WO2021073138A1

WO2021073138A1 - Audio output method and system

Info

Publication number: WO2021073138A1
Application number: PCT/CN2020/097000
Authority: WO
Inventors: 蔡继发; 宋飞豹; 倪合强; 姚寿柏
Original assignee: 苏宁易购集团股份有限公司; 苏宁云计算有限公司
Priority date: 2019-10-16
Filing date: 2020-06-19
Publication date: 2021-04-22
Also published as: CN110880316A; CA3158353A1

Abstract

An audio output method and system. The method comprises the following steps: obtaining continuous voice information collected by a client, segmenting the voice information, decoding the segmented voice information, converting the voice information into first text information, and storing the first text information (S1); filtering the first text information, then extracting a keyword, marking the keyword according to the number of times that the keyword appears, and when the number of times that the same keyword appears exceeds a preset threshold and the keyword is not in a standard word library, importing the keyword into the standard word library (S2); using a machine learning method to search for the extracted keyword and then perform matching, enabling matched characters containing the keyword to form second text information, and arranging the second text information (S3); and converting the second text information into a corresponding standard audio according to a priority order, and outputting the standard audio (S4). The present invention can recognize audio information, rapidly and accurately output processed audio information, and improve the effect of language expression.

Description

Audio output method and system

Technical field

The invention relates to the field of speech recognition, in particular to an audio output method and system.

Background technique

Audio is a necessary means to promote effective interaction between people and between people and the sounding body. Most of the existing knowledge, information, interactive entertainment, emotional expression, etc. use audio as a medium for information transmission, and these are all manifested without exception. The audio has the characteristics of efficient communication and easy acceptance. With the development of the information age, audio needs to carry more and more information. At this time, it is necessary to interact with this audio in some way. However, for some people, it is still somewhat difficult to convey what is said clearly and accurately. For example, some people have limited English proficiency, and may have substandard pronunciation and repeated words. The spoken English may be as good as possible. Less than the inner request. For example, for some people, due to problems such as typos and too fast speaking, it may be difficult for the listeners to understand this way of speaking, or even cause misunderstandings, resulting in the omission of voice information or The loss of key points is even the duplication of useless information.

The existing audio prompt method promotes interaction through text methods such as teleprompter, but its method lacks the fluency of interaction, does not achieve the rhythmic effect of language expression, and does not get the expected interactive feedback. To a certain extent Said, even the transmission of effective information will be lost. Therefore, how to provide an audio output method with higher output efficiency and improve the interactive experience between users has become an urgent problem in voice output.

Summary of the invention

In order to solve the problems of the prior art, the embodiments of the present invention provide an audio output method and system, which can automatically recognize the sent audio information intelligently, and quickly and accurately output the processed audio prompt information to improve language expression Rhythm effect.

In order to solve the above technical problems, the technical solutions adopted by the present invention are:

In the first aspect, an embodiment of the present invention provides an audio output method, including the following steps:

Acquire continuous voice information collected by the client, segment the voice information, decode the segmented voice information, and then convert the cut voice information into first text information and store it ；

The first text information is filtered, and then keywords are extracted. The keywords are marked according to the number of occurrences. When the number of times the same keyword is marked exceeds a preset threshold and the keyword is not in In the standard lexicon, import the keywords into the standard lexicon;

Using the method of machine learning, the extracted keywords are searched in the standard vocabulary and then matched, the matched words containing the keywords are formed into second text information, and the second text information is performed Prioritization of relevance;

According to the order of priority, the second text information is converted into corresponding standard audio and output.

Further, the segmentation of the voice information includes the following steps: the voice information is converted into voice coded data by pulse code modulation according to a sequence of timestamps, and two consecutive segments of the voice coded data of a fixed time length are similarly processed. Degree detection, marking the most suitable end point of the repetitive splicing area of the voice coded data, and cutting and filtering the repetitive data of the voice coded data.

Further, the filtering of the first text information includes at least: recognition of modal particles and filtering of repeated words.

Further, a preferred storage area is provided in the standard vocabulary, and the preferred storage area is used to store the candidate association content with the highest priority, and the candidate association content corresponds to the key whose number of occurrences exceeds a preset threshold. word.

