CN1223984C - Client-server based distributed speech recognition system - Google Patents

Abstract

Generally, new client-server based distributed speech recognition (DSR) systems provide an efficient method for recognizing human speech on a client device and transmit it over a network to a remote server. This system distributes the speech recognition processing between the client and server, allowing a speaker-related language model to be utilized, resulting in higher accuracy compared to traditional DSR systems. Accordingly, the client device is configured to generate a phonetic glossary by performing speech recognition using a voice model trained by the same end-user whose speech is to be recognized. The obtained phonetic glossary is sent to the server, which performs the speech processing and generates the recognized word string. Compared to designs using traditional DSR, this new DSR method and system achieves a 2-3 times lower word error rate, resulting in a more accurate recognition system.

Description

Distributed Speech Recognition System Based on Client-Server

technical field

The present invention relates to distributed speech recognition (DSR) systems and architectures. More specifically, the present invention relates to a new DSR system and method that performs the sound processing portion of speech recognition at a client device and the language processing portion at a server device.

Background technique

Since the idea of modern computers, engineers and linguists have worked together to make human speech recognition perfectly realized by a machine. One goal of automatic speech recognition is to enable a system to take input human speech, convert it to a recognizable form, and perform useful functions with the recognized speech.

Currently, there are various commercial applications for speech recognition technology. A dictation machine, for example, can "listen" to a person's dictation and transmit the "heard" text to a monitor in "real time". Another application involves machines capable of receiving and executing control commands issued by a human voice rather than via a mouse or keyboard. For example, a person can say "read my email" to a computer. The application can use speech recognition technology to recognize word strings uttered by the speaker. A series of commands to perform the desired task can then be issued, causing the computer to read the person's email.

Another application has been developed for client-server based speech systems and frameworks. Typically, the task of speech recognition is distributed between the client and the server. For example, a mobile phone or personal digital assistant (PDA) can be used as a client that captures speech, obtains speech characteristics, and sends the characteristics to a server at a central location. This communication can take place over a network such as the Internet. Once the speech features are received by the server, they are processed for voice recognition and language processing for the given person's language used.

More specifically, human speech is captured at the client side by a device such as a microphone. The speech signal is converted to digital form so that it can be analyzed by a digital computer. The digital signal is passed through a feature extraction module, which will extract the acoustic features of the speech signal, such as energy concentration at periodic sampling points. The extracted features are then quantized by mathematical models such as Mel Frequency Cepstral Coefficients. The quantized features are organized into a data packet for sending to a server.

The server then receives the data packets containing the quantified features and performs sound and language processing to provide a word string. Since the server serves multiple clients, the sound processing is simulated by a speaker-independent (SI) model.

Among the shortcomings of the traditional DSR approach is its inability to take advantage of the improved word error rate (WER) provided by speaker-dependent (SD) models. The difference between the two models is that one SD model has been trained on a specific person's speech, and as a result the WER for that specific person is lower. This is because people from different linguistic backgrounds produce significantly different vocal signals for the same word. People from different regions may have different accents and pronunciations.

In contrast, when the system is used by various speakers, such as an automated teller machine (ATM), an SI model is used and it is specified to handle any speaker regardless of the speaker's linguistic characteristics, such as the speaker's Pronunciation, the variation in speech due to gender and age and the intensity of the speaker's voice, the SD model has 2-3 times lower WER than the SI model. Since the traditional DSR method processes the voice at the server instead of the client, it is unrealistic and inefficient for the system architecture to adopt the SD voice recognition model with lower WER to improve the overall recognition accuracy.

Description of drawings

Fig. 1 shows a block diagram of a schematic communication network employing the new DSR method of one embodiment of the present invention.

FIG. 2 shows a block diagram of a schematic client-server based DSR system employing the new DSR method of one embodiment of the present invention.

Fig. 3 shows a schematic diagram of a new DSR method employing one embodiment of the present invention.

FIG. 4 shows a schematic diagram of the new DSR method using an embodiment of the present invention at a client node of a client-server based system.

