TW394926B - Speech recognition system employing multiple grammar networks - Google Patents

Abstract

The input speech is segmented using plural grammar networks, including a network that includes a filler model designed to represent noise or extraneous speech. Recognition processing results in plural lists of candidates, each list containing the N-best candidates generated. The lists are then separately aligned with the dictionary of valid names to generate two lists of valid names. The final recognition pass combines these two lists of names into a dynamic grammar and this dynamic grammar may be used to find the best candidate name using Viterbi recognition. A telephone call routing application based on the recognition system selects the best candidate name corresponding to the name spelled by the user, whether the user pronounces the name prior to spelling, or not.

Description

A7

V. Description of the Invention (/) Printed bags of the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs This case is part of the US patent application serial number: 08 / 642,766, filed on May 6, 1996, title " continuous call routing "Apparatus", applicants Jean-Claude Junqua and Michae 丨 Galler, continuous. The present invention generally relates to computer-made speech recognition. In particular, the present invention relates to a method and device for processing speech data using a multi-grammar network. The use of multiple networks leads to Different segments of voice data to facilitate the extraction of practical speech from speech that is not practical. Although the present invention has many uses, a spelled name recognition system suitable for telephone call routing application types is described here. In the embodiment, the first and second sets of grammatical networks are used to detect the N best and M best text sequences, respectively. The first set of grammatical networks is assumed that the user will immediately start pinyin in response to a system prompt Configuration. The second grammatical network assumes that the spelled name text sequence is external noise or intonation that the system cannot recognize. Started and configured> Best group of N and best text sequences of group M respectively accept dynamic programming matching for a valid name dictionary to extract groups of N corresponding to each group of best and best text sequences of group N Best and M group best name hypotheses. Then select the best candidate from these name hypothesis sets to form a recognition decision. Current speech recognition technology involves the identification of patterns in speech data and a set of patterns and systems recognized by the system Relevance of predetermined dictionary login sets. Speech recognition problems are quite challenging because of many different variants. Generally speaking, speech recognizers apply digital form of incoming speech data to transform the digital data into parameters according to a predetermined pattern. Mathematical identification procedure. It is customary that this model is pre-trained with a sufficiently large training set. The paper size is applicable to the Chinese National Standard (CNS) Α1 specification (210 × 297 mm) (please read the precautions on the back before filling this page)

