A kind of speech retrieval method and adopt the audiovisual/visual information searching system of this method
Technical field:
The present invention relates to a kind of method of retrieving by voice, the particularly method that adopts voice to retrieve to audiovisual/visual information, and the audiovisual/visual information searching system of this method of employing.
Background technology:
, information retrieval system such as AltaVista, Baidu's search engine.Classical IR problem is that the text document that needs is located in the inquiry that utilizes a set of keyword to form, and promptly locatees key word of the inquiry in the document and finds the document that mates.If comprise more query term in the document, so, it just is considered to comprise than other document of less query term and more " is correlated with ".So document can sort according to " being correlated with " degree, and is shown to the user, so that further search.General IR process designs for text, is usually used in the retrieval of phonotape and videotape or other multimedia messagess also.But phonotape and videotape or other multimedia messagess are a kind of opaque bit streams, though audiovisual/visual information can be given attributes such as name, file layout, sampling rate, but do not have identifiable speech or comparable entity among the audiovisual/visual information, therefore can not as text, search for or retrieve the content of audiovisual/visual information inside.Attribute and description based on artificial input phonotape and videotape text are the methods of carrying out the audiovisual/visual information retrieval, and the critical defect of this method is: when data volume was more and more, artificial note intensity strengthened; The people is to the perception of phonotape and videotape, as the melody of music, tone, tonequality etc., is difficult to express clearly with the literal note.
Voice are audio frequency of a kind of specific type, and it and text can be changed mutually, and can utilize speech recognition technology to carry out the audiovisual/visual information retrieval is an advantage of audiovisual/visual information retrieval.Utilize speech recognition technology to realize people's the natural language and the technology of audio and video equipment dialogue, existing certain progress at present.For example, the Chinese patent 1288222 " speech control system " of Deutsche Thomson-Brandt GmbH's application with microphone array, 1288223 " devices that are used for voice-operated adaptation direction characteristic ", be primarily characterized in that one or more microphones are integrated in the broadcast electrical equipment that the surround sound of consumer electronics reproduces, particularly in televisor, provide one or more infrared motion detectors to detect the user position, from distance far away relatively to such as televisor, video tape recorder, consumption electronic products such as DVD player are realized the voice-activated remote control, but above-mentioned technology still can't realize directly by voice audiovisual/visual information being retrieved.
Utilizing automatic speech recognition (ASR) technology is speech conversion text, adopt text searching method to carry out the phonotape and videotape retrieval, though good continuous speech recognition system can reach the word correctness more than 90% under operating carefully, but in actual applications, as phone and news broadcast etc., discrimination is not high.When some specialized vocabularies (as name, place) were not in system's dictionary, it is poorer that speech recognition system is handled the data of the voice on a large scale recognition performance of the unrestricted theme of each side.The important symbol of speech recognition is the validity to voice control command on the analysis foundation of the aural signature of the keyword of identification content topic, semantic feature, because rich and varied, fast and expansionary restriction of variation of audio-visual product content topic utilize speech recognition technology correctly to produce steering order, correct recognition rata is reduced greatly, even wrong identification occurs.
Summary of the invention:
The present invention solves in the prior art that the audio-video content of retrieval exists is difficult to defectives such as accurate description phonotape and videotape content, correct recognition rata are low, and a kind of method that voice carry out fast, accurately retrieve audio-video content and audiovisual/visual information searching system that adopts this method of adopting is provided.
Technical scheme of the present invention is: a kind of speech retrieval method comprises following step:
1), each audio-visual product is done the classification index, the classification questions record is set, and, sets up audio-visual product content default classification bibliographic database and predetermined keyword table the content topic keyword sampling of each audio-visual product;
2), classification questions record in the described default classification bibliographic database and the keyword in the described predetermined keyword table are carried out index with a received pronunciation, set up corresponding standard phonetic classification questions record model of cognition and received pronunciation key word recognition model, described received pronunciation is first voice signal, and said process is an indexing model;
3), make up phonetic classification bibliographic database and the voice antistop list of audio-visual product to the specific user by training, building process is: the specific user contrasts described received pronunciation correspondence and sends one second voice signal, train by described second voice signal, make described second voice signal and first voice signal and user be used for controlling the 3rd voice command formation one-to-one relationship that audio-visual product broadcasts, and forming phonetic classification bibliographic database and voice antistop list based on second voice signal, said process is a training mode;
4), send the 3rd voice command by the specific user, the data model of question-type voice and received pronunciation classification questions record model of cognition and received pronunciation classification questions record model of cognition and phonetic classification bibliographic database and voice antistop list is carried out matching operation, when similarity during greater than a threshold value, the audio-visual product playing device sends control signal, and said process is a standby mode.
