TW201828281A - Method and device for constructing pronunciation dictionary capable of inputting a speech acoustic feature of the target vocabulary into a speech recognition decoder - Google Patents

Abstract

The invention discloses a method for constructing a pronunciation dictionary, which is used to solve the problem of poor quality of a pronunciation dictionary constructed according to the prior technologies. The method includes: inputting a speech acoustic feature of the target vocabulary into a speech recognition decoder, wherein the pronunciation dictionary in the speech recognition decoder includes: a target vocabulary and a candidate pronunciation phoneme sequence of the target vocabulary; determining, according to the candidate pronunciation phoneme sequence output by the speech recognition decoder, a probability distribution of the target vocabulary corresponding to the output candidate pronunciation phoneme sequence; selecting, according to the probability distribution, a pronunciation phoneme sequence that is a correct pronunciation of the target vocabulary from the output candidate pronunciation phoneme sequence; and constructing a pronunciation dictionary based on the correctly pronounced phoneme sequence of the pronunciation. The invention also discloses a construction device for a pronunciation dictionary.

Description

 Method and device for constructing pronunciation dictionary  

The invention relates to the field of computer technology, and in particular to a method and a device for constructing a pronunciation dictionary.

Voice interactive technology has already appeared in the middle of the twentieth century. In recent years, with the popularization of smart phones, a large number of voice interactive products have appeared one after another, and voice interactive products have entered the daily life of ordinary users. For example, the voice input method is to receive and recognize the voice sent by the user, and then convert the voice of the user into a text, thereby eliminating the cumbersome input of typing; the caller number function can output the text in the form of voice, in the user You can get the identity of the caller without looking at the screen.

In the voice interaction technology, the pronunciation dictionary is an important part of the voice interaction system. It is the bridge between the acoustic model and the language model. Its coverage and pronunciation quality have a significant impact on the overall performance of the system.

The pronunciation dictionary contains the mapping relationship between the word and the pronunciation phoneme sequence, and the mapping relationship can usually be established by using the word conversion to the phoneme (Graphme-to-Phoneme, G2P) method. Under normal circumstances, the pronunciation dictionary is corrected by experts in linguistics related aspects, and the scale is relatively fixed, so it is impossible to cover all vocabulary. Therefore, in practical applications, it is possible to use G2P method to determine new vocabulary according to needs. The matched pronunciation phoneme sequence determines the correct pronunciation of the newly added vocabulary, and then expands the existing pronunciation dictionary according to the newly added vocabulary and the pronunciation phoneme sequence matched with it.

At present, using the G2P method, it is basically possible to accurately determine the correct pronunciation of conventional vocabulary. However, for some special words, such as vocabulary containing polyphonic words, the accuracy of the correct pronunciation of the vocabulary determined by this method is often low, thus affecting the quality of the pronunciation dictionary.

The embodiment of the invention provides a method for constructing a pronunciation dictionary, which is used to solve the problem of poor quality of a pronunciation dictionary constructed according to the prior art.

The embodiment of the invention further provides a device for constructing a pronunciation dictionary for solving the problem of poor quality of a pronunciation dictionary constructed according to the prior art.

The embodiment of the present invention adopts the following technical solution: a method for constructing a pronunciation dictionary, comprising: inputting a speech acoustic feature of a target vocabulary into a speech recognition decoder; wherein the pronunciation dictionary in the speech recognition decoder includes: a target vocabulary and a candidate pronunciation phoneme sequence of the target vocabulary; determining, according to the candidate pronunciation phoneme sequence output by the speech recognition decoder with the speech acoustic feature as an input, determining a probability distribution of the target vocabulary corresponding to the output candidate pronunciation phoneme sequence; Describe the probability distribution, select a pronunciation phoneme sequence that is the correct pronunciation of the target vocabulary from the output candidate phoneme sequence, and construct a pronunciation dictionary according to the correctly pronounced pronunciation phoneme sequence.

