DE102015213722A1 - A method of operating a speech recognition system in a vehicle and speech recognition system - Google Patents

Abstract

The invention relates to a method for operating a speech recognition system (10) in a vehicle (6) and to a speech recognition system (10). In the method according to the invention at least one speech input of a user is detected and data on the context of the speech input is detected. A recognized text of the speech input is generated and a semantic analysis of the recognized text is generated. Finally, a function is performed based on the semantic analysis.

Description

The invention relates to a method for operating a speech recognition system in a vehicle and to a speech recognition system.

Speech recognition systems find a wide range of applications for operating systems. This applies equally to the operation of devices and the completion of telephone tasks using voice computers, such as telephone banking. Also, the operation of facilities of a vehicle via voice inputs has become established as an essential feature of current user interfaces in the vehicle. For example, to input complex facts corresponding information is queried sequentially by a user through a speech dialogue system.

It is of crucial importance that the recognition and interpretation of a user's speech input is as error-free and fast as possible. Although various methods have already been proposed for this, the large computing capacity required for implementation often makes deployment difficult, for example in the field of mobile devices or in vehicles. One approach to solving this problem is to outsource the computational tasks to a server accessible via the Internet. However, this presupposes that the voice recognition system can exchange data with the server at any time, which is not always guaranteed, especially in areas with incomplete coverage by mobile radio networks.

In the in the DE 103 29 546 A1 The mechanism for automatic speech recognition described the words of a base lexicon are divided into application and user-specific phrases. The vocabulary restricted in this way reduces the number of words that can potentially be recognized.

The EP 2 273 491 A1 describes a user interface for controlling a natural language navigation system. Here, context information can be taken into account, for example to select the most sensible alternative from a large number of addresses. The interpretation of the linguistic input can be improved with the help of contextual information and learned knowledge by selecting a context-specific agent based on a first, less precise interpretation and then using it for a better interpretation.

The US 2007/0073540 A1 describes a speech recognition in which an input can be corrected. Upon inputting two utterances, it is recognized whether the second utterance matches the whole or partial first utterance, and a part recognized as being similar is corrected.

It is therefore the object of the present invention to provide a method for operating a speech recognition system and a speech recognition system with which speech inputs can be quickly and reliably recognized and interpreted.

According to the invention, this object is achieved by a method having the features of claim 1 and a speech recognition system having the features of claim 15. Advantageous embodiments and further developments emerge from the dependent claims.

In the method according to the invention, at least one speech input of a user is detected and data relating to the context of the speech input is acquired. A recognized text of the speech input is generated and a semantic analysis of the recognized text is generated. Subsequently, a function is performed based on the semantic analysis.

In particular, the voice recognition system can be operated in a vehicle.

The speech input is detected by methods known per se, for example by means of a microphone and an associated device for storing and processing the detected signal. For example, the user may signal the system to begin voice input by pressing a push-to-talk (PTT) key. The recorded audio data of the voice input can be stored in a volatile or non-volatile manner and thus be available at a later time.

After detecting the speech input, a text recognition is performed. In this case, a recognized text is determined on the basis of a speech model from the detected audio signal of the speech input. This is generated in particular machine-readable and can be used for further analysis. The language model may include, for example, a dictionary. It can also map the frequency of words individually or in combination. The speech recognition system may include multiple language models, for example, mapping a different sized vocabulary.

In a next step, a semantic analysis of the recognized text is generated, using, for example, methods from the field of "Natural Language Recognition". This concerns the interpretation of the speech input and the determination of the intention of the user. By means of semantic analysis of the recognized text can then be performed a function.

To improve this process, the above-mentioned sequential steps can be optimized individually or in combination with each other. Improved conversion of the user's verbal input into machine-readable text, for example, will result in a more reliable interpretation of the speech input. Further, the semantic analysis of the recognized text can be improved to better determine the intention of the user and then reliably perform a function according to the speech input.

The collected data contributes to the context for improving various steps by providing additional information.

According to the invention, the "context of the speech input" represents the entirety of the information concerning the driving situation in which the speech input takes place. Specifically, the data on the context of the speech input may include information concerning the context at the time of voice input; furthermore, contextual data may be included in the past. In addition to the temporal reference, a spatial reference can also create the context, for example with information about the closer surroundings of the vehicle.

In a development of the method according to the invention, the data relating to the context of the speech input relate to properties of the vehicle, the user and / or the environment of the vehicle. As a result, additional information can advantageously be provided and the individual steps of speech recognition can be improved.

