CN111583914B - Big data voice classification method based on Hadoop platform - Google Patents

Abstract

The invention relates to a big data voice classification method based on a Hadoop platform, which comprises the following steps: step 1, constructing a voice library; step 2, on the basis of the voice library, based on a Hadoop platform, subdividing the big data voice classification problem by adopting a Map function, and performing voice classification solution on the subproblems in a multi-node parallel and distributed manner to obtain corresponding voice classification results; and 3, finally, combining the voice classification results of the sub-problems by using a Reduce function so as to adapt to the online requirement of the voice classification of the big data.

Description

A Big Data Speech Classification Method Based on Hadoop Platform

technical field

The invention relates to a big data speech classification method based on Hadoop platform.

Background technique

(1) Research status of motor dysarthria:

Motor dysarthria (dysarthria) refers to a group of speech disorders caused by the disorder of muscle control due to damage to the central nervous system or peripheral nervous system. Motor dysarthria often manifests as slowed, weakened, imprecise, and uncoordinated movement of speech-related muscle tissues, and may also affect breathing, resonance, and control of laryngeal vocalization, articulation, and rhythm. Clinically, it is often referred to as dysarthria for short . Common causes of motor dysarthria include traumatic brain injury, cerebral palsy, amyotrophic lateral sclerosis, multiple sclerosis, stroke, Parkinson's disease, spinocerebellar ataxia, etc. Dysarthria can be divided into flaccid, spastic, dysarthria, hypokinetic, hyperkinetic, and mixed types according to neuroanatomical and speech acoustic characteristics. Dysarthria occurs in up to 54% of communication disorders associated with brain injury. At present, the speech-acoustic characteristics of dysarthria can be reflected from both subjective and objective aspects through the examination of voice, resonance, and rhythm in clinical practice, which is conducive to providing targeted treatment and comprehensively and scientifically clarifying the speech-acoustic pathology of dysarthria mechanism.

There are few domestic and foreign research reports on the overall incidence of motor dysarthria. Miller et al. conducted a study on 125 patients with Parkinson's disease, showing that 69.6% of the patients had lower speech intelligibility than the normal control group, and 51.2% of them had patients were one standard deviation lower, suggesting a higher prevalence of dysarthria in Parkinson's patients. Bogousslavsky et al. screened 1,000 first-time stroke patients and found that as many as 46% of patients had speech impairment, of which 12.4% were diagnosed as dysarthria. Hartelius and other studies also found that the incidence of dysarthria in patients with multiple sclerosis was 51%. This shows that the incidence of dysarthria is higher. There is no unified evaluation method for dysarthria in China at present, and there is no special evaluation standard for motor dysarthria. Most of them use the Frenchay dysarthria evaluation method or the improved method and the dysarthria checklist of the China Rehabilitation Research Center. Physicians or rehabilitation physicians check, score, record, and evaluate the degree and type of dysarthria.

(2) Current status of domestic speech database research:

With the development of information technology and computer science, speech technology enables the interaction between machine behavior and human natural language. Whether it is speech synthesis, speech recognition or speech recognition research, it must rely on the construction of an excellent back-end speech corpus. At present, the development of foreign speech databases is relatively mature, and the research on Chinese speech databases has also advanced by leaps and bounds in the past ten years. The research and establishment of speech databases have been implemented in different language and cultural contexts. However, the construction of the speech library for motor dysarthria is still in the research state.