A device for constructing a pronunciation dictionary, comprising: a decoding unit: configured to input a speech acoustic feature of a target vocabulary into a speech recognition decoder; wherein the pronunciation dictionary in the speech recognition decoder comprises: a target vocabulary and a target vocabulary a candidate pronunciation phoneme sequence; a pronunciation determining unit: configured to determine, according to the candidate pronunciation phoneme sequence output by the speech recognition decoder with the speech acoustic feature as an input, determining a probability distribution of the target vocabulary corresponding to the output candidate phoneme sequence And selecting, according to the probability distribution, a pronunciation phoneme sequence that is a correct pronunciation of the target vocabulary from the output candidate phoneme sequence; a dictionary constructing unit: configured to construct a pronunciation according to the correctly pronounced pronunciation phoneme sequence dictionary.

The above at least one technical solution adopted by the embodiment of the present invention can achieve the following beneficial effects: since the speech acoustic feature of the target word to be predicted is introduced, as one of the basis for predicting the correct pronunciation of the vocabulary, thereby relying on only the vocabulary and the pronunciation phoneme As a prior art for predicting the correct pronunciation of vocabulary, the mapping relationship of sequences can more accurately predict the correct pronunciation of the target vocabulary and improve the quality of the pronunciation dictionary based on the determined correct pronunciation construction.

21‧‧‧Decoding unit

22‧‧‧ pronunciation determination unit

23‧‧‧ dictionary construction unit

The drawings are intended to provide a further understanding of the invention, and are intended to be a part of the invention. In the accompanying drawings: FIG. 1 is a schematic diagram showing an implementation flow of a method for constructing a pronunciation dictionary according to an embodiment of the present invention; FIG. 2 is a schematic structural diagram of a device for constructing a pronunciation dictionary according to an embodiment of the present invention.

The technical solutions of the present invention will be clearly and completely described in conjunction with the specific embodiments of the present invention and the accompanying drawings. It is apparent that the described embodiments are only a part of the embodiments of the invention, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention without creative efforts are within the scope of the present invention.

The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

 Example 1  

The existing pronunciation prediction method is usually based on the G2P conversion method, and the G2P method converts the vocabulary into a pronunciation phoneme sequence by establishing a mapping relationship between vocabulary and pronunciation phonemes. Using the G2P method, the pronunciation phoneme sequence matching the regular vocabulary can be basically obtained accurately, but since the method only utilizes the mapping relationship between the vocabulary (word sequence) and the pronunciation phoneme, for some special words, such as words containing polyphonic words, The accuracy of the pronunciation phoneme sequence determined by the method is often lower, which affects the quality of the pronunciation dictionary.

In order to solve the problem that the prior art cannot accurately predict the correct pronunciation of the vocabulary and thus affect the quality of the pronunciation dictionary, Embodiment 1 of the present invention provides a method for constructing a pronunciation dictionary.

The execution body of the pronunciation dictionary construction method provided by the embodiment of the present invention may be a server or other device different from the server, and the like. The executor of the present invention is not limited to the present invention. For the convenience of description, the embodiments of the present invention are described by taking an execution subject as a server as an example.

For convenience of description, in the present embodiment, vocabulary and phonetic acoustic features having a correspondence relationship may be represented by vocabulary-speech acoustic features.

Similarly, a vocabulary (word sequence) and a phoneme sequence in which a correspondence exists, and a phonetic acoustic feature and a phoneme sequence in which a correspondence exists, can also be expressed by the above representation. For example, there are vocabulary and phoneme sequences corresponding to each other, which can be represented by a lexical-speech phoneme sequence.

The method of the present invention is provided in detail below.

The schematic diagram of the implementation process of the method is as shown in FIG. 1 , and includes the following steps: Step 11: The server inputs the speech acoustic characteristics of the target vocabulary into a speech recognition decoder embedded with a pronunciation dictionary, an acoustic model and a language model; In the embodiment, the target vocabulary may be any vocabulary, such as a Chinese vocabulary, an English vocabulary or a vocabulary in other languages. For the pronunciation dictionary existing in the speech recognition decoder, the target vocabulary may refer to a vocabulary that is not currently included in the pronunciation dictionary, that is, a new vocabulary relative to the pronunciation dictionary.