Further, before matching the keyword, it also includes processing the keyword through a fuzzy algorithm, and searching for the keyword includes searching for synonyms, homonyms, and low-proportion typos of the keyword; Prioritization methods include at least Bayesian decision-making methods.

Further, the output of the standard audio corresponds to the application sites of different clients. When it is a single-sided client, the output is at least three optional standard audios arranged according to the degree of relevance; when it is a double-sided client When, the output is the standard audio with the highest degree of relevance.

On the other hand, an embodiment of the present invention also provides an audio output system, including:

The collection and input module is used to obtain the continuous voice information collected by the client, segment the voice information, decode the voice information after the segmentation, and then convert the voice information after the segmentation into And store the first text information;

The keyword processing module is used to filter the first text information, then extract keywords, and mark the keywords according to the number of occurrences, when the number of times the same keyword is marked exceeds a preset threshold And the keyword is not in the standard thesaurus, then the keyword is imported into the standard thesaurus;

The search association module is used for matching the extracted keywords in the standard dictionary through the method of machine learning, and composing the text containing the keywords after the matching into the second text information. The second text information is arranged in priority order of relevance;

The audio output module is used for converting the second text information into corresponding standard audio and outputting it in the order of priority.

Further, the acquisition and input module includes a voice cutting unit, the voice cutting unit is used to convert the voice information into voice coded data through pulse code modulation according to a sequence of time stamps. Similarity detection is performed on the voice coded data, the repetitive splicing area of the voice coded data is marked with the most suitable endpoint, and the repetitive data of the voice coded data is cut and filtered.

Further, the keyword processing module includes a priority storage unit for setting a preferred storage area in the standard thesaurus, the preferred storage area is used to store the candidate association content with the highest priority, the candidate association The content corresponds to the keyword whose number of appearances exceeds a preset threshold.

Further, the search association module includes an intelligent search unit and a prioritization unit. The search for the keyword by the intelligent search unit includes searching for synonyms, homonyms, and low-proportion typos of the keyword. The priority ranking performed by the prioritizing unit includes at least a Bayesian decision method.

The beneficial effects brought about by the technical solutions provided by the embodiments of the present invention are:

The audio output method and system disclosed in the embodiments of the present invention first collect the user's voice information, and after the collected voice information is segmented and converted into the first text information according to the sequence of time stamps, The first text information is filtered and keywords are extracted, and the keywords are used to search for high-relevance candidate association content in the standard thesaurus. At the same time, the method of machine learning is used to search for high-frequency keywords. Reinforcement learning is performed so that the standard vocabulary can continuously store the candidate association content corresponding to the high priority, and the matching text containing the keyword is formed into the second text information through the Bayesian decision method, and the second text information is The text information is prioritized by the degree of relevance, and then the second text information is converted into the corresponding standard audio according to the order of priority, so that the application scenarios are different, and the final input is similar to the original voice information sent by the user. Audio. The embodiment of the present invention can avoid the situation of inaccurate pronunciation and wrong pronunciation of users in the process of communicating through the terminal, and can also intelligently associate the voice that the audio sender wants to express through the method of machine learning and the extraction of keywords. Through the machine learning of communication methods, the interactive experience of users when communicating is improved. In addition, the implementation of the audio output method in the embodiment of the present invention is simple, fast, and highly flexible, and has a wide range of application scenarios.

Description of the drawings

In order to more clearly describe the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained from these drawings without creative work.

FIG. 1 is a schematic flowchart of an audio output method provided by an embodiment of the present invention;

FIG. 2 is a schematic diagram of the process of searching the lexicon in the audio output method provided by the embodiment of the present invention;

Fig. 3 is a schematic structural diagram of an audio output system provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only A part of the embodiments of the present invention, but not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

Example one:

As shown in FIG. 1, an embodiment of the present invention provides an audio output method, which includes the following steps:

S1: Acquire continuous voice information collected by the client, segment the voice information, decode the segmented voice information, and then convert the segmented voice information into first text information And store

S2: Filter the first text information, then extract keywords, mark the keywords according to the number of occurrences, when the number of times the same keyword is marked exceeds a preset threshold and the keywords If it is not in the standard dictionary, then import the keywords into the standard dictionary;

S3: Using the method of machine learning, the extracted keywords are searched in the standard vocabulary and then matched, the matched words containing the keywords are formed into second text information, and the second text Prioritize the relevance of information;

S4: Convert the second text information into corresponding standard audio and output it in the order of priority.