Domestic studies on the evaluation of articulation and phonetic functions mainly focus on subjective evaluation, and only a few researchers differentiate the concepts of articulation and phonetics. Huang Zhaoming and others proposed the "Chinese Articulation Ability Test Vocabulary", which contains 50 characters. The speech rehabilitation teacher can comprehensively evaluate the subject's understanding of 21 initials and 4 kinds of words by evaluating the pronunciation of the 50 characters of the subject. The articulation ability of the tone, meanwhile, the phoneme contrast ability of the subjects was evaluated through 18 phoneme contrast and 37 smallest phoneme pairs. Chen Sanding and others evaluated the initials, finals and tones of Mandarin Chinese on 50 deaf children, revealed the developmental law of articulation and pronunciation of deaf children who speak Mandarin Chinese, and further proposed the speech rehabilitation of "early, sequential, fault-tolerant and consolidated" Educational principles. Dr. Zhang Jing from East China Normal University studied the main mistakes of hearing-impaired children in the articulation of consonants, analyzed the causes, and proposed a corresponding framework for the treatment of hearing-impaired children's consonant phonemes.

(3) Research status of big data in the medical field:

At present, the popular definition of big data is: data that exceeds the capabilities of typical database software tools to capture, store, process and analyze. Big data is different from traditional data concepts such as ultra-large-scale data and massive data. It has four basic characteristics: large number, variety, timeliness, and value. The State Council pointed out in the "Guiding Opinions on Actively Promoting the "Internet Plus" Action" that we should vigorously develop emerging models of online and offline interaction with the Internet as the carrier, and accelerate the development of Internet-based medical and health services. Kayyali B et al. studied the impact of big data on the medical industry in the United States, and pointed out that as time goes by, the value of big data to the medical industry will become more and more significant. At present, big data in the medical field mainly comes from pharmaceutical companies, clinical diagnosis data, patient medical data, health management, and social network data. For example, drug research and development is a relatively intensive process. Even for small and medium-sized enterprises, the data of a drug research and development is more than TB; the data of hospitals is also growing very fast every day, and a single dual-source CT examination of a patient has 3,000 images. , about 1.5GB of image data is generated, and a standard pathological examination image has nearly 5GB of images, plus data such as patient medical treatment and electronic medical records, which are growing rapidly every day. Research methods based on massive big data analysis have triggered people's thinking about scientific methodology. Research does not need to directly contact the research objects, but new research discoveries can be obtained through direct analysis and mining of massive data, which may give birth to a new scientific research model.

The establishment of a speech corpus is a cumbersome and complicated problem. There are still some issues to be improved in the later improvement of the speech corpus, such as making full use of the existing inter-word tone sandhi rules to reflect the actual situation of tone shift and soft voice as much as possible. For the shortage of corpus, the utilization rate of existing corpus can be improved in the preprocessing link. In view of the above reasons, the voice library should adopt an open database, so that it can be added and modified at any time, so as to improve the database. Due to the different speech situations, the establishment of a specific speech corpus will also encounter various difficulties. The issues we discuss here are just a discussion on the establishment of a speech corpus, hoping to provide some information for the study of speech. Data support plays an important role in better language development and improving the speech corpus.

In addition, the large amount of data is undoubtedly a major advantage of network big data analysis technology, but how to ensure the quality of massive data, and how to realize the cleaning, management and analysis of massive data have also become a major research topic of this topic. Technical Difficulties. Massive network big data has the characteristics of multi-source heterogeneity, interactivity, timeliness, burstiness and high noise, which leads to the characteristics of network big data, although it is huge in value, it is also noisy and low in value density. This poses a huge challenge to ensure the data quality in network big data analysis research.

Contents of the invention

The present invention designs a big data voice classification method based on the Hadoop platform. The technical problem it solves is that the large amount of data is undoubtedly a major advantage of network big data analysis technology, but how to ensure the quality of massive data and how to realize The cleaning, management and analysis of massive data has also become a major technical difficulty.

In order to solve the above-mentioned technical problems, the present invention adopts the following scheme:

A kind of big data speech classification method based on Hadoop platform, comprises the following steps:

Step 1, carry out the construction of voice database;

Step 2. On the basis of this voice library, based on the Hadoop platform, the Map function is used to subdivide the big data voice classification problem, and the sub-problems are solved in a parallel and distributed manner with multiple nodes to obtain the corresponding voice classification results;

Step 3. Finally, use the Reduce function to combine the speech classification results of the sub-problems to meet the online requirements of big data speech classification.