FIG. 5 shows a schematic phonetic word graph generated at a client node using a method according to an embodiment of the present invention.

Fig. 6A shows a schematic flow chart of sending processing of a phonetic symbol graph according to an embodiment of the present invention.

FIG. 6B shows a schematic datagram for sending a schematic phonetic graph of a method according to an embodiment of the present invention.

FIG. 7A shows a schematic diagram of a new DSR method of a method according to an embodiment of the present invention at a server node based on a client-server system.

FIG. 7B shows a schematic phonetic graph generated at a server node of a client-server network system according to a method according to an embodiment of the present invention.

FIG. 7C shows a schematic phonetic word map extended from the phonetic word map shown in FIG. 5 .

FIG. 8 shows a block diagram of a schematic system employing the new DSR method according to one embodiment of the present invention.

Detailed ways

In the following detailed description of the invention, various specific details are given in order to provide a thorough understanding of the invention. It will be apparent, however, to one of ordinary skill in the art that methods in accordance with the present invention may be practiced without these specific details. In other instances, well-known methods, processes, components and circuits have not been described in order not to obscure aspects of the invention.

The method according to the invention comprises various steps which will be described hereinafter. The steps may be implemented by hardware components, or may be embodied in machine-executable instructions, which can be used in a general purpose processor programmed with the instructions to perform the steps. Also, this step can be performed by a combination of hardware and software.

The present invention discloses a new DSR method which differs from conventional DSR methods and achieves improved recognition accuracy. This new DSR approach takes advantage of the lower WER associated with SD sound recognition models. This is accomplished by splitting and distributing the speech recognition processing to speech recognition at the client device and language processing at the server device. So after a client device captures speech, it performs sound processing according to an SD personalized sound model. This process obtains an N-best hypothesis (N-best hypothesis) about the most likely spoken content. Next, a word map packet is formed and sent to a server device through the network. Finally, the server device receives the word map packet, decodes it, and performs language processing to obtain a recognized word string.

Once the human speech is captured at a client device, it is subjected to acoustic analysis. Voice recognition involves extracting features of a captured speech signal and searching in an acoustic model for one or more possibilities between the extracted features of the captured speech and known previously recorded speech features stored in the acoustic model. match. In a preferred embodiment, a phonetic graph may be used to represent the speech. The sound model can be a personalized SD model which is personally trained by the user, for example the owner of a mobile phone or PDA. Thereby, the word graph is packaged and sent to a central server through a network. The server can then perform linguistic processing on the phonetic graph using the selected language model and generate a recognized word string.

Referring now to FIG. 1, a schematic DSR system is shown. Client 1, Client 2, and C1-C4 are examples of different client devices employing the new DSR method according to one embodiment of the present invention. Client 1 is a personal digital assistant 110 . Client 2 is a mobile phone 120, and C1-C4 are computer terminals as nodes in an illustrative LAN 138. In each case, the client device is configured to capture human speech. For example, human voice 112 for PDA 110 and for mobile phone 120. The captured signal is then acoustically processed and an obtained data packet containing a series of phonetic word maps is generated at the client device.

The data packet is then sent over the network 100, for example over the Internet. For the voice sent at the PDA 110, the processing server 150 receives the data packet (not shown), and then performs sound processing on the data contained in the data packet to generate the recognized word string 152. Similarly, host 180 and host 160 receive data packets from mobile phone 120 (client 2) and C1130 and generate identified strings 182 and 162, respectively. Client 1 can train the PDA 110 by initially reading a word string and training the PDA. Similarly, customer 2 can train mobile phone 120 by reading a word string and providing the mobile phone with the text of what he said. Once a client device has been trained, it can develop a personalized voice model that can be used by the client device as a basis for retrieving and comparing spoken content.