Above-mentioned classification index is that international decade taxonomic hierarchies classification can be notions such as relevant time of audio-visual product, publisher, broadcast unit, content body by UDC.
When above-mentioned structure specific user's audio-visual product phonetic classification bibliographic database and voice antistop list, voice control password can be set.
A kind of audiovisual/visual information searching system that realizes above-mentioned speech retrieval method, comprise voice system control module and audiovisual/visual information retrieval module, described voice system control module is electrically connected on described audiovisual/visual information retrieval module, described audiovisual/visual information retrieval module is made up of static memory and dynamic storage, and described static memory comprises:
One audio-visual data index device according to user's needs and hobby, carries out the analysis based on sound model and speech model to aural signature, the semantic feature of the content topic of various audio and video files; Automatically gather the sample of an audio-visual product content topic keyword, construct the default classification bibliographic database of the predetermined keyword table and the audio-visual product of various audio-visual product content topics in advance, and set up corresponding standard phonetic classification questions record model of cognition and received pronunciation key word recognition model by first voice signal;
One phonetic classification bibliographic database according to specific user's custom, is defined arbitrarily by specific user's voice, forms the phonetic classification bibliographic database of the audio-visual product of standard at the specific user in it;
One voice antistop list according to specific user's custom, is defined arbitrarily by specific user's voice, forms the data model of the audio-visual product of standard at specific user's voice antistop list in it;
One speech retrieval coupling subsystem is accepted specific user's voice and is putd question to; Specific user's the 3rd voice command is carried out the inspection of grammer, word etc.; Specific user's question-type is made an explanation or the processing of compiling property; The data model of question-type voice and received pronunciation classification questions record model of cognition and received pronunciation key word recognition model and phonetic classification bibliographic database and voice antistop list is carried out matching operation; Operation result is returned this voice system control module;
One speech recognition device is used to discern user's verbal speech control input command; Corresponding standard phonetic classification questions record model of cognition and received pronunciation key word recognition model are promptly set up in the foundation of first voice signal in the indexing model; The specific user responds the one second phonetic entry command signal that corresponding first voice signal sends in the training mode, sets up the voice antistop list and the phonetic classification bibliographic database of specific user's phonotape and videotape content topic; Receive in the standby mode by the retrieval of one the 3rd voice command that the specific user sent and put question to;
One dynamic storage is used for storing the 3rd voice command temporarily.
Above-mentioned voice system control module is born the communication between user and the audiovisual/visual information searching system, and the audiovisual/visual information searching system of audio-visual product playing device receives specific user's natural language this audio-visual product playing device audiovisual/visual information searching system is sent i.e. the 3rd voice command of corresponding one spoken speech retrieval order.
Above-mentioned voice system control module comprises a digital signal processor, described digital signal processor links to each other with the modulus controller with digital to analog converter respectively, described digital to analog converter links to each other with the voice sensing element by wave filter, and described digital to analog converter links to each other with loudspeaker; Described digital signal processor also links to each other with a controller, and described controller and power supply and IO are connected by communication.
The inventive method is owing at first do classification index and descriptor sampling to audio-visual product, avoided extensive work amount and unclear expression that the audio-visual product note is brought, the inventive method is set up received pronunciation classification questions record model of cognition and received pronunciation key word recognition model to classification index and descriptor then, form first speech model, make up phonetic classification bibliographic database and the voice antistop list of audio-visual product by training to the specific user, form second speech model, each the 3rd voice command comes comparison by two more sane speech models like this, coupling can make the inventive method realize retrieval more accurately, recognition performance.First voice signal is that each audio-visual product is provided with the classification questions record simultaneously, carry out the sampling of content topic keyword and the model of cognition of index formation, so this voice signal is continuous, complete, systematized, second voice signal is the signal targetedly that the specific user makes up, though be continuous, but it is not necessarily complete, the 3rd voice command then is the order that is interrupted, above-mentioned three kinds of voice signals with different levels can make by the retrieval of voice to audio/video and neither repeat, do not omit yet, so the inventive method defines by the content topic that audio-visual product is carried out reasonably and the voice control features adapts, and realizes that in conjunction with artificial intelligence voice carry out fast audio-video content, retrieval accurately.The inventive method is owing to can carry out index and training at different specific users, and therefore the speech model that obtains adapting with it is a kind of general search method, is not subjected to the restriction of concrete syntax.