1 A7 —______ B7 V. Description of the invention (2) Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs, so that the speaker variation is greatly reduced. Pattern-based recognition programs distinguish incoming data as basic building blocks, such as phonemes, which are then labeled using comparisons with trained patterns. In a type 1 recognizer, once the respective phonemes are marked, the phoneme data is compared with words previously stored in the system dictionary. This comparison is achieved via an alignment process that will accommodate inexact matches due to incorrect phoneme recognition and phoneme insertion and deletion within the given sequence. The system works on a probabilistic basis. It is customary that the speech recognizer will select the most probable word candidate derived from the segmentation, labeling, and alignment processing procedures described previously. Due to their nature, current speech recognizers select word candidates from a pre-defined dictionary and therefore they will only recognize pre-defined words. This poses a problem, especially in systems that make further decisions based on speech recognition results. Alien noise or the intonation of words not found in the dictionary are often incorrectly translated into words found in the dictionary. Subsequent decisions based on such incorrect identification may lead to erroneous system performance. To illustrate this problem, consider a phone call routing application with a spelled name. The user is synthesized with a voice indicating the name of the person who should be directed to call. If the user follows these instructions, the speech recognizer confirms each spoken text and can then align the text order with the dictionary to find the spelled name. The system then uses the routing information found in the dictionary to direct the call to the appropriate extension. However, if the user sends out external information first, such as the pronunciation of a person's name before pinyin, the recognition process is likely to fail. This is because the recognition system only expects to receive a series of spoken texts (please read the notes on the back before filling this page). The size of the paper is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) A7 _____ B7 5. Description of the invention (3) words and will try to "recognize" the name is a set of _ or multiple sets of text. The conventional system does not have the voice data to properly enter, because the basic model of the manufacturing system assumes The data are useful or meaningful equivalent units (the words) for the system. The present invention adopts and combines multiple grammar networks to generate multiple sets of identification candidates, some of which are based on the assumption that a set of patterns of foreign speech exists, And some solve the above problem based on a set of speech recognition systems that assumes that there is no set of patterns of foreign speech. The results of the two sets of patterns are used to select the most likely candidate based on the respective matching probability scores to provide the final recognition Decision "The Central Standards Bureau of the Ministry of Economy The different first and second grammatical networks that are distinguished by different speech materials are processed separately. In this way, the system extracts practical speech from non-practical speech. For each grammar network, most Identification candidates are generated. The preferred embodiment uses a set of first syntax networks to generate N sets of best candidates and a set of second syntax networks to generate M sets of best candidates, where N and M are integers greater than one and Can be the same. The first and second majority identification candidates (N-best, M-best) are transformed based on at least one set of previous restrictions on useful speech. The transformation may include, for example, Dictionary of matching candidates to the spelled name recognized by the system. The recognition decision then depends on the conversion recognition candidate. As will be explained more fully below, the present invention divides the voice data into two or more groups, each of which is Channels handled differently. — Group channels are _ '_6 used in this paper standard based on the hypothetical first grammatical network that supplies only practical tones (eg text) National Standard (CNS) A4 Specification (210X297 mm) A7 _________B7 V. Invention Description u) Printed and processed by the staff consumer cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs. The other channel uses the assumption that there is foreign and useless voice before useful voice. A different grammatical network is processed. Different grammatical networks therefore lead to different data distinctions. The identification candidates generated by each channel can be scored according to the degree to which each candidate matches the respective pattern. Two sets of channels are not required In this step comparison-in order to select the single candidate with the highest score-the two sets of identification candidates are kept separated. In this step, the identification candidates represent the best of N groups and the best text order assumptions of group M. In order to select the most Good candidate hypothesis • The two sets are individually matched to a dictionary of all names recognized by the system. This dictionary is actually a collection of one of the previous restrictions on speech that is useful to the system. It is therefore less likely that certain text order hypotheses can be scored because those text sequences do not match well with the text sequences stored in the dictionary. This preferred embodiment uses the N-best and M-best text sequences to select the N-best and M-best names from the dictionary. The contributions from the two sets of channels are therefore included in the decision-making process. Finally, the group N best and group M best names can be combined to form a simplified set of dictionary candidates to which the input intonation is applied. This reduced-scale dictionary can be used to build a set of dynamic grammars built from the best set of N and best set name candidates of the M set. This dynamic grammar will tend to favor one set of candidates or the other depending on whether the input intonation contains foreign speech. If there is a foreign voice, the grammatical network set to confirm and exclude the foreign voice will tend to produce better recognition results' and these results will be reflected as being composed of the best name candidates in group N and group M A better candidate for dynamic grammar. On the other hand, if there is no ^ 14-- (please read the notes on the back before filling this page), 1T image. This paper size applies the Chinese National Standard (CNS) A4 specification (210 × 297 mm) Employees of the Central Standards Bureau of the Ministry of Economic Affairs Printed by the Consumer Cooperative A7 ______B7 V. Description of the Invention (3 ·) Another foreign language will produce better recognition results, and it will be reflected as a better candidate in dynamic grammar. Once the dynamic grammar is constructed, the input speech data can be processed based on the dynamic grammar using a set of recognizers to extract a single most likely name candidate as a recognized name. This distinguished name is then used to access a suitable set of databases to properly direct the phone call. For a more complete understanding of the invention, its objects and advantages, reference is made to the following description and accompanying drawings. Fig. 1 is a block diagram of an exemplary system using the call routing element of the present invention; Fig. 2 is a block diagram of an exemplary embodiment of the call routing element of the present invention; The state diagram of the configured grammar network G1 at the beginning; Fig. 4 shows the state diagram of the grammar network 02 which is configured assuming that the noise of the spelling name is not in the order of words or the tone of the system is not recognized by the system; Figure 5 is a detailed block diagram of a better recognition system of the present invention; Figure 6 is a graph showing different types of recognition errors; Figure 7 is a diagram showing the optimization of PLP-RASTA filter coefficients to reduce substitution, deletion, and insertion Figure of the number of errors; Figure 8 is a figure showing the improved grid-type N-group best technique; Figure 9 is a figure that further describes how to achieve the hypothesis generated during the identified backward tracking step. The principle of the present invention will be demonstrated by a set of call routing elements and said that the paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) 11- 丨 · ---- Λ ------ IT-- ---- A (Please read the precautions on the back before filling this page) A7 ____; _B7_ V. Description of the invention (?) Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs • It reminds users to use verbatim Spelling the name into the system to provide call routing information. Therefore, in order to help understand the speech recognition system, a summary description of a set of call routing elements that can adopt the speech recognition system will be provided first. It should be noted, however, that the speech recognition system of the present invention is not limited to call routing elements. In contrast, the recognition system can be applied to a wide range of different applications where useful speech must be extracted from external noise or not useful speech. System overview and basic operation Call routing elements using continuous speech recognition will be shown in a set of embodiments that are suitable for plug-and-play connections to existing PBX switches, or incorporated into PBX equipment during manufacturing. Referring to Fig. 1, the PBX switch 210 is connected to the telephone network substructure 212 with conventional devices, such as a telephone line 214. For convenience, three sets of lines are shown in the illustrated embodiment. This is not intended to be a limitation of the present invention, since the present invention can be used in a system having a larger or smaller number of telephone lines. The PBX switch is a conventional design that directs incoming calls from network 212 to any selected telephone element, such as telephone 216. The spelled name identification call router 218 of the present invention is connected as if the telephone 216 is connected to another extension or access on the PBX switch 210. As will be discussed more fully, the preferred embodiment is connected to the PBX switch via a plurality of lines 220 carrying voice traffic and via another line 222 carrying control logic signals that allow the call router to operate integrally with the PBX system. Figure 2 shows the call router 218 »PBX switch 210 and lines 220 and 222 in more detail at the same time. According to the structure of the PBX system, the paper size applies to the Chinese National Standard (CNS) Α4 specification (210 × 297 mm)