Specifically, in this embodiment, the user's voice information is first collected through the client, and the collected voice information is segmented and converted into the first text information according to the sequence of time stamps, and then the first text information is Perform filtering and keyword extraction, use the keywords to search for high-relevance candidate association content in the standard thesaurus, and at the same time, use machine learning methods to perform reinforcement learning on high-frequency keywords, so that the The standard thesaurus can continuously store the candidate association content corresponding to the high priority, and use the Bayesian decision method to compose the matched text containing the keyword into the second text information, and prioritize the relevance of the second text information. The second text information is converted into the corresponding standard audio according to the priority order, so that the application scenarios are different, and the final input is the standard audio similar to the voice information originally sent by the user. The embodiment of the present invention can avoid the situation that the user's pronunciation is inaccurate or wrong in the process of communicating through the terminal, and through the machine learning of the communication mode, the interaction experience of the user when communicating is improved. In addition, the implementation of the audio output method in the embodiment of the present invention is simple, fast, and has a wide range of application scenarios.

When collecting audio data from the client, the length of the audio data is fixed. For example, the length of each piece of collected audio is 15 seconds, but the single sampling interval is fixed at 10 seconds, so that there will be 5 seconds for each piece of collected audio In the audio overlap area of the clock, in order to eliminate the audio in the overlap area, the voice information corresponding to the audio needs to be segmented and cut. Preferably, the segmentation of the voice information includes the following steps: The sequence of stamps is converted into speech coded data by pulse code modulation, the similarity of the speech coded data of two consecutive fixed time lengths is detected, the repetitive splicing area of the speech coded data is marked with the most suitable endpoint, and the selection is cut and filtered. The repeated data of the speech coded data.

Specifically, the standard vocabulary includes output information corresponding to all keywords; preferably, a preferred storage area is provided in the standard vocabulary, and the preferred storage area is used to store the candidate association content with the highest priority. The candidate association content corresponds to the keyword whose number of appearances exceeds a preset threshold. In the process of searching for keywords, multiple matching items in the thesaurus are searched according to keywords, and several matching items with a high degree of relevance are used as final output options; for the keyword processing and searching in steps S2 and S3 The process can be referred to simply as the process of keyword search and decoding. Among them, in the keyword search and decoding, a decoding buffer is set to record the decoding results with more occurrences of the keyword, and at the same time increase the relevance value of the selected alternative.

Figure 2 is a further description of step S3, which specifically includes the following steps:

S301: Search the filtered word segmentation for the keywords in the thesaurus, and determine whether the part of speech and frequency of the non-existent word segmentation are greater than three to determine whether to add it into the thesaurus;

S302: Use deep reinforcement learning to perform similar keyword matching for word segmentation, including matching functions such as synonyms, homophone typos, and low-proportion typos correction;

S303: According to the Bayesian decision mechanism, count the corresponding content and the time of occurrence of the high-frequency keywords, and formulate the priority search order;

S304: According to the priority rule, output the alternative content corresponding to the higher priority in sequence.

Preferably, the filtering of the first text information includes at least: recognition of modal particles and filtering of repeated words. In the keyword search and decoding, the decoding buffer also includes user-accustomed expression terms, such as expression terms repeated multiple times (three times or more) consecutively in the audio. In addition, in the keyword search and decoding, the correlation degree in the entire thesaurus is arranged from high to low as: decoding buffer, keyword search, and near (synonymous) word search. In the keyword search and decoding, there is a unified adjustment process for non-standard (three consecutive times) repeated audio and invalid audio intervals, without uploading. For example, the recognized text results are processed by word segmentation, for some existing modal particles, such as "umhhhhh", "ahhhhh", "ohhhhh", "yahhhh", etc., as well as single repeated words: " trekking trip "" Yes, yes "," ha ha ha "," la la la "and so on, as well as some of the statements replied:" Yes Yes Yes "," good good good "," right Right, right, etc., as well as colloquial words: "that that that", "six six six", "oh ye oh ye oh yeah", etc. These words are filtered and not entered as keyword searches.