Referring now to FIG. 2, a schematic block diagram of a DSR system is shown. The diagram shows how a human speech signal 200 spoken at a client device 220 is converted into a recognized word string 260 at a server device 252 . The signal 200 is captured at a client 220 node of a client-server based system, and sound processing and recognition 230 is performed at the client device 220 . As shown in FIG. 2 , the client machine 220 may be a computer terminal 210 . However, the client machine 220 could also be a mobile phone, PDA or the like. In fact, the client device is any device having the functions of receiving a person's speech signal 200, performing sound processing on it and recognizing its phonetic composition, and preparing the obtained phonetic data for transmission over a network 242 such as a communication network 240. equipment. ,

Still referring to FIG. 2, the processing server 254 handles the language processing 250 stage of the new DSR system. The processing server 254 may be a server computer system 252 capable of receiving phonetic data and performing linguistic analysis on it to obtain word strings 260 . Once the server 252 has completed language processing 250 , the server 252 generates a recognized word string 260 which can then be sent back to the client 220 over the network 242 .

Referring now to FIG. 3, there is shown a schematic diagram of a new DSR system designed according to the method of one embodiment of the present invention. The figure shows a schematic sequence of operations to which the human speech signal 300 is subjected when the new DSR method is employed.

At functional block 310 , a human voice signal 300 is received at the client device 340 . For example, a person can speak into a microphone, which will capture the person's voice signal 300 . In an exemplary embodiment, a human may be limited to word commands that can be used as control commands, ie, an automatic speech recognition (ASR) system. In another embodiment, the system may include large vocabulary continuous speech recognition (LVCSR). The method according to this embodiment employs ASR and LVCSR.

At functional block 310, the captured human speech signal 300 is converted to a digital signal by an analog-to-digital converter. In functional block 320, the features of the obtained digitized signal are extracted, for example, by a feature parameter extraction module 820 (see FIG. 8 ). Functional block 320 can be further subdivided into: end point detection as shown in functional block 322 , pre-emphasizing filtration as shown in functional block 324 , and feature computation as shown in functional block 326 . During the end point detection process, the start and end of a speech feature are detected. In other words, make a judgment about when one feature ends and another begins. During pre-emphasis filtering, the speech signal is filtered to amplify important features of the speech signal. Finally, during the feature computation, features of the speech signal are computed to form a series of possible candidates.

Accordingly, the captured human speech signal is acoustically processed at functional block 342 after the speech features have been extracted. The sound processing is to provide the matching of the speech features identified at block 320 with known phonetic units. Thus, sound processing involves taking a human speech signal and using a sound model to regenerate a series of sounds that most closely represent the input speech. The sound model may be organized by sub-word units such as phonetic level, demisyllable or syllable units. However, sound models using other phonetic units can also be applied.

One way to perform sound processing is by utilizing Hidden Markov Models (HMMs). The HMM known in the art is a stochastic finite state automatic control composed of a Markov chain of sound states. These states model the temporal structure of speech, ie, how this state changes over time. The probability functions for each of these states, the observations of simulated emissions and sound vectors, are represented by HMMs.

Once an HMM is used to represent the speech features, a search space is determined and a search of previously formed HMMs can be performed within an acoustic model. The HMM may be formed during a client device's training phase, which may occur when a person uses the client device 340 for the first time. For example, when a person buys a mobile phone, the phone may have a key that, when pressed, puts the phone into a training model. During the modeling process, the person may be asked to speak words, phonemes, or other phonetic units that appear on the screen. The mobile phone can then capture the sound produced by the user and run it through the functional blocks 322-326 of FIG. 3 to extract features related to the sound and form an HMM. During this training phase, since the client device 340 knows exactly what word the voice represents, it can store two pieces of information (the spoken word and its extracted features) and create a User personalized sound models.

By creating a personalized sound phonetic model, the mobile phone can utilize an SD sound model with a WER 2-3 times better than the SI sound model.

In functional block 334, an optimization process is configured. Any source of knowledge may be used to make judgments about the spoken word. For example, phonetic phonetic models of individual phonemes trained by users of client devices may be used alone or in combination with other sources of knowledge, such as pronunciation dictionaries. However, if the user is not the person who actually uses the client device, the person who actually uses the device should train the device, since this is the person's speech characteristics, which will result in a more accurate recognition process.