The speech acoustic feature of the target vocabulary described in the embodiment of the present invention may include, but is not limited to, a Filter Bank feature extracted from a speech signal generated by speaking the target vocabulary, and a Mel frequency cepstral coefficient. (Mel Frequency Cepstrum Coefficient; MFCC) features and at least one of Perceptual Linear Predictive (PLP) features and the like.

In the embodiment of the present invention, the voice signal may be, for example, an audio sample corresponding to the target vocabulary.

The audio sample corresponding to the target vocabulary can be obtained by using at least one of the following methods: 1. Entrusting a professional voice data provider to perform manual recording to obtain an audio sample corresponding to the target vocabulary; In the form of the user's real use experience and personal experience, the recording task is freely and voluntarily entrusted to the non-specific (and usually large) Internet mass to obtain the audio sample corresponding to the target vocabulary; The log of the user feedback is analyzed to obtain an audio sample corresponding to the target vocabulary. For example, in a voice search task, the user first inputs the target vocabulary through voice. If the voice recognition system recognizes an error and the user continues to input the correct target vocabulary through the keyboard, the series of behaviors can be recorded in the form of a log.

In the embodiment of the present invention, the speech acoustic features may be respectively obtained from the audio samples corresponding to the target vocabulary, and the obtained speech acoustic features are respectively input as the speech acoustic features of the target vocabulary into the speech recognition decoder.

The operation of the speech recognition decoder mentioned in step 11 is further described below.

Generally, a speech recognition decoder is used to search for a speech signal (or a phonetic acoustic feature) for an input, based on an acoustic model, a language model, and a pronunciation dictionary, to generate the speech signal with maximum probability (or with the acoustical characteristics of the speech) A virtual or physical device of the word that matches the voice signal).

In the field of speech recognition, the goal of decoding a speech signal is to find a word sequence W ^* (equivalent to the "word" described above), so that the probability of the corresponding acoustical feature X is maximized, which is essentially based on The machine learning problem of the Bayesian criterion is to use the Bayesian formula to calculate the optimal word sequence W ^* , as shown in the formula [1.1]:

Where P ( X | W _i ) is the acoustic model and P ( W _i ) is the language model.

The acoustic model is the probability that the speech acoustic characteristic of the word sequence W _i is X. Acoustic models can generally be trained using a large amount of data, including speech acoustic features and corresponding tag sequences.

The language model is the probability of occurrence of the word sequence W _i corresponding to the vocabulary. The probability of occurrence is generally: the probability that each word constituting a vocabulary appears in sequence according to the order in which the respective words are arranged in the vocabulary.

Considering the different pronunciation phoneme sequences that the word sequence generally corresponds to, for example, the pronunciation of a certain vocabulary (which can be represented by a word sequence) with different local accents may correspond to different pronunciation phonemes, or the words containing multi-tone words may also correspond to different Pronunciation phoneme, therefore, if assumed Is the pronunciation phoneme sequence corresponding to the word sequence W _i , then the formula [1.1] can be changed to:

Where W _i is a sequence of words; P ( X | ) is an acoustic model; P ( W _i ) is a language model; P ( | W _i ) is the pronunciation phoneme sequence of the vocabulary in the pronunciation dictionary (represented by the word sequence W _i ) The chance.

For the problem of pronunciation learning, further assuming that the word sequence W _i and the corresponding speech acoustic feature X are known, the calculation target of the formula [1.2] can be converted to find the best pronunciation phoneme sequence Q corresponding to the word sequence W _i . ^* . Thus, the formula [1.2] can be further changed to:

In the formula [1.3]: Q ^* is the maximum value of the probability phoneme sequence of the candidate phoneme sequence corresponding to the value of the right side of the formula [1.3], that is, the maximum value of the candidate phoneme sequence corresponding to the word sequence W _i ; W _i is a word sequence, i is the number of the vocabulary; X is the phonetic acoustic feature corresponding to W _i ; Q is the phonetic sequence of the pronunciation; j is the number of the phoneme sequence of the pronunciation; The pronunciation phoneme sequence numbered j in the speech phoneme sequence corresponding to the vocabulary number i .