1T (Please read the notes on the back before filling this page)-Installation. V. Description of the invention (7) A7 B7 Printed by the Consumers' Cooperative of the Central Government Bureau of the Ministry of Economic Affairs The router 218 can be configured in many different ways. In the embodiment shown, the call router has three separate audio channels connected to three sets of lines 220, respectively. Of course, the number of channels required will depend on the structure of the telephone system. Three sets of channels are shown here to show how one set of systems can simultaneously provide identification of three callers' names on each of the three sets of incoming telephone lines 214. To support additional callers, additional audio channels may be included or multiplexing circuits may be included to allow channels to be shared. Each audio channel has a set of digital signal processors (DSP) 224 and related analog-to-digital / digital-to-analog conversion circuits 226. The digital signal processor is coupled to a main processor 228 containing a set of data storage 230 that stores all references or names. The data storage 23 may be any suitable digital storage medium, such as random access memory. The data storage 230 stores a continuous speech recognition dictionary 'of all names that can be recognized by the system, and related telephone exchange numbers. As will be explained more fully below, the preferred embodiment uses a special set of speech recognizers, which are optimized recognizers for speaker-independent recognition of consecutive spelled names. There is also a set of call switching logic 232 coupled to the main processor 228 (or part of the main processor). This switching logic is connected to the signal line 222 and communicates with the PBX switching system according to the communication protocol specified by the PBX switch. Before proceeding to the detailed description of the speech recognizer, it is beneficial to briefly explain the operation of the call router 218. See Figures 1 and 2. When a group of incoming calls arrive at the PBX switch via a group of telephone lines 214, it can be handled by an operator without intervention by the call router of the present invention. But (Please read the notes on the back before filling this page) | In _ Dingliang-Private paper size applies Chinese National Standard (CNS) A4 specification (210X297 mm) Printed by A 7 ______B7 V. Description of Invention (2) Yes, if the operator is unable to handle the call (for example, the call enters without an operator after normal working hours), the PBX switch is planned to route the call to the call router 218. This can be achieved simply by the switch using one of the audio channels (to a group of lines 220) to schedule the call to the calling router according to the switching instruction transmitted on line 222. If desired, the PBX switch can be planned to jump to different signal lines on different audio channels within router 218 when the first line call is busy. After this is done, the incoming call then communicates with the selected group of DSP processors 224. The processor supplies any required audible prompts to the incoming caller (requiring the caller to spell out the name of the person required) and it also handles the name response spelled by the caller. The speech recognition algorithm used by the DSP processor 224 will be described in detail below. As part of the recognition process, the DSP processor 224 downloads a set of backup shared speech recognition resources from the main processor 228, which reflects all the Information about the names of the referees and their associated phone extensions. When N sets of optimal strategies are used for real-time recognition, the speech recognizer produced by the DSP selects the most likely candidate from the data memory 230. This candidate's name is spoken back to the caller using a DSP processor to supply a set of speech synthesis signals or replay a pre-recorded audio signal of the selected person's name. The caller is then asked to answer " yes " or " no " indicating whether the candidate is correct. If so, the main processor 228 uses the call switching logic 232 to instruct the PBX switch to transfer the call from a group of lines 220 to one of the selected telephones 216. After the switching occurs, the audio channel of the calling router can be used again Handle new incoming calls. 11 This paper size is in accordance with Chinese National Standard (CNS) A4 specification (210X297 mm) (please read the precautions on the back before filling this page). A_ Printed by the Consumers Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs A7 ______B7 V. Invention Explanation (?) Details of better speech recognition processing This better speech recognition system can be used as a multi-round step, where the final round is used only when the previous (aligned) round did not produce a single set of recognition names as output. The first and last rounds use hidden Markov pattern recognition, while the alignment rounds align with dictionary dynamic programming. As will be discussed more fully, the first round (recognition of hidden Markov patterns) is itself divided into a number of parallel sub meridians. In Figure 5, the first, second, and third rounds are shown. "It should be noted that the first round bifurcation becomes a separate hidden Markov pattern recognition block 26a and 26b. The illustrated embodiment is designed to recognize a continuously spelled name that includes a sequence of text supplied to the recognition system via input from a caller's telephone 10. To show useful and useless input examples, two sets of telephones 10 are shown. The caller uses the system correctly in a group of phones to provide a text sequence: H-A-N-S-0-N. In another set of telephones, the caller used the system by saying the spelling name and text sequence: "Hanson" H-A-N-S-0-N. As will be explained below, the system is designed to accommodate both correct and incorrect use, resulting in a more robust identification system. The identification system shown in 12 includes a set of name extraction systems shown in 13. As will be discussed, the name extraction system has the ability to form a dynamic grammar that represents a selected subset of the items found in the name dictionary. "Dynamic grammar is used when identification is not reached in the second round and processing proceeds to the third round . The input sequence of text can be fed to a suitable set of speech analysis modules 14. This module is optimized for front-end to reduce substitution, deletion and insertion errors. The 12 ^ 's scale is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) I -1-- — I-- 装 ----- -Order ------ 4 · (Please read the note at the back of the item ^ 3; write this page) A7 B7 printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs 5. The number of invention descriptions (/ 0). In consecutive spelled names, replacing errors is replacing a set of incorrect text with the correct text. Figure 6 shows examples of substitution errors in the identification of the spelled name JOHNSON at 16 and 18. Deletion error is the disappearance of one or more groups of words in consecutive spelled names. This is shown at 20 in Figure 6. The insertion error is that the consecutive spelled names contain additional text that was not previously spelled. Examples of insertion errors are shown in Figures 22 and 24 of Figure 6. The speech analysis module 14 is designed to operate on digitized speech data. So if a set of analog voice input systems is used, the analog signal should be digitized first. This can be achieved using appropriate analog-to-digital circuits that can be included in the speech analysis module 14. The preferred speech analysis module uses a set of 8-stage PLP-RASTA processing programs to compensate for the effects of the communication channel. For more information on PLP-RASTA compensation, please refer to H. Hermansky, N. Morgan, A. Bayya, and P. Kohn, EUROSPEECH'91, pp. 1367-1370, 1991. The example uses a set of 10 millisecond box shifts and a set of 20 millisecond analysis windows. The RASTA filter coefficients are optimized to reduce the number of substitution, deletion, and insertion errors. The optimal filter coefficient was chosen to be a fraction of 0.90. In determining the optimized RASTA filter coefficients, the first derivative of energy, the energy, and the first derivative of the static cepstrum coefficients C1 to C8 (calculated above the 7-tone frame) are interactively combined with the static cepstrum coefficient to form Speech parameter representation (18 coefficients in total). Figure 7 shows optimized RASTA filter coefficients that will reduce the number of substitutions, deletions, and insertion errors. In this figure, 1 > 1 ^-