Speech retrieval method of the present invention possesses three information flows, i.e. first voice signal, second voice signal and the 3rd voice command.These three stream of voice information are dynamic, renewable.The information flow of first voice signal is used for default classification bibliographic database and predetermined keyword table are made up received pronunciation classification questions record model of cognition and received pronunciation key word recognition model in continuous complete mode, the information flow of second voice signal is used for making up voice bibliographic database and the voice antistop list that satisfies user's needs or preference with the continuation mode of user selection, and the information flow of the 3rd voice command is that the user realizes also can in time locking the broadcast program content that retrieves to the retrieval of audiovisual/visual information with the spoken language order.The information flow of second voice signal serves as basis and corresponding with it with the information flow of first voice signal, and the information flow of the 3rd voice command serves as basis and corresponding with it with the information flow of second voice signal, thereby brings retrieval running accurately and efficiently and control operation.
Adopt the audiovisual/visual information retrieval module of the audiovisual/visual information searching system of the inventive method to adopt speech model sequence and user's voice input feature vector sequence to mate to the 3rd voice command of storing in the dynamic storage, distortion more between the two draws the foundation of audio-visual product content retrieval, identification judgement.
Audiovisual/visual information searching system of the present invention can provide a kind of general mechanism, discerns different specific users speech retrieval command search result's different correlativity is judged.Specific user after system trained, utilize the verbal speech retrieval command can set specific user's request program in certain one-period and singly control to automatically switch and broadcast or program request, reduce the operation of user remote controller operating key such as audio-visual product playing devices such as televisor or displays.
The technology that the present invention utilizes the artificial intelligence of computing machine to combine with information retrieval realizes the speech understanding to the audio-visual product subject content, the audiovisual/visual information searching system of a two-way interaction of Zhi Zuoing and speech retrieval method thereof thus, can effectively improve user's precision ratio, recall ratio, make and operate hommization more, intelligent, convenient more.
Description of drawings
Fig. 1 represents the block scheme of audiovisual/visual information searching system of the present invention;
Mark among Fig. 1 lists as follows:
10--audiovisual/visual information searching system 20--voice system control module
22-voice sensing element 24-wave filter
26-number converter 28-power management
30-IO communication 32-loudspeaker
34-digital to analog converter 36-digital signal processor
38-controller 40--audiovisual/visual information retrieval module
42-dynamic storage 44-speech recognition device
46-audio-visual data index device 48-voice antistop list
50-classification bibliographic database 52-speech retrieval coupling subsystem
The 54-static memory
Embodiment
Below in conjunction with accompanying drawing the present invention is described more specifically.
Audiovisual/visual information searching system 10 described in Fig. 1 by two independently module be electrically connected and form: voice system control module 20 and audiovisual/visual information retrieval module 40.