At functional block 336, an N-best hypothesis is determined after completing the search for the acoustic model. However, instead of the N-best hypotheses, a single-best hypothesis strategy can be utilized. In function block 338, a phonetic word graph (Pword graph) is generated. The main idea of the pword map is to propose phonetic alternatives in regions of the speech signal where the uncertainty about the actually spoken phoneme is high. A desirable advantage is the separation of the voice recognition processing from the application of complex language models. This language model can then be applied by methods according to embodiments of the present invention in post-processing performed at the server computer. The number of word substitution options is a design parameter that can vary depending on the level of uncertainty or precision desired by the user.

Once a Pword graph has been generated at functional block 338, the Pword graph can be packaged and sent to the server device. Any delivery medium and any packaging scheme can be used to send the Pword graph to the server. For example, an Internet Protocol datagram can be generated at functional block 354 as shown by packaging the Pword graph into a datagram. The datagram may then be sent over network 350 as indicated by functional block 352 . In a preferred embodiment of the present invention, network 350 may be the Internet, but any other type of network such as a local area network may be used.

In functional block 356, the datagram containing the Pword map is received by a server, and the Pword map is removed from the datagram. At functional block 382, linguistic processing may be performed on the Pword graph. The language processing involves organizing the series of sounds in a Pword graph and converting it into actual words. The received Pword graph is analyzed node by node. For each node, the dictionary and grammar rules are checked against the specific language model available and selected by the user. In one embodiment, the client device may have a language selection button that allows the user to speak in English or Chinese or any other language that the system may support. At functional block 390, an actual Pword map is formed from the Pword map sent by the client device (see FIG. 5). Finally, at functional block 386, a search algorithm is used to determine the recognized word string through a dictionary and a grammar dictionary.

Referring now to FIG. 4, a block diagram of a client device is shown. The client device may be a number of portable devices such as a mobile phone, PDA, portable computer, or any other device that may be used by a user communicating with another device located in a different geographic location.

Once a person decides to communicate with a remote server, that person will have the option of speaking into the receiver module of the client device, such as a microphone. However, the client device performs a series of operations on the captured human voice signal 400 . These operations are shown in functional blocks 420 and 450 in FIG. 4 . The operations performed on the human speech signal 400 can generally be divided into two types. In a first series of functional processes represented at functional blocks 422, 424, 426 and 428, a human speech signal is subjected to a process in which the human speech signal is converted to a digital signal according to methods known in the art. Then, at functional block 412, the digitized signal is presented to a feature extraction module, which extracts the features present in the person's speech signal. These features may represent the concentrated energy in the speech signal being measured periodically, and may be represented as a sum of sound vectors, eg shown in function block 428 as x1, x2, . . . , xT. However, other features of the sound signal known in the art may also be extracted.

At functional block 450, the sound vectors x1, x2, . To accomplish this task, the sound processor may refer to a sound model containing sound vectors for various speech sounds previously uttered by a person using the client device. The model can be easily trained by the person who will initially use the client device. For example, a person may program or train the client device when purchasing it for the first time. The device could have a "train me" switch that, when activated, would flash text on its screen prompting the user to pronounce the text. The device can, for example, flash words, phonemes, syllables, demisyllables or other units of any word according to certain design parameters. The selection of the phonetic symbol unit has no influence on the method based on the embodiment of the present invention.

So, for example, the device flashes the word "apple" and the user speaks "apple". The device will capture the voice signal produced by the user's spoken voice, for example through a microphone. Those of ordinary skill in the art know that this signal is an analog signal, which may resemble speech signal 400 when viewed on an oscilloscope. After capturing the signal, the sound processor may use the functions at functional blocks 412, 422, 424, 426, and 428 to extract features of the signal produced by the user speaking the word "apple," resulting in the generation of a set of sounds vector. This expression is then stored in a database along with the representation of the word "apple". This processing can be performed continuously word by word. The more words that are displayed to the device, the more complete the sound model for that user or device owner. Once the model is complete, the device is ready for voice recognition which occurs at functional block 450 .