P ( X | Acoustic model The corresponding speech acoustic feature is the probability of X.

At present, the acoustic model used in the related speech recognition technology is generally obtained by training a hybrid model of the Hidden Markov Model-Deep Neural Network (HMM-DNN), or it may be Trained on the DNN model. In the embodiment of the present invention, the hybrid model or the DNN model of the HMM-DNN can be trained to obtain an acoustic model through a huge amount of speech acoustic features, and is set in the speech recognition decoder described in the embodiment of the present invention.

P ( W _i ) is a language model--the language model in this embodiment may be an N-Gram model based on the assumption that the appearance of the Nth word is only related to the previous N-1 words, and other No word is relevant. The probability of a whole sentence is the product of the probability of occurrence of each word. The probability of occurrence of each word can be obtained by counting the number of simultaneous occurrences of N words directly from the corpus. The language model in this embodiment may also be a language model based on a conditional random domain or a deep neural network strategy. The language model can be pre-generated and arranged in the speech recognition decoder described in the embodiment of the present invention.

P ( | W _i ) is the pronunciation phoneme sequence based on the vocabulary in the given pronunciation dictionary (represented by the word sequence W _i ) The chance.

The pronunciation dictionary mentioned here may be, for example, a pronunciation dictionary in which each candidate phoneme sequence corresponding to the target vocabulary is added.

The candidate pronunciation phoneme sequence of the target vocabulary refers to the pronunciation phoneme sequence that may be correctly pronounced as the target vocabulary. In the embodiment of the present invention, the G2P method may be used to generate a pronunciation phoneme sequence (referred to as "candidate phoneme sequence" in the embodiment of the present invention) for the target vocabulary, and the target vocabulary and the generated candidate phoneme sequence are generated. , added to the pronunciation dictionary.

The adding the target vocabulary and the generated candidate phoneme sequence to the pronunciation dictionary may refer to adding the term including the target vocabulary-candidate phoneme sequence to the pronunciation dictionary.

It should be noted that when the pronunciation dictionary is not currently present, adding the term to the pronunciation dictionary may refer to constructing a pronunciation dictionary according to the term; when the pronunciation dictionary currently exists, adding the term In the pronunciation dictionary, the existing pronunciation dictionary may be updated according to the term to obtain an updated pronunciation dictionary.

For convenience of description, it is assumed in the embodiment of the present invention that a pronunciation dictionary currently exists. In such a scenario, the target vocabulary is a new vocabulary relative to the currently existing pronunciation dictionary.

In this embodiment, the number of corresponding candidate phoneme sequences generated for the target vocabulary depends on the actual situation.

For example, using the G2P method, more than ten candidate phoneme sequences can be generated for the target vocabulary "Alibaba." Taking one of the pronunciation phoneme sequences as an example, it can be expressed as "a1/li3/ba1/ba1/". In the pronunciation phoneme sequence, the symbol "/" is used to distinguish different phonemes, that is, the symbols before and after the "/" indicate different phonemes. For example, a1 and li3 are different phonemes. The number in the phoneme represents the tone, that is, 1 represents a tone, 2 represents a tone, 3 represents a tone, and 4 represents a tone of four.

Based on the acoustic model P ( X | embedded in the above pronunciation dictionary, as shown in the formula [1.3] And a speech recognition decoder of the language model P ( W _i ), in the embodiment of the present invention, inputting a speech acoustic feature of the target vocabulary into the speech recognition decoder, which can trigger the speech recognition decoder to transmit acoustic characteristics to the speech sample Decoding, outputting the pronunciation phoneme sequence corresponding to the acoustic feature of the speech sample.

The subsequent steps included in the method provided by the embodiments of the present invention are further described below.

