Patent Grant
7240003
This application is a National Stage Filing Under 35 U.S.C. 371 of International Application No. PCT/GB01/04331, filed Sep. 28, 2001, and published in English as International Publication No. WO 02/27546 A2, on Apr. 4, 2002.
The present invention relates to the annotation of data files which are to be stored in a database for facilitating their subsequent retrieval. The present invention is also concerned with a system for generating the annotation data which is added to the data file and to a system for searching the annotation data in the database to retrieve a desired data file in response to a user's input query. The invention also relates to a system for translating an unordered list of nodes and links into an ordered and blocked list of nodes and links.
Databases of information are well known and suffer from the problem of how to locate and retrieve the desired information from the database quickly and efficiently. Existing database search tools allow the user to search the database using typed keywords. Whilst this is quick and efficient, this type of searching is not suitable for various kinds of databases, such as video or audio databases.
According to one aspect, the present invention aims to provide a data structure for the annotation of data files within a database which will allow a quick and efficient search to be carried out in response to a user's input query.
According to another aspect, the present invention According to another aspect, the present invention provides data defining a phoneme and word lattice for use as annotation data for annotating data files to be stored within a database. Preferably, the data defines a plurality of nodes and a plurality of links connecting the nodes, and further data associates a plurality of phonemes with a respective plurality of links and further data associates at least one word with at least one of said links, and further data defines a block arrangement for the nodes such that the links may only extend over a given maximum number of blocks. It is further preferred that the links may only extend into a following block.
According to another aspect, the present invention provides an apparatus for searching a database which employs the annotation data discussed above for annotating data filed therein. Preferably, the apparatus is arranged to generate phoneme data in response to a user's query or input, and to search the database using the generated phoneme data. It is further preferred that word data is also generated from the user's input or query.
According to another aspect, the present invention provides an apparatus for generating a phoneme and word lattice corresponding to received phoneme and word data, comprising means for defining a plurality of links and a plurality of nodes between which the links extend, means for associating the links with phonemes or words, and means for arranging the nodes in a sequence of time ordered blocks in which the links only extend up to a maximum given number of blocks later in the sequence. Preferably, the maximum extension allowed for a link is to extend into a following block. It is further preferred that the apparatus is arranged to add nodes or links incrementally as it forms the lattice, and to split an existing block of nodes into at least two blocks of nodes.
According to another aspect, the present invention provides an apparatus for adding phonemes or words to a phoneme and word lattice of any of the types discussed above, and arranged to analyse which data defining the current phoneme and word lattice needs to be modified in dependence upon the extent to which the links are permitted to extend from one block to another. Preferably, this analysis is further dependent upon the location within the lattice of a point identifying the latest node in each block to which any link originating in the preceding block extends and a point identifying the earliest node in each block from which a link extends into the succeeding block.
According to another aspect, the present invention provides a method of adding phonemes or words to a phoneme and word lattice of any of the types discussed above, comprising analysing which data defining the current phoneme and word lattice needs to be modified in dependence upon the extent to which the links are Preferably, this analysis is further dependent upon the location within the lattice of respective points identifying the latest node in each block to which any link originating in the preceding block extends.
According to another aspect, a method and apparatus are provided for converting an unordered list of nodes and links into an ordered and blocked list of nodes and links. The blocks are formed by filling and splitting: successive nodes are inserted into a block until it is full, then a new block is begun. If new nodes would overfill an already full block, that block is split into two or more blocks. Constraints on the links regarding which block they can lead to are used to speed up the block splitting process, and identify which nodes remain in the old block and which go into the new block.
Exemplary embodiments of the present invention will now be described with reference to the accompanying figures, in which:
a is a schematic diagram of a phoneme lattice for an example audio string from the input video data file;
b is a schematic diagram of a word and phoneme lattice embodying one aspect of the present invention, for an example audio string from the input video data file;
a is a schematic diagram of a word and phoneme lattice showing relative timings of the nodes of the lattice;
b is a schematic diagram showing the nodes of a word and phoneme lattice divided into blocks.
a is a schematic diagram illustrating the format of data corresponding to one node of a word and phoneme lattice;
b is a schematic diagram illustrating a data stream defining a word and phoneme lattice;
a to 18h are schematic diagrams illustrating the build-up of a word and phoneme lattice;
a to 19h are schematic diagrams illustrating the build-up of a data stream defining a word and phoneme lattice;
a to 20c are schematic diagrams showing the updating of a word and phoneme lattice on insertion of a long link;
a to 21d are schematic diagrams illustrating the updating of a word and phoneme lattice on insertion of additional nodes;
a and 23b are schematic diagrams illustrating the application of a block splitting procedure to a word and phoneme lattice; and
Embodiments of the present invention can be implemented using dedicated hardware circuits, but the embodiment to be described is implemented in computer software or code, which is run in conjunction with processing hardware such as a personal computer, work station, photocopier, facsimile machine, personal digital assistant (PDA) or the like.