rasta ^ mmm '> mmmmmww.nk'SL 13 This paper size is applicable to China National Standard (CNS) A4 specification (2 丨 0X297 mm) ---- I--I ------ ^ --- --- 1T ------ 2 /.·»<· (Please read the notes on the back before filling this page) A7 B7 V. Description of the Invention (//) The first derivative of the static cepstrum coefficient combination. Although PLP-RASTA optimization is currently the better one, other forms of optimization can be used. For example, a dark frequency cepstrum coefficient (MFCC) analysis may be used in addition. A 14th order MFCC analysis may be used to obtain appropriate results. In the MFCC analysis, 11 sets of static cepstrum coefficients (including C0) are calculated with a frame shift of 16 milliseconds and an analysis window of 32 milliseconds. "Different recognition accuracy can be obtained using different feature sets." These characteristics Sets can contain static and dynamic features separately and in combination. In order to demonstrate the robustness of the parameters used in the present invention, erasure and filtering data are used. In order to obtain filtered data for testing in the preferred embodiment, a set of distortion filters is used and the test data is filtered to manually generate a set of mismatches between the training set and the test set. In this regard, see H. Murveit 1 J. Butzberger and H. Weintraub, Darpa 讨 Symposium, Speech and Natural Language, pp. 280-284, February 992. Printed by the Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs (please read the notes on the back before filling out this page). Return to Figure 5. The output of the voice analysis module 14 is divided into two groups of channels. The Markov pattern-related recognition block 26a and another set are the recognition block 26b »recognition blocks 26a and 28a related to a set of pre-defined text syntaxes G1-, which function as hidden Markov patterns. The different sets of pre-defined text grammars G2—functions shown in the recognition blocks 26b and 28b constitute the grammatical networks shown in Figs. 3 and 4, respectively. These grammatical networks are graphs containing nodes related to various possible texts and possible transitions from node-to-node. These grammars include silent nodes followed by text loops, where any text can follow any text. The third grammar G1 starts with a set of Sil nodes (Sil) 50, and is transferred to a separate paper. This paper size applies the Chinese National Standard (CNS) A4 specification (210X297). Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs. System A7 ___B7 V. Description of the invention (A?) Initial text A, B, C * »The grammar G2 shown in Figure 4, with a set of filler nodes 52 representing foreign speech or miscellaneous words before Pinyin News began. The charger node is transferred to the mute node 52 and then to the respective text node, such as G1- In this preferred production, the identification blocks 26 a and 26 b are frame-synchronized, first-order, continuous-density hidden markoff using Viterbi decoding Pattern recognizer. This preferred embodiment uses a modified Viterbi decoder that produces N-best or M-best hypotheses (instead of a single hypothesis). The Viterbi decoder is usually designed based on the probability of a match between the HMM mode and the test tone. Provide only the best assumptions. In the present invention, this standard Viterbi decoder is modified so that it provides the N-best or M-best hypotheses depending on the highest probability of matching between the HMM pattern and the test tone. The identification blocks 26a and 26b each produce their unique N-group best or M-group best hypotheses. Although the same number is used in the preferred embodiment (for example, N = M = 10), the two sets of identification blocks need not generate the same number of hypotheses if desired. Therefore, in Fig. 5, the identification block 26a generates N sets of best hypotheses and the identification block 26b generates M sets of best hypotheses. As stated previously, the symbols N and M can be any integer greater than one. The exact value chosen for the integers N and M may depend on the processor speed and memory size. The technique for generating the best (or best in the M group) text candidate for the N group will be discussed more fully below. It will be appreciated that the technique for generating the hypothesis of group N (or group M) is roughly the same in both cases. The hidden Markov pattern recognizer adopted in 26a and 26b has a beam search capability designed to limit the search space, so that the recognizer can process the incoming speech more quickly. 15 This paper size is applicable to China National Standard (CNS) A4 specification (210X297 mm)-: --- 1 --------- 1T ------ ^ j · (Please read the note on the back first Please fill in this page for further information.) Printed by the Central Standards Bureau of the Ministry of Economic Affairs, Shellfish Consumer Cooperatives. A7 _____ ^ _ B7_ V. Description of the invention (0) A set of points for the possibility of matching between the input voice and the reference voice. When there is no beam mechanism, the recognizer must score all possible channels in each frame during the search procedure. With a beam search identifier, only those channels whose scores are offset from the best score by an amount equal to the beam width are considered. Instead of finding the entire search space, make a beam search so that the least likely search channel is trimmed so that only the best hypothesis is available. The N-best (or M-best) assumptions obtained from the identifiers 263 and 26b are then sent to the dynamic program planning (DP) alignment modules 38a and 38b, respectively. The dynamic programming alignment module has access to a set of related name dictionaries 39 for comparing one of the N-best (or M-best) hypotheses. Dynamic programming is used to plug, replace, and delete errors. In some examples, the result of dynamic programming alignment will generate a set of single