Preferably, the step 2 includes the following content: (1) Client submits a voice classification task to the Job Tracker of the Hadoop platform, and the Job Tracker copies the voice feature data to the local distributed file processing system;

(2) Initialize the task of speech classification, put the task into the task queue, and the Job Tracker assigns the task to the corresponding node according to the processing capabilities of different nodes, that is, the Task Tracker;

(3) Each Task Tracker uses a support vector machine to fit the relationship between the speech features to be classified and the speech feature library according to the assigned tasks, and obtains the corresponding category of the speech;

(4) Save the corresponding voice category as Key/Value to the local file disk;

(5) If the Key/Value of the intermediate results of speech classification are the same, they are merged, and the merged result is handed over to Reduce for processing to obtain the result of speech classification, and the result is written into the distributed file processing system;

(6) Job Tracker clears the task status, and the user gets the result of voice classification from the distributed file processing system.

Preferably, step 1, carrying out the construction of voice bank comprises the following steps: step 11, the design of pronunciation text; Step 12, voice recording; Step 13, the mark of voice file; Step 14, the acoustic parameter analysis to voice file; Step 15 , The establishment of the database management system.

Preferably, the design of the pronunciation text in the step 11 includes the selection of the pronunciation text, and the selection principle of the corpus of the pronunciation text includes one or more of the following:

a. The words in the corpus are required to include all the phonetic phenomena as much as possible, which can better and more conveniently reflect the phonological characteristics of the voices of different patients;

b. The vocabulary in the corpus is based on the common Chinese survey table, so it can be easily compared with Mandarin Chinese;

c. The sentences in the corpus are mainly obtained from conversations with patients based on several related topics, so they are more in line with the real situation faced by speech recognition; "several related topics" include daily life topics or medical history topics, such as asking the time of first onset and medical history.

d. The sentences in the corpus are complete in content and semantics, so it can reflect the prosodic information of a sentence as much as possible;

e. The selection of triphones without classification can effectively solve the problem of sparse training data.

Preferably, the design of the pronunciation text in the step 11 also includes the compilation of the pronunciation text, and the compilation principles of the pronunciation text include one or more of the following:

a. Character part: use the initials, finals, and tones listed in the survey word list as the main recording materials for this speech database;

b. Vocabulary part: Based on at least a 4,000-word vocabulary, record related words according to the original conclusions about related phonology, and strive to fully reflect their phonetic characteristics, including the characteristics of sound quality and super-tone quality, for some very distinctive sounds Phenomenon, example words can be added to reflect its characteristics; "Conclusion Record Related Words of Related Phonetics" refers to common vocabulary summarized according to the characteristics of sounds, combination rules, rhythm and intonation used in the same language. "Characteristic phonetic phenomena" refer to dialects that are prone to mispronunciation, such as flat-tongue sounds that are hard to distinguish, and f and h that are not distinguished.

c. Sentence materials: The number of corpus is determined according to the language mastery of different speakers. When selecting, it is necessary to ensure that the range of corpus is as wide as possible, and to make it representative; "representative" here refers to the ability to Reflect the language characteristics of motor dysarthria and have universal sentences.

d. Natural dialogue part: Daily life as the topic, in the form of answering questions and free conversation, recording 20-40 minutes of audio materials from the speaker, involving vocabulary in daily spoken language that is different from Mandarin, and requiring the speaker to speak in dialect.

Preferably, the voice recording in step 12 includes the determination of the speaker, and the selection principle of the speaker is to choose articulate and moderate speech rate ("moderate speech rate" refers to a moderate rate of speech, controlled at 120-150 words/ Minutes), a native speaker who is proficient in using the local language and willing to actively cooperate with the survey, and must ensure that the language environment in which he or she lives is relatively stable, and at the same time must have a level of education;

Or/and, the voice recording also includes voice collection by a voice collector, and the voice collection adopts two methods: one is reading aloud with a prompt text, the prompt is a Chinese text material, and the speaker converts it into The other is natural speech, where the speaker uses prompts to tell folk stories, ethnic living conditions and humming of local folk songs.