Specifically, before matching the keyword, it also includes processing the keyword through a fuzzy algorithm, and searching for the keyword includes synonyms, homophones, and low-proportion typos of the search keyword; the priority The arrangement method includes at least Bayesian decision method. Among them, in the keyword search and decoding, define the synonyms and synonyms of the current audio content to match the keywords in the thesaurus, at this time the priority is reduced by one level; in the keyword search and decoding, intelligent fuzzy matching of the recognized content , For example, "Once bitten by a snake" is recognized as "being bitten by a snake once", and "four, four, five" is recognized as "rounding", etc.;

Preferably, the output of the standard audio corresponds to the application places of different clients. When it is a single-sided client, the output is at least three optional standard audios arranged according to the degree of relevance; when it is a double-sided client When, the output is the standard audio with the highest degree of relevance. When the corresponding client is the same device, it is a single-sided client at this time, and the output of the client can be selected because the input and input at this time are the same audio sender. Specifically, for the output alternative text, under the constraints of the Bayesian decision mechanism, the alternative text is sorted by priority, and the priority is defined in order from high to low: the content of the alternatives that appear more frequently and have been selected ( The first item), the content of the current keyword search alternative (the first item), the content of the alternative item corresponding to the subsequent keyword (the first item), the content of the alternative item that appears frequently and has been selected (the second item), the current Keyword search for alternative content (the second item), subsequent keywords corresponding to the alternative content (second item), ..., and so on; then filter out repeated alternative texts, and comprehensively judge the first three alternative texts with the highest relevance , Generate the corresponding audio from these three alternative texts, for example, each audio length does not exceed 5s (otherwise only the audio before 5s), and then convert the audio into MP3 format audio, and then transmit it after further processing To the client, if the similarity between the audio of the time stamp and the alternative audio reaches a certain threshold, all alternative audio data with the same time stamp as the alternative audio are cleared at this time. In the case of two different clients, that is, when the audio input terminal is the audio sender and the audio output terminal is the audio receiver, at this time, there is only one standard with the highest degree of relevance for the audio sending Audio, the standard audio output through the client at this time is the input audio of the audio sender by default.

Embodiment two:

As shown in FIG. 3, an embodiment of the present invention also provides an audio output system, including:

The collection and input module 1 is used to obtain the continuous voice information collected by the client terminal 5, segment the voice information, decode the voice information after the segmentation, and then convert the voice information after the segmentation Convert it into the first text message and store it;

The keyword processing module 2 is used for filtering the first text information, and then extracting keywords, marking the keywords according to the number of occurrences, when the number of times the same keyword is marked exceeds a preset Threshold and the keyword is not in the standard dictionary, then import the keyword into the standard dictionary;

The search association module 3 is used for using the method of machine learning to match the extracted keywords in the standard vocabulary, and compose the matched words containing the keywords into the second text information, Prioritizing the degree of relevance of the second text information;

The audio output module 4 is used to convert the second text information into a corresponding standard audio and output it in the order of priority.

Preferably, the acquisition and input module 1 includes a voice cutting unit, which is used to convert the voice information into voice coded data through pulse code modulation according to a sequence of time stamps. Perform similarity detection on the voice coded data, mark the repetitive splicing area of the voice coded data with the most suitable endpoint, and cut and filter the repetitive data of the voice coded data. For example, when collecting audio data transmitted from the client, the audio data length is fixed. For example, the length of each piece of collected audio is 15 seconds, but the single sampling interval is fixed at 10 seconds, so that there will be some audio data for each piece of collected audio. In the audio overlap area of 5 seconds, in order to eliminate the audio in the overlap area, the voice information corresponding to the audio needs to be segmented.

Preferably, the keyword processing module 2 includes a priority storage unit for setting a preferred storage area in the standard vocabulary, and the preferred storage area is used to store the candidate association content with the highest priority. The associative content corresponds to the keyword whose appearance count exceeds a preset threshold. Among them, in the process of searching for keywords, multiple matching items in the thesaurus are searched according to keywords, and several matching items with a high degree of relevance are used as final output options; in addition, in the keyword search and decoding, A decoding buffer is set in the priority storage unit, and the decoding results with more occurrences of keywords are recorded, and at the same time, the relevance value of the selected alternatives is improved.