names without other candidates. ”When DP alignment only generates a set of candidates, the decision strategy module 40 detects this name and provides the distinguished name as an output. . However, in most cases, no single candidate is generated. In this case, the decision strategy module transmits the N-best and M-best assumptions to the module 42 to establish a set of dynamic grammars. Module 42 uses the N-best and M-best candidates provided by the DP alignment module to create a set of grammars. The highly constrained recognizer 44 is then activated to use the dynamic grammar 42 to measure the best set of N and the best candidates of set M. The recognizer 44 may also be a set of hidden Markov pattern recognizers. Even if highly restricted, because the dynamic syntax is small and because the parameter representation (calculated in 14) does not need to be recalculated, the data transmitted via this recognizer is not time-consuming. ”If necessary, a set of neural network discriminators Can be applied to the identifiers 26a and 26b or 16 This paper size is applicable to China National Standard (CNS) A4 specification (210X297 mm) j-- 丨 M ---- install -------- order ----- -扃 (Please read the precautions on the back before filling this page) A7 ___B7__ 5. Description of the invention (/ 4) The output of the identifier 44. The list in Appendix A shows how the system of the present invention recognizes the spelled name WILSON. The part designated as [First Pass] in the list shows all hypotheses generated by the two syntaxes. It does not have the name WILSON. In the section labeled [DP Alignment], the top (best) candidate is listed: the name WILSON (the first candidate out of 10) is included in the list. In the section labeled [Costly Constrained Pass], the input intonation is compared only with the candidate selected during DP alignment. In this case, the recognizer correctly detects the name WILSON. Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economics and Technology of the N Group (please read the notes on the back before filling this page)-The best candidate of the N group or the best candidate of the M group uses the best choice of the N group Algorithm. For technical details, see R. Schwartz and Steve Austin, "Effective, High-Performance Algorithms for the Best Search in N Groups," DARPA Speech Recognition Symposium, pp. 6-11, 1990. In recognition, the incoming speech data is decomposed into time frames and analyzed on a frame-by-frame basis. There are many possible hypotheses for any given intonation. The preferred N-best (or M-best) algorithm selects the best start time for a group of characters based on previous characters only and not on characters preceding the previous characters. As each text is spoken and analyzed, the hidden Markf pattern recognizer will generate probability scores for each pattern. Because the system's final purpose is to select the most probable text sequence | the system stores the most channels that represent possible combinations of text. In order to make the system better as a set of real-time identifiers • Two different sets of Ί7 This paper size applies to Chinese National Standard (CNS) Α4 specifications (210 × 297 mm) Printed by A7 _ B7 of the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs The data pruning level of the invention description (/ 5) is made. Two-level pruning techniques involve comparing the probability of a given hypothesis with a set of probability thresholds. If the probability given to the channel is below the threshold, it is discarded. More specifically, pruning occurs at a regional level and at a wide area level. Regional level pruning involves discarding channels that represent low-probability matches at the text level and wide-area pruning involves discarding channels that represent low-probability matches from the beginning of the intonation to the last found text. Therefore. At the end of the spelled name, a roving backtracking is reached to extract the N-best (or M-best) name hypothesis. When backtracking operations are performed, regional and wide-area pruning has reduced the size of the memory space that needs to be analyzed. In addition to regional and wide-area pruning, the preferred system can also use an adaptive threshold so that the pruning threshold is dynamically adjusted during system execution. In the case of confusing words, the recognizer uses state links to help focus the distinguishing part of words and reduce the number of evaluation parameters. These links are (m, η), (1, r), Cp, t), and (b, d). In the preferred embodiment, all characters except the character W are represented in a six-group state HMM mode. The text W is expressed in twelve groups of states in the HMM mode and the mute mode is expressed in one group of states. Depending on the degree of text obfuscation, the text patterns have different numbers of Gaussian densities. " E-group " texts ... b, c, d, e, g, p, t, v, and z, and texts m, n, s, and f are all modeled with 6 sets of Gaussian density. The rest of the text is modeled with 3 sets of Gaussian density. Fig. 8 shows another preferred technique of the present invention for performing N-best (or M-best) hypothesis analysis. This is called the grid-type N group best technology. It is applicable to Chinese national standard (〇 阳) 8 4 size (210 乂 297 mm) on 18 paper sizes. I- ^ 丨 丨 ------- f-- ---- IT ------ Λ. (Please read the precautions on the back before filling out this page) Printed by the Consumers' Cooperative of the Central Government Bureau of the Ministry of Economic Affairs Α7 Β7 V. Description of Invention (/ Office) Each frame This step calculates the probability of each grammar node and stores the best response to reach that node. It then stores the probability and the number of frames that are supposed to work. This technique therefore preserves the N-group best (or M-group best) hypothesis and the best group is transmitted by the node, all other groups are included by the most likely channel.