The programme instructions which make the PC 1 operate in accordance with the present invention may be supplied for use with an existing PC 1 on, for example, a storage device such as a magnetic disc 13, or by downloading the software from the Internet (not shown) via the internal modem and telephone line 9.
Data File Annotation
A system has been proposed to generate N-Best word lists for an audio stream as annotation data by passing the audio data from a video data file through an automatic speech recognition unit. However, such word-based systems suffer from a number of problems. These include (i) that state of the art speech recognition systems still make basic mistakes in recognition; (ii) that state of the art automatic speech recognition systems use a dictionary of perhaps 20,000 to 100,000 words and cannot produce words outside that vocabulary; and (iii) that the production of N-Best lists grows exponentially with the number of hypothesis at each stage, therefore resulting in the annotation data becoming prohibitively large for long utterances.
The first of these problems may not be that significant if the same automatic speech recognition system is used to generate the annotation data and to subsequently retrieve the corresponding data file, since the same decoding error could occur. However, with advances in automatic speech recognition systems being made each year, it is likely that in the future the same type of error may not occur, resulting in the inability to be able to retrieve the corresponding data file at that later date. With regard to the second problem, this is particularly significant in video data applications, since users are likely to use names and places (which may not be in the speech recognition dictionary) as input query terms. In place of these names, the automatic speech recognition system will typically replace the out of vocabulary words with a phonetically similar word or words within the vocabulary, often corrupting nearby decodings. This can also result in the failure to retrieve the required data file upon subsequent request.
In contrast, with the proposed phoneme and word lattice annotation data, a quick and efficient search using the word data in the database 29 can be carried out and, if this fails to provide the required data file, then a further search using the more robust phoneme data can be performed. The phoneme and word lattice is an acyclic directed graph with a single entry point and a single exit point. It represents different parses of the audio stream within the data file. It is not simply a sequence of words with alternatives since each word does not have to be replaced by a single alternative, one word can be substituted for two or more words or phonemes, and the whole structure can form a substitution for one or more words or phonemes. Therefore, the density of data within the phoneme and word lattice essentially remains linear throughout the audio data, rather than growing exponentially as in the case of the N-Best technique discussed above. As those skilled in the art of speech recognition will realise, the use of phoneme data is more robust, because phonemes are dictionary independent and allow the system to cope with out of vocabulary words, such as names, places, foreign words etc. The use of phoneme data is also capable of making the system future proof, since it allows data files which are placed into the database to be retrieved even when the words were not understood by the original automatic speech recognition system.
The way in which this phoneme and word lattice annotation data can be generated for a video data file will now be described with reference to
As shown in
a illustrates the form of the phoneme lattice data output by the speech recognition unit 33, for the input audio corresponding to the phrase '. . . now is the winter of our . . . '. The automatic speech recognition unit 33 identifies a number of different possible phoneme strings which correspond to this input audio utterance. For example, the speech recognition system considers that the first phoneme in the audio string is either an /m/ or an /n/. For clarity, only the alternatives for the first phoneme are shown. As is well known in the art of speech recognition, these different possibilities can have their own weighting which is generated by the speech recognition unit 33 and is indicative of the confidence of the speech recognition unit's output. For example, the phoneme /n/ may be given a weighting of 0.9 and the phoneme /m/ may be given a weighting of 0.1, indicating that the speech recognition system is fairly confident that the corresponding portion of audio represents the phoneme /n/, but that it still may be the phoneme /m/.
In this embodiment, however, this weighting of the phonemes is not performed.
As shown in
In this embodiment, the annotation data 31-3 stored in the database 29 has the following general form:
The time of start data in the header can identify the time and date of transmission of the data. For example, if the video file is a news broadcast, then the time of start may include the exact time of the broadcast and the date on which it was broadcast.
The flag identifying if the annotation data is word annotation data, phoneme annotation data or if it is mixed is provided since not all the data files within the database will include the combined phoneme and word lattice annotation data discussed above, and in this case, a different search strategy would be used to search this annotation data.
In this embodiment, the annotation data is divided into blocks in order to allow the search to jump into the middle of the annotation data for a given audio data stream. The header therefore includes a time index which associates the location of the blocks of annotation data within the memory to a given time offset between the time of start and the time corresponding to the beginning of the block.
The header also includes data defining the word set used (i.e. the dictionary), the phoneme set used and the language to which the vocabulary pertains. The header may also include details of the automatic speech recognition system used to generate the annotation data and any appropriate settings thereof which were used during the generation of the annotation data.