The sound processor is now responsible for the task of recognizing spoken speech. It does this by searching a database containing trained sound models. At functional block 446, a search is performed to find one or more matches for the speech. The judgment of the spoken word can be realized through an optimization process. Several search processing methods have been developed and are known in the art. For example, a directed search strategy with modification options can be used. Alternatively, a dictionary of trees or a one-shot algorithm can be applied. The choice of a particular search strategy does not affect or alter the method according to embodiments of the present invention.

At functional block 442, a database containing sound models is collected. The training phase of the speech recognition system of this embodiment occurs at this functional block. At functional block 444 , a language model is considered to link to the search strategy used at functional block 446 . However, adding a language model on the client side can be a design choice. It is not necessary to include a language model to implement the method according to this embodiment.

The search results are generated at functional block 448 . Here, an N-best hypothesis is generated. Although, a single best hypothesis could also be used in a preferred embodiment, the N-best hypothesis yields higher accuracy because it provides not just a single guess at what was said, but multiple guesses. At functional block 452, a word map may be generated from this information. The main idea of a word graph is word replacement. Word graphs must be proven efficient in situations where high precision is required. In fact, a word graph shown in Figure 5 shows words with similar sounds, or features, or sound vectors. This similarity can be confusing. For example, the words "duo" and "dao" and "yao" in Chinese look almost the same on a spectrum analyzer. Similarly, referring to FIG. 5, the words "dai", "nai" and "mai" are similar except for one letter or phoneme. These general similarities in most languages can be further analyzed by using a grammar lexicon given in the language model discussed below with reference to Figure 7A.

Referring to Figure 4, once a word map representing alternative options for a spoken word is generated, the device can send this information as a binary file to a remote server. The word map can be represented in a datagram as shown in Figure 6B. However, any other form of packing of the data may be employed.

Referring now to FIG. 5, an example of a word graph having two levels of substitution option capacity is shown. In this example, the Chinese of the actually read word is: "wo yao mai zhong ke jian", which means "I want to buy Zhongkejian" (Zhongkejian is the name of a stock on the Chinese stock market ). This word map is the output of the sound processor as shown in FIG. 4 at functional block 452 . The sound processor compares the sound vectors captured by the device with the sound model and provides the sound processor with three alternatives for each word. Words 512, 511 and 510 represent "yao" and its replacement options. Words 514, 515, and 516 represent "mai" and its replacement options. Accordingly, the word graph shown in FIG. The single best sentence represented by the graph. Applying a language model to the word graph can be done at a server node, since the language processing is a rather complex process and has nothing to do with the voice recognition process. Therefore, the method of this embodiment takes advantage of the SD sound model by generating word graphs with two levels of word replacement options. The word graph will be transmitted to a server which then completes the recognition process and determines the single best sentence.

Referring now to FIG. 6A, there is shown a transmission process according to one embodiment of the present invention. At functional block 602, a phonetic word map is generated by the client device. At functional block 604, the word graph is converted to a binary file and packaged for sending over the network. For example, at functional block 604, a TCP/IP datagram is used as a vehicle for transmission. However, any other method of packaging the wordmap for transmission may be used, and this particular choice has no effect on methods according to embodiments of the present invention. At functional block 606, the datagram is sent to the server, and at functional block 608, the datagram is received at the server.

Referring now to FIG. 6B, an exemplary Internet Protocol datagram is shown. In the header 612 portion of the datagram 600, the logical address of the client device and the logical address of the server device and any other control information known in the art are contained. The data area 610 may include a binary representation of the phonetic graph generated by the client device.

Referring now to Figure 7a, a schematic block diagram of a server node 700 is shown. At functional block 710, a TCP/IP datagram is received by server 700 as shown in FIG. 6B. At functional block 712, the word graph is decoded from its binary form and an actual word graph representation of what was spoken at the corresponding client node (not shown) is formed. At this point, the server has the equivalent of N hypothetical representations of the speech. As known in the art, the server can use a language model at block 720 and a dictionary as shown at block 718 to search and determine the most likely speech.