Preferably, the keyword processing module 2 for filtering the first text information includes at least: recognition of modal particles and filtering of repeated words. When the keyword processing module 2 searches and decodes keywords, there is a unified adjustment process for non-standard (three consecutive times) repeated audio and invalid audio intervals without uploading. For example, the recognized text result is processed through word segmentation, for some existing modal particles, such as "umhhhh", "ahhhhh", "ohhhhh", "yahyah", etc., as well as single repeated words: trekking trip "" Yes, yes "," ha ha ha "," la la la "and so on, as well as some of the statements replied:" Yes Yes Yes "," good good good "," right Right, right, etc., as well as colloquial words: "that that that", "six six six", "oh ye oh ye oh yeah", etc. These words are filtered and not entered as keyword searches.

Preferably, the search association module 3 includes an intelligent search unit and a prioritization unit. The intelligent search unit is used to match the keywords, including searching for synonyms, homophones, and low-proportion words of the keywords. A typo, the priority ranking performed by the prioritizing unit includes at least a Bayesian decision method. Specifically, before matching the keyword by the search association module 3, it also includes processing the keyword through a fuzzy algorithm, and searching for the keyword includes synonyms, homonyms, and low-proportion typos of the keyword. ; Among them, in the keyword search and decoding, define synonyms and synonyms that identify the current audio content to match the keywords in the thesaurus, at this time the priority is reduced by one level; in the keyword search and decoding, the recognition content is intelligently fuzzy Matching, for example, "Once bitten by a snake" is recognized as "Snake bitten once", "Four, Four, and Five" is recognized as "Rounding" and so on.

A structural schematic diagram of the audio output system shown in Figure 3, the client 5 in the figure is the same device, where, for different applications, the client 5 can also be set as two different devices, one It is the audio input terminal and the other is the audio output terminal. These two clients are held by two different users, so that the two clients can directly communicate directly, which can prevent users from communicating through the terminal. Inaccurate pronunciation, incorrect pronunciation in the situation, through the machine learning of the communication method, to enhance the user's interactive experience when communicating. In the case of a single-sided client, the output is at least three optional standard audios arranged according to the degree of relevance; in the case of a double-sided client, the output is the standard audio with the highest degree of relevance. When the corresponding client is the same device, it is a single-sided client at this time, and the output of the client can be selected because the input and input at this time are the same audio sender. Specifically, for the output alternative text, under the constraints of the Bayesian decision mechanism, the alternative text is sorted by priority, and the priority is defined in order from high to low: the content of the alternatives that appear more frequently and have been selected ( The first item), the content of the current keyword search alternative (the first item), the content of the alternative item corresponding to the subsequent keyword (the first item), the content of the alternative item that appears frequently and has been selected (the second item), the current Keyword search for alternative content (the second item), subsequent keywords corresponding to the alternative content (second item), ..., and so on; then filter out repeated alternative texts, and comprehensively judge the first three alternative texts with the highest relevance , Generate the corresponding audio from these three alternative texts, for example, each audio length does not exceed 5s (otherwise only the audio before 5s), and then convert the audio into MP3 format audio, and then transmit it after further processing To the client terminal 5, if the similarity between the audio of the time stamp and the alternative audio reaches a certain threshold, all alternative audio data with the same time stamp as the alternative audio are cleared at this time. In the case of two different clients, that is, when the audio input terminal is the audio sender and the audio output terminal is the audio receiver, at this time, there is only one standard with the highest degree of relevance for the audio sending Audio, the standard audio output through the client at this time is the input audio of the audio sender by default.

All the above optional technical solutions can be combined in any way to form an optional embodiment of the present invention, which will not be repeated here.

It should be noted that the audio output system provided in the above embodiment only uses the division of the above functional modules for example when the audio is output. In actual applications, the above functions can be allocated by different functional modules according to needs. , The internal structure of the audio output system is divided into different functional modules to complete all or part of the functions described above. In addition, the audio output system provided in the foregoing embodiment and the audio output method embodiment belong to the same concept. For the specific implementation process, please refer to the method embodiment, which will not be repeated here.