The phoneme probability data defines the probability of insertions, deletions, misrecognitions and decodings for the system, such as an automatic speech recognition system, which generated the annotation data.
The blocks of annotation data then follow the header and identify, for each node in the block, the time offset of the node from the start of the block, the phoneme links which connect that node to other nodes by phonemes and word links which connect that node to other nodes by words. Each phoneme link and word link identifies the phoneme or word which is associated with the link. They also identify the offset to the current node. For example, if node N50 is linked to node N55 by a phoneme link, then the offset to node N50 is 5. As those skilled in the art will appreciate, using an offset indication like this allows the division of the continuous annotation data into separate blocks.
In an embodiment where an automatic speech recognition unit outputs weightings indicative of the confidence of the speech recognition units output, these weightings or confidence scores would also be included within the data structure. In particular, a confidence score would be provided for each node which is indicative of the confidence of arriving at the node and each of the phoneme and word links would include a transition score depending upon the weighting given to the corresponding phoneme or word. These weightings would then be used to control the search and retrieval of the data files by discarding those matches which have a low confidence score.
Data File Retrieval
As those skilled in the art will appreciate, this type of phonetic and word annotation of data files in a database provides a convenient and powerful way to allow a user to search the database by voice. In the illustrated embodiment, a single audio data stream was annotated and stored in the database for subsequent retrieval by the user. As those skilled in the art will appreciate, when the input data file corresponds to a video data file, the audio data within the data file will usually include audio data for different speakers. Instead of generating a single stream of annotation data for the audio data, separate phoneme and word lattice annotation data can be generated for the audio data of each speaker. This may be achieved by identifying, from the pitch or from another distinguishing feature of the speech signals, the audio data which corresponds to each of the speakers and then by annotating the different speaker's audio separately. This may also be achieved if the audio data was recorded in stereo or if an array of microphones were used in generating the audio data, since it is then possible to process the audio data to extract the data for each speaker.
In the above embodiments, a speech recognition system was used to generate the annotation data for annotating a data file in the database. As those skilled in the art will appreciate, other techniques can be used to generate this annotation data. For example, a human operator can listen to the audio data and generate a phonetic and word transcription to thereby manually generate the annotation data.
In the above embodiments, the annotation data was generated from audio stored in the data file itself. As those skilled in the art will appreciate, other techniques can be used to input the annotation data.
The automatic speech recognition unit 51 generates the phoneme and word lattice annotation data by (i) generating a phoneme lattice for the input utterance; (ii) then identifying words within the phoneme lattice; and (iii) finally by combining the two.
As those skilled in the art will appreciate, this embodiment can be used to annotate any kind of image such as x-rays of patients, 3D videos of, for example, NMR scans, ultrasound scans etc. It can also be used to annotate one-dimensional data, such as audio data or seismic data.
In the above embodiment, a data file was annotated from a voiced annotation. As those skilled in the art will appreciate, other techniques can be used to input the annotation. For example,
In the above embodiment, a phonetic transcription unit 107 was used for generating the annotation data for annotating the image or text data. As those skilled in the art will appreciate, other techniques can be used. For example, a human operator can manually generate this annotation data from the image of the document itself.
In the first embodiment, the audio data from the data file 31 was passed through an automatic speech recognition unit in order the generate the phoneme annotation data. In some situations, a transcript of the audio data will be present in the data file. Such an embodiment is illustrated in
In the above embodiments, a phoneme (or phoneme-like) and word lattice was used to annotate a data file. As those skilled in the art of speech recognition and speech processing will realise, the word “phoneme” in the description and claims is not limited to its linguistic meaning but includes the various sub-word units that are identified and used in standard speech recognition systems, such as phonemes, syllables, Katakana (Japanese alphabet) etc.
Lattice Generation
In the above description, generation of the phoneme and word lattice data structure shown in
a shows the timing of each node of the lattice relative to a common zero time, which in the present example is set such that the first node occurs at a time of 0.10 seconds. It is noted that
In the present embodiment, the nodes are divided into three blocks as shown in
The time of each node is provided relative to the time of the start of its respective block. This does not affect the timings of the nodes in block 0. However, for the further blocks the new off-set timings are different from each node's absolute relative timing as per
The use of time off-sets determined relative to the start of each block rather than from the start of the whole lattice provides advantages with respect to dynamic range as follows. As the total time of a lattice increases, the dynamic range of the data type used to record the timing values in the lattice structure will need to increase accordingly, which will consume large amounts of memory. This will become exacerbated when the lattice structure is being provided for a data file of unknown length, for example if a common lattice structure is desired to be usable for annotating either a one minute television commercial or a film or television programme lasting a number of hours. In contrast, the dynamic range of the corresponding data type for the lattice structure divided into blocks is significantly reduced by only needing to accommodate a maximum expected time off-set of a single block, and moreover this remains the same irrespective of the total duration of the data file. In the present embodiment the data type employed provides integer values where each value of the integer represents the off-set time measured in hundredths of a second.
b also shows certain parts of the lattice structure identified as alpha (α) and beta (β). The significance of these items will be explained later.