Preferably, the acoustic parameter analysis of the voice file described in step 14 includes the voice annotation of the voice library, and the basic voice annotation includes the segmentation and alignment of the consonants and vowels of each syllable, and the labeling of the sound and rhyme tone, including two parts:

The first part is text labeling, Chinese characters + pinyin means phonetic transliteration, and the voice information is recorded in Chinese characters, so as to provide it to the recognition system and provide materials for linguistic research; text labeling must indicate basic text information and paralinguistics Phenomena, the paralinguistic phenomena in the basic notation can be represented by general paralinguistic symbols;

The second part is syllable labeling. Standard Mandarin syllable labeling is used for standard Mandarin syllable labeling, and syllable labeling is tone labeling; 0 in the tone labeling indicates soft tone, 1 indicates Yinping, 2 indicates Yangping, 3 indicates Shangsheng, and 4 indicates Qusheng.

Preferably, the acoustic parameter analysis of the voice file described in step 14 also includes the extraction of acoustic parameters;

First, the recorded voice is segmented and the silent segment is eliminated to ensure that the object of analysis is a single word, phrase, sentence, and dialogue; Marking; finally, the corresponding fundamental frequency and formant acoustic analysis parameter data are obtained according to the autocorrelation algorithm.

Preferably, the establishment of the database management system described in step 15 includes the selection of the database, and the sql database management system that is easier to implement is selected for use;

Or/and, four kinds of materials need to be stored in the establishment of the database management system described in step 15: the one, the speaker attribute material; The third is the actual voice data material, which is used to save the original parameters of the recorded voice waveform graphics; the fourth is the acoustic analysis parameter data, that is, the storage of the acoustic parameters extracted from the processed voice waveform.

This big data voice classification method based on Hadoop platform has the following beneficial effects:

(1) The big data voice classification method based on the Hadoop platform of the present invention solves technical problems such as time-consuming big data voice classification and poor real-time performance. The big data speech classification mechanism based on Hadoop platform is used to take advantage of cloud computing technology to obtain the ideal big data speech classification result as the goal, which overcomes the disadvantages of the current speech classification mechanism and greatly shortens the time of speech classification , the classification speed can adapt to the online requirements of big data speech classification, and the overall effect of speech classification is obviously better than other current speech classification mechanisms.

(1) This invention aims to study the speech characteristics of patients with motor dysarthria caused by nervous system diseases. Relying on the advantages of the open network platform, it can realize the measurement covering large-scale groups and the collection of related information, and realize the The establishment of sound databases such as the voice of healthy people and the voice of patients, and on this basis, the establishment of a lexicon to meet the diagnosis of patients with motor dysarthria.

(2) With the continuous expansion of the voice database, the present invention finally establishes a rich data resource center based on information such as Mandarin, dialects, different medical histories, and different illnesses, and provides a way for patients with neurological diseases to self-diagnose on the Internet. It can also assist Doctors conduct clinical diagnosis and treatment, and provide a rich and accurate data platform for the quantification of neurological diseases.

(3) On the basis of the voice library, the present invention uses the Map function to subdivide the big data voice classification problem based on the Hadoop platform, uses multiple nodes to solve the voice classification problem in parallel and distributed manner, and obtains the corresponding voice classification result ; Finally, the Reduce function is used to combine the speech classification results of the sub-problems to meet the online requirements of big data speech classification.

Description of drawings

Figure 1: An example of voice annotation of "bao" in the embodiment of the present invention.

Figure 2: Formant data of "bao" speech in an embodiment of the present invention.