During this process, for each phonetic word graph node (see Figures 7b and 7c), the server 700 looks up the selected phonetic word by checking the dictionary and the grammar dictionary. However, the present invention is not limited to this dictionary and grammar model. Any other language models such as cache-based language models, trigger-based language models, and long-range trigram language models (lexicographically-encoded context-free grammars) can also be used. Regardless of the particular language model used, the result at function block 720 is a recognized string that can be stored, or used as a command to the server 700 .

Referring now to Figure 7b, a schematic graph of real phonetic characters is shown. The phonetic character graph represents the spoken word string "I want to buy Zhong Ke Jian (wo yao mai zhong ke jian)". From this word map, a corresponding word map as shown in Fig. 7c can be generated. In this process, the server searches for words that sound similar to "yao" for each phonetic word map (for example, "yao"). As another example, for the phonetic character "zhong", the actual character might be "zhong" or "zhong" or "kind" (which is a character that sounds like "zhong"). The pronunciation of these characters is not similar in English, but they are similar in Chinese. The method according to the present embodiment is not limited to English or Chinese. Any language capable of constructing a language model can be used.

Referring again to FIG. 7b, once the word replacement options are obtained for each phonetic symbol node, the server can generate a plurality of actual word nodes according to the found words. This topological relationship is then replicated in the phonetic graph to obtain the extended phonetic graph as shown in Figure 7c, which shows the phonetic graph obtained from the graph shown in Figure 7b.

Referring now to Figure 7c, the extended phonetic graph is shown. In this case, the server takes into account the different possibilities of the read sequence on the basis of a language model. For example, after the word "I" (the subject), the language model might detect a verb such as "want" instead of a noun. Accordingly, a revision strategy can be employed wherein nouns following "I" are not considered further, eg the noun "medicine" may not follow the word "I" as subject. In this way, the resulting search space can be greatly reduced. Similarly, a dictionary can be used to eliminate other similar-sounding words. Here, a language model based on a bigram language model or a trigram language model can be used. The choice of the bigram or trigram language model does not affect the method according to the embodiments.

Referring now to FIG. 8, there is shown a schematic block diagram of a speech recognition system including client devices, server devices and a communication network. Voice input 800 may be a user's name, such as John. The speech input 800 may be captured by a microphone connected to a client device belonging to John, such as John's mobile phone or PDA. John can use his device's trained model to train his device to recognize his speech. The acoustic model 824 located at the client device 810 is used in the training model. When John is prompted to speak different words, phrases or sentences, the language model collects data corresponding to each language. When John is ready to communicate with a remote server 850 via a communication network 840, he can switch off the training mode and start talking as if he were having a normal conversation with another person. The client device 810 will capture John's speech and pass it through the feature extraction module 822 to perform a series of front-end processing on the simulated human speech signal 800 as known to those of ordinary skill in the art. Whereas in prior art models, according to one embodiment of the invention, the extracted features are sent to the server for language processing, an additional function occurs at the server device, i.e., resulting in the generation of a phonetic graph sound processing. Thus, an embodiment takes advantage of the SD sound model, since John is able to train the device individually, thus resulting in an SD personalized sound model. Prior art cannot take advantage of the lower WER associated with SD models, where features collected at the client are sent directly to the server, and the server performs sound recognition and analysis. Using the SD model with this prior art is impractical because the server serves many users without knowing their identities. Therefore, this prior art is limited to the SI model, which easily leads to a higher error rate.

Once the voice processor receives the extracted features, it searches for a speaker-dependent voice model trained by John's voice. The obtained matches are known phonetic units 830 that can be sent to the server 850 in a datagram. The datagram is received at server 850 and provided to a language processor 855 in conjunction with a pronunciation dictionary 857 and language model 859 to determine the recognized word string.

Claims (22)

1. a distributed speech recognition method comprises:

Voice signal a client node recipient;

Discern the feature of described people's voice signal;

Identification is corresponding to the known phonetic symbol unit of the described feature that is identified;

Formation comprises the packet of at least one described known phonetic symbol unit; And

Described packet is sent to a server node.