The format in which the data is held for each respective node in the preferred form of the phoneme and word lattice data structure will now be explained with reference to
The first data component 210 specifies the time off-set of the node from the start of the block. In the present example, the value is 0.10 seconds, and is implemented by means of the integer data type described earlier above.
The second data component 212 represents the word link “NOW”, which is shown in
The fourth data component 216 represents the phoneme /n/ which extends from the first node to the second node, entailing therefore a nodal off-set of one which leads directly to the value 001 for the fifth data component 218 as shown in
The manner in which the data components 212, 216 and 220 represent the respective word or phoneme associated with their link can be implemented in any appropriate manner. In the present embodiment the data components 212, 216 and 220 consist of an integer value which corresponds to a word index entry value (in the case of a word link) or a phoneme index entry value (in the case of a phoneme link). The index entry value serves to identify an entry in a corresponding word or phoneme index containing a list of words or phonemes as appropriate. In the present embodiment the corresponding word or phoneme index is held in the header part of the annotation data 31-3 described earlier. In other embodiments the header may itself only contain a further cross-reference identification to a separate database storing one or more word or phoneme indices.
Generally, the different links corresponding to a given node can be placed in the data format of
The data for each node, in the form shown in
Earlier, with reference to
The block arrangement shown in
By virtue of the blocks being implemented so as to obey the above described criteria, the following advantages are achieved. If further data is later to be inserted into the phoneme and word lattice structure, this may involve the insertion of one or more additional nodes. In this event, any existing link “passing over” a newly inserted node will require its nodal off-set to be increased by one, as the newly inserted node will need to be included in the count of the number of nodes over which the existing link extends. For example, if a new node were inserted at a time of 0.50 seconds into block 2, then it can be seen from
During insertion of such additional nodes and processing of the consequential changes to the nodal off-sets, it is necessary to search back through the lattice data structure from the point of the newly inserted node in order to analyse the earlier existing nodes to determine which of them have links having a nodal off-set sufficiently large to extend beyond the newly inserted node. An advantage of the blocks of the lattice data structure being arranged according to the present criteria is that it reduces the number of earlier existing nodes that need to be analysed. More particularly, it is only necessary to analyse those nodes in the same block in which the node is inserted which precede the inserted node plus the nodes in the neighbouring block directly preceding the block in which the new node has been inserted. For example, if a new node is to be inserted at 0.50 seconds in block 2, it is only necessary to analyse the four existing nodes in block 2 that precede the newly inserted node plus the five nodes of block 1. It is not necessary to search any of the nodes in block 0 in view of the block criteria discussed above.
This advantage becomes increasingly beneficial as the length of the lattice increases and the number of blocks formed increases. Furthermore, the advantage not only applies to the insertion of new nodes into an otherwise complete lattice, it also applies to the ongoing procedure of constructing the lattice, which may occur when nodes are not necessarily inserted into a lattice in strict time order.
Yet further, it is noted that the particular choice of the criteria to only allow links to extend into a neighbouring block may be varied, for example the criteria may allow links extending only as far as four blocks away, it then being necessary to search back only a maximum of four blocks. This still provides a significant advantage in terms of reducing the level of processing required in the case of large lattices, particularly lattices with hundreds or thousands of blocks. The skilled practitioners will appreciate that any appropriate number of blocks can be chosen as the limit in the criteria, it merely being necessary to commensurately adapt the number of blocks that are searched back through.
The lattice data structure of the present embodiment contains a further preferred refinement which is also related to the extension of the word or phoneme links into neighbouring blocks. In particular the lattice data structure further includes data specifying two characteristic points of each block. The two characteristic points for each block are shown as alpha (α) and beta (β) in
Beta for a given block is defined as the time of the latest node in the given block to which any link originating from the previous block extends. Thus, in the case of block 1, beta is at the first node in the block (i.e. the node to which the phoneme link /z/ and the word link “IS” extend), since there are no links originating in block 0 that extend further than the first node of block 1. In the case of block 2, beta is at the third node, since the word link “WINTER” extends to that node from block 1. In the case of the first block of the lattice structure i.e. block zero, there are intrinsically no links extending into that block. Therefore, beta for this block is defined as occurring before the start of the lattice.