Fig. 3: the basic frame of Hadoop platform in the embodiment of the present invention.

Fig. 4: The present invention is based on the big data speech classification process of Hadoop platform.

Detailed ways

Below in conjunction with Fig. 1 to Fig. 4, the present invention will be further described:

The speech library is composed of unvoiced sound library, voiced sound library, tone library, speech synthesis program, and Chinese-Pinyin conversion program.

1. Establishment of voiceless library:

According to the characteristics of unvoiced sounds, in order to improve the quality of synthesized speech. The voiceless library is established by direct sampling method. That is, the unvoiced part in front of the voiced segment in various pinyin combinations is sampled to form the unvoiced bank. Since the unvoiced sounds in one syllable actually only account for a very small part, the actual storage space occupied by the unvoiced sounds library composed of unvoiced sounds extracted from more than 400 atonal syllables is very small.

2. Establishment of voiced sound library:

The voiced sound is synthesized by calling the VTFR corresponding to the voiced sound by the voiced sound synthesis program. The voiced sound library is actually composed of VTFRs of various voiced sounds. The VTFRs of various voiced sounds are sequentially extracted by using the extraction VTFR program, and the voiced sound library is formed by storing the VTFRs of various voiced sounds and the voiced sound synthesis program in one data package. The actually extracted VTFR is only one curve, and the space occupied by the voiced sound library formed in this way is very small.

The establishment of the speech corpus of the present invention mainly includes the following four main processes: design of pronunciation text; speech recording; parameter analysis of speech files; establishment of a database management system; data analysis of big data technology.

1. Design of pronunciation text;

1.1 Selection of pronunciation text:

How to select corpus is the key to corpus construction. In order to ensure the orderly and effective construction of the database and the quality of the corpus, before the construction of the corpus, the selection principles of the corpus must first be studied and formulated. The selection principles of this speech corpus include: 1. The single characters in the corpus are required to include all phonetic phenomena as much as possible, so it can better and more conveniently reflect the phonological characteristics of the dialect speech; 2. The vocabulary in the corpus is based on the commonly used table in Chinese survey Therefore, it can be easily compared with Mandarin Chinese; 3. The sentences in the corpus are mainly selected from the spoken corpus! Therefore, it is more in line with the real situation faced by speech recognition; 4. The sentences in the corpus are both content and semantic. It is complete, so it can reflect the prosodic information of a sentence as much as possible; 5. We do not classify triphones, which can effectively solve the problem of sparse training data.

1.2 Preparation of pronunciation text:

The compilation of the pronunciation text is one of the key links in the establishment of the speech database. When we determine the pronunciation material, we follow the selection principle of the pronunciation text, including five parts: one is the word part. The initials, finals and tones listed in the survey word list are used as the main recording materials for this speech database; the second is the vocabulary part. Based on but not limited to a 4,000-word vocabulary, record related words according to the original conclusions about related phonology, and strive to fully reflect their phonetic characteristics, including the characteristics of sound quality and super-quality sound, and aim at some very distinctive phonetic phenomena , you can add example words to reflect its characteristics; the third is the sentence material part, the number of corpus is determined according to the language mastery of different speakers. When selecting, it is necessary to ensure that the scope of the corpus is as wide as possible, and it must also be representative to a certain extent; The fourth is the natural dialogue part, with daily life as the topic. It adopts the form of answering questions and free conversation, and records about half an hour of audio materials from the speaker.

2. Voice recording;

2.1 Determination of speaker:

The selection principle of speakers is to select native speakers who are articulate, moderate in speed, proficient in using the local language, and willing to actively cooperate with the survey. They must also ensure that their language environment is relatively stable, and at the same time, they must have a certain level of education.

2.2 Voice collection:

The way you speak during the recording directly determines the purpose of the voice bank. Due to the particularity of the collected corpus, according to different research purposes, two methods are used: one is reading aloud with prompt text, and the prompt is written in Chinese! The speaker converts it into his native language and reads it aloud; the other is Natural voice, the speaker can use prompts to tell folk stories, ethnic living conditions and humming of local folk songs, etc.