The Viterbi transmission algorithm calculates the probability of each state. This is done in a box-by-box manner with all the data in the input buffer, and the probability is stored in the state data structure. The preferred lattice-type N-group optimal technique is a modified Viterbi algorithm that produces N-group best (or M-group best) candidates, but only transmits the maximum probability in the next mode. So in each frame, the routine calculates the probability of each grammatical node and then stores the best response into that node. Referring to Figure 8, a group of network nodes n are shown. Three sets of hypotheses W ,, " 3 enter node η. Of these assumptions, only the highest probability (highest probability) is sent forward. Therefore, the node η is assumed to produce the word hypothesis Wj, Wk, and W, from the maximum probability of the node η. The Viterbi transmission algorithm stores a set of indicators of probability, persistence (currently assumed number of active boxes), and hypotheses that generate specific network nodes. When the probabilistic data in the state data structure is analyzed, this information is used by the backtracking algorithm. Refer to Figure 9 for the backtracking algorithm. This preferred backtracking algorithm ranks all N best ending nodes into a set of priority queues, shown in the vertical row in line 9 280 »If, for example, in the last box of the first round When 10 sets of hypotheses are transmitted in (N = 10), there will be 10 sets of end nodes (η〆n2, ..., η ,.) in the priority queue 280 "Conclusion 9 This paper standard applies Chinese national standards (CNS) Α4 specifications (210X297 mm) (Please read the notes on the back before filling this page) Order Λ. A7 __B7 V. Description of the invention (, 7) Bundle nodes are sorted in descending order, making the The first ending node represents the group with the highest probability score. For the sake of brevity, it is assumed in Fig. 9 that node 1 ^ is the node with the highest score. This node was traced back in the past and expanded to place the hypothesis (W, w2 or w3) that gave the given node '. When confirming this hypothesis, it is allowed to trace back the routine to confirm the node that generated the identification hypothesis. In Figure 9, if it is assumed that _W2 is responsible for generating node η, then node ^ is identified by backtracking. Node nb then replaces node η, somewhere in priority queue 280. After replacement, the priority queues will be maintained by categories in descending order. If the replaced new node nb has the highest score, it will occupy the position previously occupied by node n ,. Of course, a group of other nodes in the priority queue may also have a higher score than the new node being replaced. The node with the highest score in this case, and not replaced by the new node nb, will be used in the next backtracking operation. Backtracking proceeds in the above manner until a set of start nodes (a set of nodes corresponding to the first frame of speech) are encountered during the backtracking. When a set of starting nodes is reached, a set of hypotheses is discovered. The backtracking algorithm stores sets of symbols encountered during backtracking. These symbols can be stored as a string of symbols read in reverse to produce the most likely candidate names. Printed by the Consumer Cooperatives of the Central Government Bureau of the Ministry of Economic Affairs. Figure 9 shows the general situation with multiple N-group optimal assumptions (N = 10). To further illustrate the priority queue backtracking steps, see the example in Appendix B. This example shows the backtracking steps for the name " JONES ". From the description above, we can understand that the call routing element of the present invention is extremely suitable for several 20 paper sizes. It is applicable to the Chinese National Standard (CNS) A4 specification (210X 297 mm) A7 B7 Printed by the Shell Standard Consumer Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs Note (/ 5) Plug-and-match connection for almost any office telephone network or PBX system. The routing element employs a skilled speaker-irrelevant, continuous speech recognition technology that allows incoming callers to spell the required recipient name and the system will automatically determine the appropriate telephone extension for the recipient and enable the existing telephone network Or the PBX system connects the incoming caller to the recipient's extension. The invention does not need to communicate with the telephone network via complicated key commands, thus making the system extremely suitable for the visually impaired. This speech recognition technology is highly resilient: incoming callers can spell it out at their unique natural pinyin rate and the system will automatically prompt the user to confirm the name selection after the caller exits pinyin. If the incoming caller pauses significantly and then continues to pinyin, the system will automatically determine the speech recognition. "This multi-round speech recognition step works well even through a noisy phone channel." This step is lost between rounds. N sets of best hypotheses and delay the most expensive calculation steps until the final round, at which time the list of candidates for possible names is greatly reduced. As a result of the best N-group, multi-round identification procedure, the present invention can be manufactured using low-cost DSP circuits. Although the invention has been described in its preferred embodiments, it will be understood that the invention may be modified without departing from the spirit of the scope of the appended patents. Therefore, for example | different configurations can be connected with different types of current and future telephone systems, including analog systems and digital systems '': ------------------- (Please read the note on the back first Please fill in this page for further information), tr This paper size applies Chinese National Standard (CNS) A4 specification (210X297 mm) V. Description of the invention (/ y) A7 B7 Component reference table 10 Caller phone 12 Identification system 13 Name removal System 14 Speech analysis module 26a, 28a Identification block 26b, 28b Identification block 38a, 38b Dynamic programming planning alignment surface 39 Name dictionary 40 Decision strategy module 42 Dynamic syntax establishment module 44 Identifier 210 PBX switch 212 Fiber under the telephone network 214 telephone line 216 telephone 218 call router 220 line 222 line 224 digital signal processor 226 analog-to-digital digital-to-ratio conversion circuit 228 main processor 230 data storage 232 call switching logic M ^ — (Please read first Note on the back, please fill out this page again) Order Λ Printed by the Central Consumers Bureau of the Ministry of Economic Affairs Printed by the Consumer Cooperatives 22 This paper size applies to Chinese National Standards (CNS) Α4 specifications (210X297 mm) 5. Description of the invention (Μ;