Alpha for a given block is defined as the time of the earliest node in the given block from which a link extends into the next block. In the case of block 0, two links extend into block 1, namely word link “IS” and the phoneme link /z/. Of these, the node from which the word link “IS” extends is earlier in block 0 than the node from which the phoneme link /z/ extends, hence alpha is at the node from which the word link “IS” extends. Similarly, alpha for block 1 is located at the node where the word link “WINTER” originates from. In the case of the last block of the lattice, in this case block 2, there are intrinsically no links extending into any further block, hence alpha is specially defined as being at the last node in the block. Thus it can be appreciated that conceptually beta represents the latest point in a block before which there are nodes which interact with the previous block, and alpha represents the earliest point in a block after which there are nodes which interact with the next block.
As those skilled in the art will appreciate, each alpha and beta can be specified by identification of a particular node or by specification in terms of time. In the present embodiment identification is specified by nodes. The data specifying alpha and beta within the lattice data structure can be stored in a number of different ways. For example, data components of the type shown in
The specification of alpha and beta for each block firstly provides certain advantages with respect to analysing the nodal off-sets of previous nodes in a lattice when a new node is inserted. In particular, when a new node is inserted at a location after beta in a given block, it follows that it is only necessary to analyse the preceding nodes in the given block, and it is no longer necessary to analyse the nodes in the block preceding the given block. This is because it is already known that by virtue of the new inserted node being after beta within the given block, there can by definition be no links that extend from the previous block beyond the newly inserted node, since the position of beta defines the greatest extent which any links extend from the previous block. Thus the need to search and analyse any of the nodes of the preceding block has been avoided, which becomes particularly advantageous as the average size of blocks increases. If alternatively a new node is inserted into a given block at a location before beta of the given block, then it is now necessary to consider links originating from the preceding block as well, but only those nodes at or after alpha in the preceding block. This is due to the fact that from the definition of alpha, it is already known that none of the nodes in the preceding block that come before the preceding block's alpha have links which extend into the given block. Thus processing is again reduced, and the reduction will again become more marked as the size of individual blocks is increased. Moreover, the position of alpha in any given block will tend to be towards the end of that block, so that in the case of long blocks the majority of the processing resource that would otherwise have been used analysing the whole of the preceding block is saved.
The specification of alpha and beta for each block secondly provides certain advantages with respect to employing alpha and beta in procedures to re-define blocks within an existing lattice so as to provide smaller or more evenly arranged blocks whilst maintaining compliance with the earlier mentioned criteria that no link may extend further than one block. In these procedures, existing blocks are essentially split, according to the relative position of alpha and beta within an existing block. In one approach, provided alpha occurs after beta within a given block, the given block can be divided into two blocks by splitting it somewhere between beta and alpha. Similarly, the data specifying beta and alpha is advantageously employed to determine when existing blocks can be split into smaller blocks in the course of a preferred procedure for constructing the lattice data structure.
It was mentioned earlier above that in the present embodiment the longest link from a given node is positioned first in the sequence of data components for any given node as shown in
A preferred method of generating the above described lattice data structure will now be described with reference to
In overview, as each set of data components is added to the lattice, the various ends of blocks, alphas and betas are updated. When the number of nodes in a block reaches a critical value, in this example 9, the locations of alpha and beta are analysed and if suitable the block is split into two smaller blocks. The various alphas and betas are again updated, and the process then continues in the same manner with the addition of further data components.
The process steps laid out in
Referring to
At step S63 the automatic speech recognition unit 33 sets an incremental counter n equal to 1.
At step S65 the automatic speech recognition unit 33 inserts the first set of data components into the data stream defining the lattice data structure. More particularly, the automatic speech recognition unit 33 collects the data corresponding to the first two nodes of the lattice and any direct phoneme links therebetween (in this case phoneme links /n/ and /m/). It then additionally collects any words that have been identified by the word decoder 37 as being associated with a link between these two nodes, although in the case of the first two nodes no such word has been identified. It then inserts the corresponding data components into the data stream. In particular, referring again to
In the present case, the process then moves on to step S71, where the automatic speech recognition unit 33 defines the end of the last block to be immediately after the newly inserted node which is at the end of the lattice. At this stage of the procedure there is only one block, hence in defining the end of the last block, the end of the sole block is in fact defined. This newly defined current end of the block is shown as item 203 in
The automatic speech recognition unit 33 then determines all of the alpha and beta points. At the present stage there is only one block so only one alpha and one beta is determined. The procedure for determining alpha and beta in the first block was described earlier above. The resulting positions are shown in
As step S79 the automatic speech recognition unit 33 determines whether any of the alpha and beta values are “invalid”, in the sense of being either indeterminate or positioned such as to contravene the earlier described criteria that no link may extend further than into a directly neighbouring block. At the present stage of building up the lattice this determination step obviously determines that there is no such invalidity, and hence the process moves to step S81. At step S81 the automatic speech recognition unit determines whether the number of nodes in any blocks that have just had nodes inserted in them has reached or exceeded a predetermined critical number. The predetermined critical number is set for the purpose of defining a minimum number of nodes that must be in a block before the block structure will be analysed or altered for the purposes of giving smaller block sizes or more even block spacings. There is an effective overhead cost in terms of resources that are required when carrying out block division, data storage of the block flag data, and so on. Hence block division for blocks containing less than the critical number of nodes would tend to be counter productive. The choice of the value of the critical number will depend on the particular characteristics of the lattice or data file being considered. As mentioned above, in the present embodiment the number is set at nine. Hence at the present stage of the process, where only two nodes have been inserted in total, the answer to the determination step S81 is no.