3. Analysis of the parameters of the voice file:

After recording the pronunciation text, it is necessary to analyze and process the voice data to obtain different characteristics of the voice signal. This is the key to the design of the voice corpus and the necessary basis for the later voice processing. The present invention focuses on the study of voice information, so it is necessary to mark the basic attributes of the voice signal waveform and extract relevant acoustic parameters at the same time.

3.1 Information labeling of the voice library:

Speech annotation uses Praat software, and refers to the Chinese speech segment annotation system SAMPA-C for hierarchical annotation. The annotation of the speech database includes two parts: text annotation and tonal syllable annotation. Here, the pronunciation "bao" is taken as an example, as shown in Figure 1.

The first part is text labeling, Chinese characters + pinyin means phonetic transliteration, and the voice information is recorded in Chinese characters, so that it can be used by the recognition system, and can also provide materials for linguistic research. Text annotation must indicate basic text information and paralinguistic phenomena, and the paralinguistic phenomena in basic annotations can be represented by general paralinguistic symbols.

The second part is the syllable labeling. The standard Mandarin syllable labeling is adopted for the standard Mandarin syllable labeling, and the syllable labeling is tonal labeling. In the tone label, 0 means soft tone, 1 means Yinping tone, 2 means Yangping tone, 3 means Shang tone, and 4 means Qu tone.

3.2 Extraction of acoustic parameters:

For the recorded speech signal, it is also necessary to extract the acoustic parameters of each segment. In actual operation, the recorded speech is first segmented and the silent segment is eliminated to ensure that the object of analysis is a single word; In the voice waveform data, the start and end segments of the voice signal are judged, and the range of finals is marked; finally, the corresponding fundamental frequency and formant data are obtained according to the autocorrelation algorithm. Taking the voice "bao" as an example, as shown in Figure 2.

4. Establishment of database management system:

4.1 Selection of database

For the choice of database, because in the voice library, a large amount of voice waveform data needs to be stored, which is characterized by a large amount of data, variable length, and low requirements for transaction processing and recovery, security, and network support. Therefore, we can choose the sql database management system which is easier to implement.

4.2 Establishment of database management system

The establishment of the database management system in the speech corpus needs to store four kinds of materials: one is the speaker’s attribute material, such as the speaker’s age, gender, education, mastery of Chinese, and the use of the mother tongue by the person; the other is the pronunciation text material, Input and store the pronunciation material of the speaker and its corresponding dialect pronunciation and Mandarin International Phonetic Alphabet and other text materials; the third is the actual voice data material, which is mainly used to save the original parameters of the recorded voice waveform graphics; the fourth is the acoustic analysis parameter data, namely Save the acoustic parameters extracted from the processed speech waveform.

5. Data analysis of big data technology

Big data is a collection of data that is so large that it exceeds the capabilities of traditional database software tools in terms of acquisition, storage, management, and analysis. great feature. The strategic significance of big data technology lies not in mastering huge data information, but in professional processing of these meaningful data. In other words, if big data is compared to an industry, the key to achieving profitability in this industry lies in improving the "processing capability" of data and realizing the "value-added" of data through "processing". In the construction of the lexicon, the important value of using big data technology is to achieve the purpose of evaluating the pros and cons of phonetic elements in the lexicon through targeted analysis and research on the data, so as to make the lexicon construction more perfect.

The lexicon is shared through the network platform to facilitate the testing of different groups of people. At the same time, more data samples will be obtained to enrich the voice library. In the future, more targeted sports articulation can be established according to different regions and dialects. The lexicon of patients with disorders provides more abundant and reliable data samples for subsequent automatic identification of disease classification and grading.