FIRST PASS G1 grammar: A7 B7

APPENDIX A

Hypothesis 1: ocfeylson Hypothesis 2: onseylson G2 grammar: Letters spotted at frame 104

Hypothesis 1: wylson (Please read the notes on the back before filling out this page)

DICTIONARY ALIGNMENT PASS N-best candidates from Gl:

, 1T of of of of of of 8: 8: 8: 8: 8: 8: 8: 8:

Candidate

Candidate

Candidate

Candidate

Candidate

Candidate

Candidate

Candidate neilson masterson nielson andersson carlson nelson anderson patterson M-best candidates from G2:

Candidate 1 Candidate 2 Candidate 3 Candidate 4 Candidate 5 Candidate 6 Candidate 7 Candidate 8 Candidate 9 ooooooooo lllllxxxl fffffffff ooooooooo

Wilson walton wasson watson nelson folsom urmson bylsma olson 23 This paper size applies to Chinese National Standard (CNS) A4 specification (210X297 mm) A7 B7 V. Description of the invention (2 /)

Candidate 10 of 10: sisson

COSTLY CONSTRAINED PASS

Hypothesis 1: wilson 310 frames (3.1 seconds) in signal. N II i I Binding II Binding — ^ | 旅 '.: (Please read the precautions on the back before filling out this page) Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs 24 This paper size applies to China National Standard (CNS) A4 specification (210X297 mm) A7 B7 V. Description of invention (n)

Appendix B

Example of Priority * Queue Backtrack Procedure for " JONES " • Last Frame Number is 10; Three Hypothesis Nodes Saved 1. node 6 (s) 2. node 6 (r) 3. node 4 (d) prob 0.9 prob 0.8 prob 0.6 duration 18 frames duration 20 frames duration 12 frames prev. node 3 prev. node 3 prev. node 2 frame # 100 frame # 100 frame # 100 • Build Priority Queue: (Order of Decreasing Probability): (Please read the note on the back first (Please fill in this page again)

Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs node 3 (e) prob. 0.9 (prob. Of parent) duration 10 prev. Node 1 frame: 100-18 = 82 node 2 (a) prob. 0.7 prob. Of parent ( 0.9) duration 10 prev. Node 8 frame: 100-18 = 82

Insert New Hypothesis Node in Priority Queue:

(Children Replace Parent Node in Queue) 25

Expand Maximum Likelihood Node, Extending Sequence Backward: This paper size applies to China National Standard (CNS) A4 (21 × 297 mm) Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs A7 B7 V. Description of the invention (θ)

Appendix B (cont.)

Begin Procedure BackTrack {Initialize backtrack priority-queue Q}

For each grammar-terminal state S Begin

If S has active list hl, h2 ..... hn of hypotheses in final frame T

Begin

For each active hypothesis h Begin generate node N FLscore <-h.score N.sequence <-h.symbol N.duration <-h.duration N.predecessor <-h.predecessor N.time <-T enqueue N in Q End For

End If End For {Process priority-queue Q, generating n-best sequences.}

NumSequences <-0

While Queue nonempty and NumSequences < n Begin

Dequeue first (top-scoring) node N from Q

If N's predecessor is grammar-initial state Begin send N.sequence to output NumSequences <-NumSequences + 1

End If {Expand N to generate child nodes.

Add child nodes to priority queue Q} T <-N.time-N.duration S <-N.predecessor

For each active hypothesis h for state S in frame T Begin

generate node C C.score <-N.score-(best score for S in frame T-h.score) Appendix B (cont.) C.sequence <-concatenation of h.symbol and N.sequence C.duration <-h.duration II C.predecessor <-h.predecessor 'C.time <-T enqueue C in Q End For 26 This paper size applies Chinese National Standard (CNS) A4 specification (2 丨 〇X297 mm ) --- 1 ------- install --- '(Please read the precautions on the back before filling in this page), -ιτ

A A7 B7 V. Invention Description U4)

End WhileEnd Procedure BackTrack (Please read the precautions on the back before filling out this page)

.1T Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economics 27 This paper size applies to China National Standard (CNS) A4 (210X297 mm)

Claims (1)