The process steps are thus completed for the first set of data components to be inserted, and the current form of the lattice and data stream is shown in
The procedure then moves to step S89, where the automatic speech recognition unit determines that more sets of data components are to be added, and hence at step S91 increments the value of n by one and the process steps beginning at steps S65 are repeated for the next set of data components. In the present case the next set of data components consists of data (item 270 in
As the procedure continues, the fourth node and the two links which end at that node, namely the phoneme link /w/ and the word link “NOW”, representing the next set of data components, are inserted. The process steps from S65 onwards are followed as described for the previous sets of data components, resulting in the lattice structure shown in
The procedure continues as described above without variation for the insertion of the fifth, sixth, seventh and eighth nodes providing the lattice structure and data stream shown in
It can thus be appreciated that the basic approach of the present method is that when the number of nodes in a block reaches nine or more, the block will be divided into two blocks, provided that alpha is greater than beta. The reason for waiting until a certain number of nodes has been reached is due to the cost in overhead resource, as was explained-earlier above. The reason for the criteria that alpha be greater than beta is to ensure that each of the two blocks formed by the division of an original block will obey the earlier described criteria that no link is permitted to extend into any block beyond a directly neighbouring block.
Therefore, in the present case, the procedure moves to step S85 in which the automatic speech recognition unit splits the sole block of
At step S87 updated values of alpha and beta are determined by the automatic speech recognition unit. Given there are now two blocks, there are two betas and two alphas to be determined. The new locations of these alphas and betas are shown in
The procedure of
The next set of data components inserted consists of the fourteenth node and the phoneme link /oh/ ending at that node. The situation after steps S65 to S79 are implemented for this set of data components is shown in
The procedure shown in
At this stage, the automatic speech recognition unit 33 determines at step S89 that no more sets of data components are available to be inserted, and hence the current lattice data structure is complete, and indeed corresponds to the lattice shown in
An example will now be given to demonstrate the merging of two blocks due to the later insertion of a long link that extends beyond a neighbouring block. This situation did not arise in the earlier example because the data was added into the lattice on a fully time ordered sequential basis. In contrast, in the following example, after the lattice of
However, in the present example, the insertion of the earlier timed link is essentially part of the original on-going construction of the lattice, although the data component consisting of the additional link is processed separately at the end because it constitutes a word recognised by the automatic speech recognition unit 33 when passing the phoneme data through a second speech recognition vocabulary. In the present example, the second vocabulary consists of a specialised name place vocabulary that has been optionally selected by a user. Hence, in the present example, at step S89 it is determined that a further set of data components is to be inserted, and following incrementing of the value of n at step S91, the data is inserted at step S65. The data consists of the word link “ESTONIA” and extends from the fourth node of block 0 to the third node of block 2, as shown in
At step S67 the automatic speech recognition unit 33 recognises that no new node has been inserted, hence the process moves to step S75 where it determines updated locations of alpha and beta. However, because the newly inserted link extends from block 0 right over block 1 to end in block 2, it contravenes the earlier described criteria barring link extensions beyond directly neighbouring blocks, and moreover does not produce a valid alpha or beta for block 1. This is represented in
The procedure therefore moves to step S77 which consists of merging blocks. Any suitable criteria can be used to choose which blocks should be merged together, for example the criteria can be based on providing the most evenly spaced blocks, or could consist of merging the offending block with its preceding block. However, in the present embodiment the choice is always to merge the offending block with its following block, i.e. in the present example block 1 will be merged with block 2.