As shown in Figure 3, a speech classification mechanism based on the Hadoop platform is proposed. Firstly, a large number of images are collected, an image database is constructed, and effective features of image classification are extracted. Then, based on the Hadoop platform, the Map function is used to solve the big data speech classification problem. Subdivision, use multi-node parallel and distributed to solve the sub-problems for speech classification, and obtain the corresponding speech classification results; finally use the Reduce function to combine the speech classification results of the sub-problems to meet the online requirements of big data speech classification.

As shown in Figure 4, the big data voice classification process based on the Hadoop platform, the specific steps are as follows:

(1) Client submits a voice classification task to the Job Tracker of the Hadoop platform, and the Job Tracker copies the voice feature data to the local distributed file processing system;

(2) Initialize the task of speech classification, put the task into the task queue, and the Job Tracker assigns the task to the corresponding node according to the processing capabilities of different nodes, that is, the Task Tracker;

(3) Each Task Tracker uses a support vector machine to fit the relationship between the speech features to be classified and the speech feature library according to the assigned tasks, and obtains the corresponding category of the speech;

(4) Save the corresponding voice category as Key/Value to the local file disk;

(5) If the Key/Value of the intermediate results of speech classification are the same, they are merged, and the merged result is handed over to Reduce for processing to obtain the result of speech classification, and the result is written into the distributed file processing system;

(6) Job Tracker clears the task status, and the user gets the result of voice classification from the distributed file processing system.

Above, the present invention has been exemplarily described in conjunction with the accompanying drawings. Obviously, the realization of the present invention is not limited by the above-mentioned manner, as long as various improvements of the method concept and technical solutions of the present invention are adopted, or the present invention is implemented without improvement. The ideas and technical schemes directly applied to other occasions are within the protection scope of the present invention.

Claims (9)