Audiovisual/visual information retrieval module 40 is made up of static memory 54 and dynamic storage 42, static memory 54 comprises: an audio-visual data index device 46, according to user's needs and hobby, aural signature, the semantic feature of the content topic of various audio and video files are carried out the analysis based on sound model and speech model; Automatically gather the sample of an audio-visual product content topic keyword, construct the default classification bibliographic database of the predetermined keyword table and the audio-visual product of various audio-visual product content topics in advance, and set up corresponding standard phonetic classification questions record model of cognition and received pronunciation key word recognition model by first voice signal; One phonetic classification bibliographic database 50 according to specific user's custom, is defined arbitrarily by specific user's voice, forms the phonetic classification bibliographic database of the audio-visual product of standard at the specific user; One voice antistop list 48 according to specific user's custom, is defined arbitrarily by specific user's voice, forms the voice antistop list of the audio-visual product of standard at the specific user; One speech retrieval coupling subsystem 52 is accepted specific user's voice and is putd question to; Specific user's the 3rd voice command is carried out the inspection of grammer, word etc.; Specific user's source question formula is made an explanation or the processing of compiling property; The data model of question-type voice and received pronunciation classification questions record model of cognition and received pronunciation key word recognition model and phonetic classification bibliographic database and voice antistop list is carried out matching operation; Operation result is returned this system control module; One speech recognition device 44 is used for discerning user's verbal speech control input command: the foundation of indexing model first voice signal, promptly set up corresponding standard phonetic classification questions record model of cognition and received pronunciation key word recognition model; The one second phonetic entry command signal that corresponding first voice signal of specific user sends in the training mode is set up the voice antistop list and the phonetic classification bibliographic database of specific user's phonotape and videotape content topic; Receive in the standby mode by the retrieval of one the 3rd voice command that the specific user sent and put question to.
Dynamic storage 42 is used for storing the 3rd voice command temporarily.
Voice system control module 20 comprises a digital signal processor 36, digital signal processor 36 links to each other with modulus controller 26 with digital to analog converter 34 respectively, analog to digital converter 26 links to each other with voice sensing element 22 by wave filter 24, and digital to analog converter 34 links to each other with loudspeaker 32; Digital signal processor 36 also links to each other with a controller 38, and controller 38 links to each other with IO communication 30 with power management 28.
Come the power supply of ringing tone by power management 28 as information retrieval system 10.Lean on to such an extent that enough closely speak so that picked up the 3rd voice command that audiovisual/visual information searching system 10 is provided by voice sensing element 22 by the user.Voice sensing element 22 is a simulating signal with user's language conversion, and what be connected to voice sensing element 22 is wave filter 24, and this wave filter 24 is with the noise signals filtering in the non-genus voice frequency range in the simulating signal of voice sensing element 22 generations.What be connected to wave filter 24 is analog to digital converter 26, and this analog to digital converter 26 becomes digital signal with the analog signal conversion of wave filter 24 filterings.Digital signal sends to digital signal processor 36 by analog to digital converter 26, and this digital signal processor 36 deposits signal in the dynamic storage 42 in the audiovisual/visual information retrieval module 40 in.Then, in most preferred embodiment of the present invention, digital signal processor 36 calls speech recognition device 44, so that the digital signal that is stored in the dynamic storage 42 is carried out a series of frequency domain transformations; Speech recognition device 44 produces model of cognition, and this model of cognition is spectrum transformation, compares with the model of cognition (also being spectrum transformation) of order in being stored in keyword command list 48.If occurrence is arranged, controller 38 is visited the instruction set in the keyword command list 48 that is connected to the command recognition model so.The professional and technical personnel can know, being used for any other proper method of recognizing voice model can be with replacing spectrum transformation in the present invention.
Before playing an audio-video content, with the voice sensing element 22 of audiovisual/visual information searching system 10 near playback equipments, so that descriptive narration or useful voice messaging carry out the voice index to audio-video subject content in the automatic typing phonotape and videotape of the audio-visual data index device content.In one embodiment of the invention, in order to differentiate voice indexing model (this is first voice signal of standard quite continuous in one long period), voice training pattern (this is quite continuous specific user's second voice signal in one long period) and voice control input command are standby mode (this is the 3rd voice signal of one or two keyword normally), the user can be set under the voice index purpose state by the phonetic entry purpose item of IO communication 30 expressions, IO communication 30 setting signals are sent to digital signal processor 36, and indicating following voice signal is voice index order.After the user finished the voice index, the user set IO communication 30, and the input signal of voice state of a control is sent to controller 38, and this indicates the end of voice index signal.The voice index signal that is obtained, and is connected with the data file of audio-video content topic voice antistop list 48 and phonetic classification bibliographic database 50 in audio-visual data index device 46 as voice index file storage.