Step 12: determining a candidate pronunciation phoneme sequence output by the speech recognition decoder with the speech acoustic feature described in step 11 as an input; and determining a target vocabulary corresponding according to a statistical rule of the target vocabulary corresponding to the output candidate pronunciation phoneme sequence a probability distribution of the candidate pronunciation phoneme sequence outputted; selecting, according to the probability distribution, a pronunciation phoneme sequence that is the correct pronunciation of the target vocabulary from the output candidate pronunciation phoneme sequence; for example, if a candidate corresponding to the target vocabulary T is assumed There are two pronunciation phoneme sequences, A1 A2 and B1 B2, respectively, and they are added to the pronunciation dictionary included in the speech recognition decoder. Further, if it is assumed that there are 100 audio samples of the collected T, so that the respective acoustic acoustic features of the 100 audio samples (a total of 100 speech acoustic features) can be obtained, by performing step 11, the 100 speech acoustic features are obtained. They are input into a speech recognition decoder embedded in a pronunciation dictionary, an acoustic model, and a language model, respectively.

Then, the speech recognition decoder identifies and decodes the 100 speech acoustic features, and can output a candidate pronunciation phoneme sequence, such as a combination of outputs A1, A2, B1, B2.

Further, it is assumed that according to a pronunciation dictionary set in the speech recognition decoder, a statistical rule of determining a target vocabulary corresponding to the output candidate phoneme sequence is: among the 100 speech acoustic features: 75 speech acoustic features are The term "T-A1 A2" through the pronunciation dictionary is mapped to T, and 25 speech acoustic features are mapped to T through the term "T-B1 B2" of the pronunciation dictionary.

Then, according to the statistical law, the following probability distribution can be obtained: the probability that T corresponds to A1 A2 is 75/100=0.75

The probability that T corresponds to B1 B2 is 25/100=0.25

Generally, the server may determine the candidate phoneme sequence corresponding to the maximum probability value in the probability distribution as the pronunciation phoneme sequence of the correct pronunciation of the target vocabulary.

Following the above example, the server may determine the candidate phoneme sequence A1 A2 corresponding to the maximum probability value of 0.75 in the probability distribution as the pronunciation phoneme sequence of T correctly.

Step 13: Construct a pronunciation dictionary based on the pronunciation phoneme sequence that is correctly pronounced as the target vocabulary.

Specifically, the server may delete other candidate phoneme sequences corresponding to the target vocabulary other than the pronunciation phoneme sequence that is correctly pronounced as the target vocabulary, for example, from the pronunciation dictionary in which each candidate phoneme sequence corresponding to the target vocabulary is added. Alternatively, the server may reconstruct a new pronunciation dictionary based on the pronunciation phoneme sequence that is correctly pronounced as the target vocabulary.

According to the above method provided by Embodiment 1 of the present invention, since the speech acoustic characteristics of the target word to be predicted are introduced, as one of the basis for predicting the correct pronunciation of the vocabulary, the mapping relationship between the vocabulary and the pronunciation phoneme sequence is performed. In the prior art for predicting the correct pronunciation of vocabulary, the correct pronunciation of the target vocabulary can be predicted more accurately, thereby improving the quality of the speech dictionary.

 Example 2  

In order to solve the problem that the accuracy of the pronunciation phoneme sequence matching the vocabulary is low by using the prior art, the embodiment of the present invention provides a device for constructing a pronunciation dictionary. The schematic diagram of the vocabulary pronunciation prediction apparatus is shown in FIG. 3, and mainly includes the following functional unit: a decoding unit 21, configured to input a speech acoustic feature of a target vocabulary into a speech recognition decoder; wherein, in the speech recognition decoder The pronunciation dictionary includes: a candidate pronunciation phoneme sequence of the target vocabulary and the target vocabulary; the pronunciation determining unit 22 is configured to determine the target according to the candidate pronunciation phoneme sequence output by the speech recognition decoder with the speech acoustic feature as an input The vocabulary corresponds to a probability distribution of the output candidate phoneme sequence; according to the probability distribution, a pronunciation phoneme sequence that is the correct pronunciation of the target vocabulary is selected from the output candidate phoneme sequence; the dictionary constructing unit 23 A pronunciation dictionary is constructed based on the pronunciation phoneme sequence of the correct pronunciation.