1. A big data voice classification method based on Hadoop platform, comprising the following steps: Step 1, carry out the construction of voice library; Step 2. On the basis of this voice library, based on the Hadoop platform, the Map function is used to subdivide the big data voice classification problem, and the sub-problems are solved in a parallel and distributed manner with multiple nodes to obtain the corresponding voice classification results; Step 3. Finally, use the Reduce function to combine the speech classification results of the sub-problems to meet the online requirements of big data speech classification. 2. the big data voice classification method based on Hadoop platform according to claim 1, is characterized in that: described step 2 comprises the following content: (1) Client submits a voice classification task to the Job Tracker of the Hadoop platform, and the Job Tracker copies the voice feature data to the local distributed file processing system; (2) Initialize the task of speech classification, put the task into the task queue, and the Job Tracker assigns the task to the corresponding node according to the processing capabilities of different nodes, that is, the Task Tracker; (3) Each Task Tracker uses a support vector machine to fit the relationship between the speech features to be classified and the speech feature library according to the assigned tasks, and obtains the corresponding category of the speech; (4) Save the corresponding voice category as Key/Value to the local file disk; (5) If the Key/Value of the intermediate results of speech classification are the same, they are merged, and the merged result is handed over to Reduce for processing to obtain the result of speech classification, and the result is written into the distributed file processing system; (6) Job Tracker clears the task status, and the user gets the result of voice classification from the distributed file processing system. 3. the big data voice classification method based on Hadoop platform according to claim 2, is characterized in that: Step 1, carrying out the construction of voice bank comprises the following steps: step 11, the design of pronunciation text; Step 12, voice recording; Step 13, the mark of voice file; Step 14, the acoustic parameter analysis to voice file; Step 15, database management System establishment. 4. the big data voice classification method based on Hadoop platform according to claim 3, is characterized in that: The design of pronunciation text in the described step 11 comprises the selection of pronunciation text, and the selection principle of the corpus of described pronunciation text comprises following one or more: a. The words in the corpus are required to include all the phonetic phenomena as much as possible, which can better and more conveniently reflect the phonological characteristics of the voices of different patients; b. The vocabulary in the corpus is based on the common Chinese survey table, so it can be easily compared with Mandarin Chinese; c. The sentences in the corpus are mainly obtained from conversations with patients based on several related topics, so they are more in line with the real situation faced by speech recognition; d. The sentences in the corpus are complete in content and semantics, so it can reflect the prosodic information of a sentence as much as possible; e. The selection of triphones without classification can effectively solve the problem of sparse training data. 5. the big data voice classification method based on Hadoop platform according to claim 4, is characterized in that: The design of the pronunciation text described in the step 11 also includes the preparation of the pronunciation text, and the compilation principle of the pronunciation text includes the following one or more: a. Character part: use the initials, finals, and tones listed in the survey word list as the main recording materials for this speech database; b. Vocabulary part: Based on but not limited to a 4,000-word vocabulary, record related words according to the original conclusions about related phonology, and strive to fully reflect its phonetic characteristics, including sound quality and super-tone quality characteristics, for some very For characteristic phonetic phenomena, example words can be added to reflect its characteristics; c. Sentence material part: the number of corpus is determined according to the language mastery of different speakers. When selecting, it is necessary to ensure that the range of corpus is as wide as possible and make it representative; d. Natural dialogue part: Daily life as the topic, in the form of answering questions and free conversation, recording 20-40 minutes of audio materials from the speaker, involving vocabulary in daily spoken language that is different from Mandarin, and requiring the speaker to speak in dialect. 6. the big data voice classification method based on Hadoop platform according to claim 5, is characterized in that: The voice recording in step 12 includes the determination of the speaker. The selection principle of the speaker is to select a native speaker who is articulate, moderate in speed, proficient in using the local language and willing to actively cooperate with the investigation. The language environment is relatively stable, and at the same time, there must be a level of education; Or/and, the voice recording also includes voice collection by a voice collector, and the voice collection adopts two methods: one is reading aloud with a prompt text, the prompt is a Chinese text material, and the speaker converts it into The other is natural speech, where the speaker uses prompts to tell folk stories, ethnic living conditions and humming of local folk songs. 7. the big data voice classification method based on Hadoop platform according to claim 6, is characterized in that: The acoustic parameter analysis of the voice file described in step 14 includes the voice annotation of the voice library, and the basic voice annotation includes the segmentation and alignment of the consonants of each syllable, and the labeling of the sound and tone, including two parts: The first part is text labeling, Chinese characters + pinyin means phonetic transliteration, and the voice information is recorded in Chinese characters, so as to provide it to the recognition system and provide materials for linguistic research; text labeling must indicate basic text information and paralinguistics Phenomena, the paralinguistic phenomena in the basic notation can be represented by general paralinguistic symbols; The second part is syllable labeling. Standard Mandarin syllable labeling is used for standard Mandarin syllable labeling, and syllable labeling is tone labeling; 0 in the tone labeling indicates soft tone, 1 indicates Yinping, 2 indicates Yangping, 3 indicates Shangsheng, and 4 indicates Qusheng. 8. the big data voice classification method based on Hadoop platform according to claim 7, is characterized in that: described in the step 14 to the acoustic parameter analysis of voice file also comprises the extraction of acoustic parameter; First, the recorded voice is segmented and the silent segment is eliminated to ensure that the object of analysis is a single word, phrase, sentence, and dialogue; Marking; finally, the corresponding fundamental frequency and formant acoustic analysis parameter data are obtained according to the autocorrelation algorithm. 9. the big data voice classification method based on Hadoop platform according to claim 8, is characterized in that: the establishment of database management system described in the step 15 comprises the selection of database, selects the sql database management system that is easier to realize; Four kinds of materials need to be stored in the establishment of the database management system described in step 15: the one, the pronunciation person attribute material; It is the actual voice data material, which is used to save the original parameters of the recorded voice waveform graphics; the fourth is the acoustic analysis parameter data, that is, the storage of the acoustic parameters extracted from the processed voice waveform.

Images