A person of ordinary skill in the art can understand that all or part of the steps in the above embodiments can be implemented by hardware, or by a program to instruct relevant hardware. The program can be stored in a computer-readable storage medium. The storage medium mentioned can be a read-only memory, a magnetic disk or an optical disk, etc.

The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection of the present invention. Within range.

Claims

An audio output method, characterized in that it comprises the following steps:

Acquire continuous voice information collected by the client, segment the voice information, decode the segmented voice information, and then convert the cut voice information into first text information and store it ；

The first text information is filtered, and then keywords are extracted. The keywords are marked according to the number of occurrences. When the number of times the same keyword is marked exceeds a preset threshold and the keyword is not in In the standard lexicon, import the keywords into the standard lexicon;

Using the method of machine learning, the extracted keywords are searched in the standard vocabulary and then matched, the matched words containing the keywords are formed into second text information, and the second text information is performed Prioritization of relevance;

According to the order of priority, the second text information is converted into corresponding standard audio and output.
The audio output method according to claim 1, wherein the segmentation of the voice information includes the following steps: the voice information is converted into voice coded data by pulse code modulation according to a sequence of timestamps, and continuous Perform similarity detection on two segments of the voice coded data with a fixed length of time, mark the repetitive splicing area of the voice coded data with the most suitable endpoint, and cut and filter the repetitive data of the voice coded data.
The audio output method according to claim 1, wherein the filtering of the first text information includes at least: recognition of modal particles and filtering of repeated words.
The audio output method according to claim 1, wherein a preferred storage area is provided in the standard vocabulary, and the preferred storage area is used to store the candidate association content with the highest priority, and the candidate association The content corresponds to the keyword whose number of appearances exceeds a preset threshold.
The audio output method according to claim 1, characterized in that, before matching the keyword, it further includes processing the keyword by a fuzzy algorithm, and searching for the keyword includes searching for the keyword. Synonyms, homophones, and low-proportion typos; the priority ranking method includes at least a Bayesian decision method.
The audio output method according to claim 1, wherein the output of the standard audio corresponds to the application sites of different clients, and when it is a single-sided client, the output is at least three options according to the degree of relevance. The standard audio arranged; when it is a two-sided client, the standard audio with the highest degree of relevance is output.
An audio output system, characterized in that it comprises:

The collection and input module is used to obtain the continuous voice information collected by the client, segment the voice information, decode the voice information after the segmentation, and then convert the voice information after the segmentation into And store the first text information;

The keyword processing module is used to filter the first text information, then extract keywords, and mark the keywords according to the number of occurrences, when the number of times the same keyword is marked exceeds a preset threshold And the keyword is not in the standard thesaurus, then the keyword is imported into the standard thesaurus;

The search association module is used for matching the extracted keywords in the standard dictionary through the method of machine learning, and composing the text containing the keywords after the matching into the second text information. The second text information is arranged in priority order of relevance;

The audio output module is used for converting the second text information into corresponding standard audio and outputting it in the order of priority.
The audio output system according to claim 7, wherein the collection and input module comprises a voice cutting unit, and the voice cutting unit is used to convert the voice information into voice through pulse code modulation according to a sequence of time stamps. For encoding data, similarity detection is performed on two consecutive segments of the voice coded data of a fixed time length, the repetitive splicing area of the voice coded data is marked with the most suitable endpoint, and the repetitive data of the voice coded data is cut and filtered.
The audio output system according to claim 7, wherein the keyword processing module includes a priority storage unit, configured to set a preferred storage area in the standard vocabulary, and the preferred storage area is used to store priority The candidate association content with the highest rank, and the candidate association content corresponds to the keyword whose number of appearances exceeds a preset threshold.
The audio output system according to claim 7, wherein the search association module includes an intelligent search unit and a prioritization unit, and the search for the keyword by the intelligent search unit includes a search for the keyword Synonyms, homophones, and low-proportion typos, the priority ranking performed by the prioritizing unit includes at least a Bayesian decision method.