In an embodiment, the apparatus provided by the embodiment of the present invention may further include a phoneme sequence processing unit. The unit is configured to obtain a candidate phoneme sequence of the target vocabulary before inputting the phonetic acoustic feature of the target vocabulary into the speech recognition decoder; and adding the target vocabulary and the obtained candidate phoneme sequence to the speech recognition decoder In the pronunciation dictionary.

In an embodiment, the phoneme sequence processing unit may be specifically configured to obtain a candidate phoneme sequence of the target vocabulary by using the G2P method.

In an embodiment, the decoding unit 21 may be specifically configured to collect an audio sample corresponding to the target vocabulary; obtain the voice acoustic feature according to the audio sample; and input the obtained voice acoustic feature into the voice Identify the decoder.

In an embodiment, the pronunciation determining unit 22 may be specifically configured to determine a maximum probability value in the probability distribution; and select, from the output candidate pronunciation phoneme sequence, a candidate pronunciation phoneme corresponding to the maximum probability value. A sequence, a phoneme sequence that is the correct pronunciation of the target vocabulary.

In an embodiment, the dictionary constructing unit 23 may be specifically configured to delete the target vocabulary from the pronunciation dictionary that adds the target vocabulary and the obtained candidate phoneme sequence according to the pronunciation phoneme sequence that is correctly pronounced as the target vocabulary. Corresponding to other candidate phoneme sequences other than the correctly pronounced pronunciation phoneme sequence.

According to the above apparatus provided in Embodiment 2 of the present invention, since the speech acoustic characteristics of the target word to be predicted are introduced, as one of the basis for predicting the correct pronunciation of the vocabulary, the mapping relationship between the vocabulary and the pronunciation phoneme sequence is performed. In the prior art for predicting the correct pronunciation of vocabulary, the correct pronunciation of the target vocabulary can be predicted more accurately.

Those skilled in the art will appreciate that embodiments of the present invention can be provided as a method, system, or computer program product. Thus, the present invention can take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining soft and hardware aspects. Moreover, the present invention can take the form of a computer program product embodied on one or more computer usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable code therein. .

The present invention has been described with reference to flowchart illustrations and/or block diagrams of a method, a device (system), and a computer program product according to an embodiment of the invention. It will be understood that each flow and/or block of the flowcharts and/or <RTIgt; The computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing device to produce a machine for executing instructions by a processor of a computer or other programmable data processing device Means are generated for implementing the functions specified in one or more flows of the flowchart or in a block or blocks of the block diagram.

The computer program instructions can also be stored in a computer readable memory that can boot a computer or other programmable data processing device to operate in a particular manner, such that instructions stored in the computer readable memory produce an article of manufacture including the instruction device. The instruction means implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.

These computer program instructions can also be loaded onto a computer or other programmable data processing device to perform a series of operational steps on a computer or other programmable device to produce computer-implemented processing for use in a computer or other programmable device The instructions executed on the steps provide steps for implementing the functions specified in one or more flows of the flowchart or in a block or blocks of the flowchart.

The above description is only an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and changes can be made in the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims (14)