In one embodiment of the invention, one of voice control input command is voice index order.After sending voice control input index first voice signal, the voice signal that is used for the index purpose is obtained subsequently, and be used as voice index file storage, be connected with the audio-video content topic predetermined keyword table and the default data file of classifying bibliographic database of structure in advance in being stored in audio-visual data index device 46.When time that the user pipes down surpass the preset time section, such as at 5 seconds between 10 seconds the time, system does not have phonetic entry to be interpreted as the end of index voice signal this section in the schedule time.In this indexing model, the inventory of index audio-visual product document is treated in acquisition in advance, thereby constructs the predetermined keyword table and the default classification bibliographic database of the content topic of a large amount of audio-visual products.
When receiving and identifying index first voice signal, perhaps, the following phonetic entry of IO communication 30 expressions is for after the index purpose, is obtained by voice sensing element 22 from the user's voice signal, and is converted to first simulating signal.Wave filter 24 is with the noise signals filtering in the non-genus voice frequency range in the simulating signal of voice sensing element 22 generations, and analog to digital converter 26 becomes first digital signal with first analog signal conversion of wave filter 24 filterings.First digital signal is sent to digital signal processor 36.Digital signal processor 36 calls speech recognition device 44, so that first digital signal is carried out producing first model of cognition behind a series of frequency domain transformations, with voice index first digital signal as independent voice index file storage in audio-visual data index device 46, and be connected with the data file of audio-video theme voice antistop list 48 and phonetic classification bibliographic database 50.
Audiovisual/visual information searching system 10 can copy on another equipment by being connected to controller 38 connections, such as on the computing machine audio-visual data being carried out voice index or training.By speech recognition device 44, audiovisual/visual information searching system 10 training identification users' oral second voice signal of specific user.Speech recognition device 44 can be arranged in the audiovisual/visual information searching system, perhaps is arranged in the host computer system, when the audiovisual/visual information searching system is connected to host computer system, for the visit of audiovisual/visual information searching system.In preferred embodiment of the present invention, use on-line mode, so that utilize obtainable bigger computing power in the host computer system, and reduce the complicacy of audiovisual/visual information searching system.
In another embodiment of the present invention, one of voice control input command is the voice training order.After sending specific user's second voice signal, the user can require the audio-video content topic voice antistop list 48 of audiovisual/visual information searching system 10 audio playbacks prompting standard and the data file in the phonetic classification bibliographic database 50.Controller 38 is the required voice training file of retrieval from speech retrieval coupling subsystem 52, and it is passed to the digital to analog converter 34 that digital signal is converted to simulating signal, and simulating signal is passed to the loudspeaker 32 that produces audio frequency output.In addition, specific command instruction collection can utilize digital to analog converter 34 and loudspeaker 32, sends audible output signal to the user, to confirm to receive described order.
For example, when in training mode, using speech recognition device 44, the voice antistop list 48 of audiovisual/visual information searching system 10 audio playbacks prompting standard and the data file in the phonetic classification bibliographic database 50, the specific user trains by second voice signal.At every turn the repeating of specific user's second voice signal all picked up by audiovisual/visual information searching system 10, and is identified as the corresponding voice that comprise user's second voice signal in voice antistop list 48 and the phonetic classification bibliographic database 50 by speech recognition device 44 and changes model of cognition with tone.The model of cognition that the specific user is selected to be used for calling all second voice signals of various functions all is stored in the voice keyword command list 48 of audiovisual/visual information searching system.That model of cognition in voice antistop list 48 links separately is predetermined, also be stored in the instruction set of the various functions in the voice antistop list 48.Therefore, when oral the 3rd voice command of user was received and discerns by the audiovisual/visual information searching system, the instruction set relevant with this keyword order just was performed.Because the instruction set of function depends on training and speech recognition that user's keyword is selected and subsequently that keyword selected, so present embodiment is irrelevant with category of language, therefore can be with foreign language as voice control input command.For example, the user may revise the pronunciation of his or her order keyword " TV ", can discern up to the audiovisual/visual information searching system and resemble the order of saying the user.Therefore, in this embodiment, system will aim at a kind of specific language earlier, and in this language, caused action represented in the order keyword.Can make the foreign language form of this system for the user who uses the foreign language keyword of representing caused action.
The above, it only is preferred embodiment of the present invention, be not that the present invention is done any pro forma restriction, every foundation technical spirit of the present invention all still belongs in the scope of technical solution of the present invention any simple modification, equivalent variations and modification that above embodiment did.