A8 B8 C8 D8 VI. Patent Application Scope 1. A method for processing speech data for word recognition, which includes: processing the speech data using a set of recognizers according to the first grammatical mode to extract the first set of most recognition candidates : Using a set of recognizers to process the speech data according to a second grammatical mode different from the first grammatical mode to extract a second set of majority recognition candidates; the first set of majority recognition candidates and a predetermined word dictionary Align to generate a first list of word candidates; align the second set of plurality of recognition candidates with the predetermined word dictionary to generate a second list of word candidates; from the first sum of word candidates The second list establishes a set of dynamic grammar patterns: and uses a set of recognizers to process the speech data to extract recognition words based on the dynamic grammar patterns. 2. The method of claim 1 in the scope of patent application, wherein the first and second grammatical modes are network modes including a plurality of interconnected text modes. 3. The method according to item 2 of the patent application, wherein the text mode is represented by a hidden markoff mode. Printed by the Central Bureau of Standards of the Ministry of Economic Affairs, Shellfish Consumer Cooperatives 11 ^ ------- Wood-I ——II Nails (Please read the note on the back before filling in this page) 4. If the scope of patent application is the first Item method, wherein the first and second grammatical modes are a network mode including a plurality of interconnected text modes and the second grammatical mode includes at least a tone representing a tone that is not defined by the text mode—group charger mode 5- The method according to item 1 of the scope of patent application, wherein the first and second grammatical modes are a network mode including a plurality of interconnected text modes and the second grammatical mode includes at least a tone representing an undefined text mode 28 paper sizes of this paper use the Chinese National Standard (CNS) A4 specification (210X297 mm) Printed by A8 B8 C8 D8 of the Bayong Consumer Cooperative of the Central Rubbing Bureau of the Ministry of Economic Affairs 6. Scope of patent application—group charger model and representative of this A paused set of silent modes in audio data. 6. The method of claim 1 in the scope of patent application, wherein the second syntax mode defines a text layout syntax. 7. The method of claim 1 in the scope of patent application, wherein the first and second grammatical patterns include a plurality of different nodes and wherein the first and second majority recognition candidates are matched according to the voice data The nodes are extracted by a recognition program that scores the node; and wherein the first and second majority candidates are selected by selecting the node that is scored and has the best match for the speech data. 8. The method according to item 1 of the patent application scope, wherein the second syntax mode includes at least one set of nodes for representing noise. 9. The method according to item 1 of the scope of patent application, further comprising using the recognition word to guide a group of telephone calls. 10 · —A method for processing voice data for recognition of the spelled name, which includes: According to the first grammar mode Use a set of recognizers to process the speech data to extract a first set of text candidates: use a set of recognizers to process the speech data to extract a second set of text candidates according to a second grammatical pattern; Align a plurality of text candidates with a predetermined name dictionary to generate a first list of name candidates; Align the second plurality of text candidates with the predetermined name dictionary to generate a second list of name candidates; 29 pieces of paper Standards apply to China National Standards (CNS > Λ4 specification (210X297 mm) I---·· * 1 ....... ·---I---1 ^ .Water-_lli--— II 1 C Please refer to the notes on the back of Mt * before writing this page) A8 B8 C8 D8 The scope of patent application is to build a set of dynamic grammar patterns from the first and second lists of name candidates: A group of recognizers process the speech data to extract the distinguished name. "For example, the method of patent application No. 10, wherein the first and second grammatical modes are different." 12, such as the method of patent application No. 10 , Where the first and second grammatical patterns are network patterns containing a plurality of interconnected text patterns. 13. The method according to item I2 of the patent application, wherein the text mode is represented by a hidden markoff mode. 14. The method of claim 10 in the scope of patent application, wherein the first and second grammatical patterns are a network pattern including a majority: an interconnected text pattern and the second grammatical pattern at least includes a representation that is not defined by the text pattern. A tone set of charger modes. 15. The method of claim 10 in the scope of patent application, wherein the first and second alternative patterns are network patterns including a plurality of interconnected text patterns and the second grammatical pattern includes at least a representation that is not defined by the text pattern A set of charger modes for the intonation and a set of silent modes representing pauses in the voice data. Printed by the Shelling Consumer Cooperative of the Central Rubbing Bureau of the Ministry of Economic Affairs (please read the note “$” on the back before filling out this page) 16. For the method of applying for item 10 of the patent scope, where the second grammar mode defines a text configuration grammar . 17. For the method of claim 10, in which the first and second grammatical patterns include a plurality of different nodes and in which the first and second plurality of recognition candidates are matched according to the voice data Sections 30 This paper size is applicable to China National Standards (CNS) A4 specifications (210X297 mm) 6. Scope of patent application ABCD Central Government Standards Bureau of the Ministry of Economic Affairs prints dots on the printing and marking of this node to identify this node The program extracts; and wherein the first and second plurality of identification candidates are extracted by selecting a node that is scored and has a best match with the voice data. 18. The method of claim 10, wherein the second syntax mode includes at least one set of nodes for representing noise. 19. The method of claim 10, further comprising using the recognition word to direct a group of telephone calls. 20.-A method for processing speech data for recognition, comprising: processing the speech data using a set of recognizers according to a set of first grammatical networks to distinguish and extract a first set of majority recognition candidates according to a set of first speech input criteria By: using a set of recognizers to process the speech data according to a set of second grammatical networks to distinguish and extract a second set of majority recognition candidates according to a set of second speech input criteria; converting at least one set of predictive restrictions on the speech data The plurality of identification candidates of the first group and the second group become converted candidates; an identification decision is formed according to the converted candidates. 21. A method for processing speech data for recognition, comprising: processing the speech data separately using different first and second grammatical networks that cause different distinctions of the speech data to extract practical speech from speech that is not practical : Using the first grammatical network to generate a first set of majority recognition candidates and using the second grammatical network to generate a second set of majority recognition candidates: according to at least one set of predictive limit conversions of the practical speech data (Please read the note on the back before filling in this page). Name · Order-A8 B8 C8 D8 VI. Patent application scope The majority of the first and second groups of identification candidates to generate converted identification candidates ; And form an identification decision based on these converted identification candidates. (Please read the note on the back before filling in this page) Order Printed by Zhengong Consumer Cooperative, Central Bureau of Standards, Ministry of Economic Affairs This paper size applies to China National Standard (CNS) A4 (210X297 mm)