 A method for constructing a pronunciation dictionary, comprising: inputting a speech acoustic feature of a target vocabulary into a speech recognition decoder; wherein the pronunciation dictionary in the speech recognition decoder comprises: a candidate for a target vocabulary and a target vocabulary a phoneme sequence; a probability pronunciation phoneme sequence output by the speech recognition decoder with the speech acoustic feature as an input, determining a probability distribution of the target vocabulary corresponding to the output candidate phoneme sequence; according to the probability distribution, the output is In the candidate pronunciation phoneme sequence, the pronunciation phoneme sequence which is the correct pronunciation of the target vocabulary is selected; and the pronunciation dictionary is constructed based on the correctly pronounced pronunciation phoneme sequence.    The method of claim 1, wherein the voice acoustic feature is input to the speech recognition decoder, the method further comprising: obtaining a candidate phoneme sequence of the target vocabulary; and selecting the target vocabulary and the obtained candidate The pronunciation phoneme sequence is added to the pronunciation dictionary in the speech recognition decoder.    The method of claim 2, wherein obtaining the candidate phoneme sequence of the target vocabulary comprises: converting the word into a phoneme G2P method to obtain a candidate phoneme sequence of the target vocabulary.    The method of claim 1, wherein the acoustic model embedded in the speech recognition decoder is obtained by training a deep neural network.    The method of claim 1, wherein the inputting the speech acoustic feature of the target vocabulary into the speech recognition decoder comprises: collecting an audio sample corresponding to the target vocabulary; and obtaining the acoustical acoustic feature according to the audio sample The obtained speech acoustic feature is input to the speech recognition decoder.    The method of claim 1, wherein, according to the probability distribution, selecting a pronunciation phoneme sequence that is the correct pronunciation of the target vocabulary from the output candidate phoneme sequence includes: determining the probability distribution The maximum probability value; from the output candidate phoneme sequence, the candidate phoneme sequence corresponding to the maximum probability value is selected as the pronunciation phoneme sequence of the correct pronunciation of the target vocabulary.    The method according to any one of claims 1 to 6, wherein the pronunciation pronunciation dictionary is constructed according to the correctly pronounced pronunciation phoneme sequence, comprising: joining according to the pronunciation phoneme sequence which is correctly pronounced as the target vocabulary In the pronunciation dictionary of the target vocabulary and the obtained candidate phoneme sequence, the candidate phoneme sequence corresponding to the correctly pronounced pronunciation phoneme sequence corresponding to the target vocabulary is deleted.    A device for constructing a pronunciation dictionary, comprising: a decoding unit: configured to input a speech acoustic feature of a target vocabulary into a speech recognition decoder; wherein the pronunciation dictionary in the speech recognition decoder comprises: a target vocabulary And a candidate pronunciation phoneme sequence of the target vocabulary; a pronunciation determining unit: configured to determine, according to the candidate pronunciation phoneme sequence output by the speech recognition decoder with the speech acoustic feature as an input, the target vocabulary corresponding to the output candidate pronunciation phoneme sequence a probability distribution; according to the probability distribution, a pronunciation phoneme sequence that is a correct pronunciation of the target vocabulary is selected from the output candidate phoneme sequence; a dictionary construction unit: configured to construct a pronunciation dictionary according to the correctly pronounced pronunciation phoneme sequence.    The device of claim 8, wherein the device further comprises: a phoneme sequence processing unit, configured to obtain a candidate phoneme of the target vocabulary before inputting the phonetic acoustic feature of the target vocabulary into the speech recognition decoder a sequence; and the target vocabulary and the obtained candidate phoneme sequence are added to the pronunciation dictionary in the speech recognition decoder.    The device of claim 9, wherein the phoneme sequence processing unit is specifically configured to: use a word conversion to a phoneme G2P method to obtain a candidate phoneme sequence of the target vocabulary.    The apparatus of claim 8, wherein the acoustic model embedded in the speech recognition decoder is obtained by training a deep neural network.    The device of claim 8, wherein: the decoding unit is configured to collect an audio sample corresponding to the target vocabulary; obtain the speech acoustic feature according to the audio sample; and input the obtained acoustic acoustic feature into the In the speech recognition decoder.    The device of claim 8, wherein the pronunciation determining unit is specifically configured to: determine a maximum probability value in the probability distribution; and select, from the output candidate phoneme sequence, the maximum probability value. The candidate pronunciation phoneme sequence, which is the pronunciation phoneme sequence of the correct pronunciation of the target vocabulary.    The device according to any one of the preceding claims, wherein the dictionary construction unit is specifically configured to: add a target vocabulary and obtain a candidate pronunciation according to a pronunciation phoneme sequence that is correctly pronounced as the target vocabulary In the pronunciation dictionary of the phoneme sequence, other candidate phoneme sequences corresponding to the correctly pronounced pronunciation phoneme sequence corresponding to the target vocabulary are deleted.