This is implemented by removal of the block marker dividing block 1 from block 2, resulting in two blocks only, as shown in
At step S79 the automatic speech recognition unit 33 determines that alpha and beta are now valid, so the procedure moves to step S81. In the present example, because there are now twelve nodes in block 1 and because alpha is greater than beta, the procedure moves to step S85 and block 1 is split using the same procedure as described earlier above. However, the earlier employed criteria specifying where to locate the new block division, namely half way in terms of nodes between beta and alpha, contains in the present example a refinement that when the block to be split has greater than nine nodes, splitting should, where possible, leave the earlier of the two resulting blocks with no more than eight nodes. This is to avoid inefficient repetitions of the block splitting process. Hence in the present example the new block marker is inserted immediately after the eighth node of the block being split, as shown in
In the above procedure described with reference to
A further example demonstrating the processing of data according to the procedure laid out in the flow chart of
In this further example, additional data is added via a keyboard and a phonetic transcription unit, of the same form as the keyboard 3 and phonetic transcription unit 75 shown in
Referring again to
Following adjustment of the off-sets, the procedure of
The procedure then continues in the same fashion resulting in the insertion of data components (iv), (v) and (vi) up to reaching step S81 during processing of data component (vi). At this stage, the lattice is of the form shown in
The above-mentioned preferred procedure for implementing step S73 of adjusting the off-sets will now be described with reference to the flow chart of
The above procedure is implemented by the process steps shown in
An alternative way of splitting a block will now be described. When the number of nodes in a given block has reached the critical number and alpha is later than beta for the given block, then the given block and the preceding block are adjusted to form three new blocks in place of those two blocks. This procedure will now be described more fully with reference to
a shows a sequence of nodes within a lattice, linked by phoneme links for example phoneme link 412, the end part of a word link 414 and a further word link 416. The nodes are divided into blocks by block markers 402, 404 and 406, forming blocks n and (n+1) of the lattice.
The positions of alpha and beta for block n and block (n+1) respectively are shown also.
In an alternative version of the embodiments described in the preceding paragraph, the two new block dividers may be positioned at nodes relatively close, compared to the number of nodes in each block, to the position of beta of block n and beta of block (n+1) respectively, instead of at those two beta positions as such.
In the above embodiments, the timing of each node of the lattice is provided, prior to arrangement in blocks, relative to a common zero time set such that the first node occurs at a time of 0.10 seconds. The start time for the first block is set equal to the common zero time. The start time for each of the other blocks is the time of the last node of the preceding block. However, in an alternative embodiment the timing of each node may be provided in an absolute form, and the block marker demarcating the start of each block may be given a Universal Standard Time (UST) time stamp, corresponding to the absolute time of the first node of that block rounded down to the nearest whole second. The UST time stamp may be implemented as a 4 byte integer representing a count of the number of seconds since 1 Jan. 1, 1970. The times of the nodes in each block are then determined and stored as offset times relative to the rounded UST time of the start of the block. Because in this embodiment each block time is rounded to the nearest second, if block durations of less than 1 second were to be permitted, then two or more blocks could be allocated the same time stamp value. Therefore, when UST time stamps are employed, block durations less than 1 second are not permitted. This is implemented by specifying a predetermined block duration, e.g. 1 second, that a current block must exceed before splitting of the current block is performed. This requirement will operate in addition to the earlier described requirement that the current block must contain greater than a predetermined number of nodes before splitting of the current block is performed. Alternatively, shorter block durations may be accommodated, by employing a time stamp convention other than UST and then rounding down the block marker times more precisely than the minimum allowed duration of a block.
In the above embodiments the phoneme and word lattice structure was determined and generated by the automatic speech recognition unit 33, configured with the requisite functionality. As will readily be appreciated by those skilled in the art, a standard automatic speech recognition unit can be used instead, in conjunction with a separate lattice creation unit comprising the functionality for determining and generating the above described phoneme and word lattice structure. An embodiment employing a standard automatic speech recognition unit 40, which outputs a sequence of phonemes is shown in
In the above embodiments, the phoneme and word data was associated with the links of the lattice. As those skilled in the art will appreciate, the word and/or the phoneme data can be associated with the nodes instead. In this case the data associated with each node would preferably include a start and an end time for each word or phoneme associated therewith.
A technique has been described above for organising an unordered list of nodes and links into an ordered and blocked list. The technique has been described for the particular application of the ordering of an unordered list of phonemes and words. However, as those skilled in the art will appreciate, this technique can be applied to other types of data lattices. For example the technique can be applied to a lattice which only has phonemes or a lattice which only has words. Alternatively still, it can be applied to a lattice generated from a hand writing recognition system which produces a lattice of possible characters as a result of a character recognition process. In this case, the nodes and links would not be ordered in time, but would be spatially ordered so that the characters appear in the ordered lattice at a position which corresponds to the character's position on the page relative to the other characters.