2. distributed speech recognition method according to claim 1, wherein said client node is selected from mobile phone, personal digital assistant and portable computer system.

3. distributed speech recognition method according to claim 1, the feature of wherein discerning described people's voice signal comprises:

Voice signal to described people is carried out end point detection;

Filtering is emphasized in voice signal execution to described people in advance; And

Quantize described people's voice signal.

4. distributed speech recognition method according to claim 1 is wherein discerned corresponding to the known phonetic symbol unit pack of the described feature that is identified and is drawn together sound model of search, comprising:

The sub-word cell of simulating by the hidden Markov model of the Markov chain that comprises a sound status.

5. distributed speech recognition method according to claim 1 is wherein discerned corresponding to the known phonetic symbol unit pack of the described feature that is identified and is drawn together sound model relevant with the speaker of use.

6. distributed speech recognition method according to claim 1, wherein said known phonetic symbol unit forms a phonetic symbol word figure.

7. distributed speech recognition method according to claim 1, wherein said packet comprise the binary representation of a source address, destination address and described known phonetic symbol unit.

8. distributed speech recognition method according to claim 1, wherein said packet is sent by the internet.

9. a distributed speech recognition system comprises:

Client node, comprising:

The characteristic extracting module of the feature of identification people's voice signal,

Be connected to the acoustic processing module of described characteristic extracting module, this acoustic processing module is from the known phonetic symbol of the described feature identification that is identified unit, and

Be connected to the transmitter module of described acoustic processing module, this transmitter module forms and comprises the packet of at least one described phonetic symbol unit and described packet is sent to a server; And

Server, comprising:

Receiver module is used for receiving described packet and removes described at least one described known phonetic symbol unit from described packet; And

Language processing module is used to discern and described at least one relevant word in described known phonetic symbol unit.

10. distributed speech recognition system according to claim 9, wherein said client node is selected from mobile phone, personal digital assistant and portable computer system.

11. distributed speech recognition system according to claim 9, wherein said characteristic extracting module also are configured to described people's voice signal is carried out end point detection, emphasized filtering and quantification in advance.

12. distributed speech recognition system according to claim 9, wherein said acoustic processing module comprises a sound model relevant with the speaker.

13. distributed speech recognition system according to claim 9, wherein said acoustic processing module forms a phonetic symbol word figure according to described known phonetic symbol unit.

14. distributed speech recognition system according to claim 9, wherein said transmitter module forms the binary representation of described word figure, and before sending described word figure, described binary representation and source address and destination address are together placed a datagram.

15. a client devices, comprising:

Receiver module is used for recipient's voice signal;

Characteristic extracting module, it is connected to described receiver module, is used to discern the feature of described people's voice signal;

The acoustic processing module, it is connected to described characteristic extracting module, is used for discerning known phonetic symbol unit from the described feature that is identified, and forms the packet that comprises at least one described phonetic symbol unit; And

Transmitter module, it is connected to described acoustic processing module, is used for described packet is sent to a server node.

16. client devices according to claim 15, wherein said characteristic extracting module also are configured to described people's voice signal is carried out end point detection, emphasized filtering and quantification in advance.

17. client devices according to claim 15, wherein said acoustic processing module comprises a sound model relevant with the speaker.

18. client devices according to claim 15, wherein said acoustic processing module forms a word figure from the described known phonetic symbol unit that is identified.

19. client devices according to claim 18, wherein said word figure are phonetic symbol word figure.

20. a server, comprising:

Receiver module is used for comprising the packet of at least one known phonetic symbol unit from a client node reception, and removes described at least one known phonetic symbol unit from described packet; And

Language processing module, it is connected to described receiver module, is used for determining and described at least one relevant word in known phonetic symbol unit.

21. server according to claim 20 wherein receives described packet by the internet.

22. server according to claim 20, wherein said packet are datagrams, wherein comprise:

Header portion with address and described known phonetic symbol unit of the address of described client node, described server.

Images