text to obtain entities in the recognized text; Matching the entity in the recognized text with the person-event relationship resource library to determine the target entity in the recognized text and the time corresponding to the target entity; Determine at least one age range and its credibility by using the target entity in the identified text and the time corresponding to the target entity; The at least one age range and its credibility are fused to obtain a final age range and a final credibility as the character-event relationship feature.

8. The method according to claim 1, characterized in that: The step of determining the age prediction result of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted includes: Determining the gender of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted; The age prediction reference feature of the speaker to be predicted is input into an age prediction model corresponding to the gender of the speaker to be predicted to obtain an age prediction result of the speaker to be predicted.

9. A speech-based age prediction device, characterized in that: include: An acquisition unit, used for acquiring speech data of a speaker to be predicted; A speech recognition unit, used to recognize the speech data to obtain a recognized text; an age prediction reference feature determination unit, configured to determine age prediction reference features of the speaker to be predicted based on the speech data and the recognition text, wherein the age prediction reference features include acoustic features and character event relationship features; the character event relationship features are used to characterize the age range and credibility of the speaker to be predicted determined based on the character events corresponding to the speaker to be predicted; The age prediction result determining unit is used to determine the age prediction result of the speaker to be predicted based on the age prediction reference feature of the speaker to be predicted.

10. A computer program product, characterized in that The invention comprises computer-readable instructions, which, when executed on an electronic device, enable the electronic device to implement the age prediction method based on speech as claimed in any one of claims 1 to 8.

11. An electronic device, characterized in that: The method comprises at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the voice-based age prediction method as described in any one of claims 1 to 8.

12. A computer-readable storage medium, characterized in that: The storage medium carries one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the voice-based age prediction method as described in any one of claims 1 to 8.

Images