| Number | Date | Country | Kind |
|---|---|---|---|
| 002390.1 | Sep 2000 | GB | national |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/GB01/04331 | 9/28/2001 | WO | 00 | 3/7/2003 |
| Publishing Document | Publishing Date | Country | Kind |
|---|---|---|---|
| WO02/27546 | 4/4/2002 | WO | A |
| Number | Name | Date | Kind |
|---|---|---|---|
| 4227176 | Moshier | Oct 1980 | A |
| 4736429 | Niyada et al. | Apr 1988 | A |
| 4903305 | Gillick et al. | Feb 1990 | A |
| 4975959 | Benbassat | Dec 1990 | A |
| 4980918 | Bahl et al. | Dec 1990 | A |
| 4985924 | Matsuura | Jan 1991 | A |
| 5075896 | Wilcox et al. | Dec 1991 | A |
| 5131043 | Fujii et al. | Jul 1992 | A |
| 5136655 | Bronson | Aug 1992 | A |
| 5202952 | Gillick et al. | Apr 1993 | A |
| 5333275 | Wheatley et al. | Jul 1994 | A |
| 5390278 | Gupta et al. | Feb 1995 | A |
| 5500920 | Kupiec | Mar 1996 | A |
| 5577249 | Califano | Nov 1996 | A |
| 5594641 | Kaplan et al. | Jan 1997 | A |
| 5638425 | Meador, III et al. | Jun 1997 | A |
| 5640487 | Lau et al. | Jun 1997 | A |
| 5649060 | Ellozy et al. | Jul 1997 | A |
| 5675706 | Lee et al. | Oct 1997 | A |
| 5680605 | Torres | Oct 1997 | A |
| 5684925 | Morin et al. | Nov 1997 | A |
| 5708759 | Kemeny | Jan 1998 | A |
| 5721939 | Kaplan | Feb 1998 | A |
| 5729741 | Liaguno et al. | Mar 1998 | A |
| 5737489 | Chou et al. | Apr 1998 | A |
| 5737723 | Riley et al. | Apr 1998 | A |
| 5752227 | Lyberg | May 1998 | A |
| 5781884 | Pereira et al. | Jul 1998 | A |
| 5787414 | Miike et al. | Jul 1998 | A |
| 5799267 | Siegel | Aug 1998 | A |
| 5835667 | Wactlar et al. | Nov 1998 | A |
| 5852822 | Srinivasan et al. | Dec 1998 | A |
| 5870740 | Rose et al. | Feb 1999 | A |
| 5873061 | Häb-Umbach et al. | Feb 1999 | A |
| 5907821 | Kaji et al. | May 1999 | A |
| 5983177 | Wu et al. | Nov 1999 | A |
| 5999902 | Scahill et al. | Dec 1999 | A |
| 6023536 | Visser | Feb 2000 | A |
| 6026398 | Brown et al. | Feb 2000 | A |
| 6061679 | Bournas et al. | May 2000 | A |
| 6070140 | Tran | May 2000 | A |
| 6122613 | Baker | Sep 2000 | A |
| 6172675 | Ahmad et al. | Jan 2001 | B1 |
| 6182039 | Rigazio et al. | Jan 2001 | B1 |
| 6192337 | Ittycheriah et al. | Feb 2001 | B1 |
| 6236964 | Tamura et al. | May 2001 | B1 |
| 6243680 | Gupta et al. | Jun 2001 | B1 |
| 6272242 | Saitoh et al. | Aug 2001 | B1 |
| 6289140 | Oliver | Sep 2001 | B1 |
| 6314400 | Klakow | Nov 2001 | B1 |
| 6321226 | Garber et al. | Nov 2001 | B1 |
| 6389395 | Ringland | May 2002 | B1 |
| 6463413 | Applebaum et al. | Oct 2002 | B1 |
| 6487532 | Schoofs et al. | Nov 2002 | B1 |
| 6490563 | Hon et al. | Dec 2002 | B2 |
| 6535850 | Bayya | Mar 2003 | B1 |
| 6567778 | Chao Chang et al. | May 2003 | B1 |
| 6567816 | Desai et al. | May 2003 | B1 |
| 6662180 | Aref et al. | Dec 2003 | B1 |
| 6990448 | Charlesworth et al. | Jan 2006 | B2 |
| 20020022960 | Charlesworth et al. | Feb 2002 | A1 |
| 20020052740 | Charlesworth et al. | May 2002 | A1 |
| Number | Date | Country |
|---|---|---|
| 597798 | May 1994 | EP |
| 649144 | Apr 1995 | EP |
| 689153 | Dec 1995 | EP |
| 789349 | Aug 1997 | EP |
| 849723 | Jun 1998 | EP |
| 2302199 | Jan 1997 | GB |
| 2331816 | Jun 1999 | GB |
| 2349260 | Oct 2000 | GB |
| WO 9847084 | Oct 1998 | WO |
| WO 9905681 | Feb 1999 | WO |
| 0031723 | Jun 2000 | WO |
| WO 0054168 | Sep 2000 | WO |
| WO 0131627 | May 2001 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 20030177108 A1 | Sep 2003 | US |