The captured data on the context may be provided, for example, by the sensors of the vehicle. Furthermore, they can be predetermined by the manufacturer, for example, and can be provided by a device of the vehicle, for example a storage and computing unit of the vehicle. They can also be detected by a computer which is connected, for example, by an Internet connection with other arithmetic or can exchange data through a local area network, for example with a mobile user device of the user. Further, inputs may be captured by the user, such as by input devices of the vehicle.

For example, the data about the vehicle may relate to the speed and position of the vehicle. Furthermore, data may be included on design-related properties of the vehicle, such as the vehicle type, the dimensions, the required fuel type or the maintenance intervals. Information about the use of the vehicle, such as the distance traveled within a time interval, may also be included.

The user of the speech detection system may in particular be the driver of the vehicle. The data about the user may include, for example, an identification of the user or previous interactions of the user. This allows user-specific adjustments. In addition to the user, data on further vehicle occupants can also be recorded in a vehicle, for example their identity and number as well as their interactions with the voice recognition system. The data on the context may be captured by various devices, but may also be provided by the speech recognition system itself, for example as part of a learning function in which past interactions with the user are stored by the speech recognition system.

The data about the environment of the vehicle may, for example, include information about other vehicles in the vicinity of the vehicle, such as their position and speed. Also, the traffic situation can be detected, for example, a traffic jam situation can be detected. Further, information about the traveled road may be detected, such as the road type, the maximum speed allowed, the condition of the road surface or the density of the traffic.

In a further development, the data on the context of the speech input comprise further data generated by or stored on devices. As a result, the context can advantageously be enriched by additional information. The devices by which data can be generated or stored are, in particular, devices of the vehicle or mobile user devices. For example, it can be detected which devices are active at the moment of speech input and to which the speech input relates. Data may be collected, for example, from an electronic calendar, a navigation function, a media player or a telephone device.

In particular, operating devices can provide data on the context, for example, it can be detected whether the user has activated the operation of a particular device and whether he is currently performing an operation. If, for example, a navigation device is activated, then the system can conclude that voice input takes place in connection with this device.

In one embodiment of the invention, after the generation of the semantic analysis, a recognition quality of the recognized text is further determined and a language model selected from a plurality of language models depending on the semantic analysis and the data on the context of the speech input. Based on the selected language model, another recognized text of the speech input is generated. Subsequently, a further semantic analysis of the further recognized text is generated and said steps are repeated until the recognition quality exceeds a predetermined value. As a result, the speech recognition can advantageously be optimized.

According to this embodiment, the first speech recognition, by which a recognized text is generated, and a first semantic analysis of this recognized text, are followed by a further step, in which the recognition quality of the recognized text is checked. In particular, the first text recognition takes place on the basis of a general language model, which may include, for example, a basic vocabulary and basic sentence structures. It is therefore expected that the first speech recognition generally can not be performed optimally. A recognition quality is determined, which is a measure of how well the text was recognized. For this purpose, the semantic analysis based on this text is used in combination with the acquired data about the context of the speech input. For example, it can be determined whether the recognized text contradicts the context mutatis mutandis. Furthermore, according to the invention, based on the data on the context and the semantic analysis, a language model is determined which is adapted to the determined context. For example, a language model can be selected that is adapted to the respective situation. For example, it can be recognized from the first semantic analysis that the user is about to enter a shopping list. In order to recognize the text that follows, a language model can then be loaded that includes, for example, a food-adapted vocabulary. Furthermore, for example, in the operation of a navigation system, a language model can be selected in which place and street names are known. Furthermore, for example, the driving situation can be used to determine a suitable language model. For example, if it is registered that the vehicle is in a traffic jam, a voice model may be selected to detect messages announcing an imminent delay.

On the basis of the thus determined further language model, speech recognition of the speech input is again carried out, followed by a new semantic analysis. In an iterative process, in turn, the recognition quality is determined and a new language model is chosen until a text is generated that meets the predetermined recognition quality requirements.

In a further development, at least one user intervention class is determined on the basis of the semantic analysis. The data on the context of the speech input is detected depending on the user's intention class, and at least one more semantic analysis is generated depending on the data on the context of the speech input. This advantageously achieves an improved semantic analysis.

In particular, this can be done iteratively by gradually specifying and improving the semantic analysis step-by-step by incorporating the context.

In determining the user Intent class, the input intent of the user is roughly estimated. It is now possible to record such data about the context of the speech input that are relevant for inputs of the determined user's intention class. For example, in the case of an input relating to a navigation device, it can be detected which appointments are recorded in a calendar device, which energy reserves of the vehicle are available and what information about the traffic situation in the surroundings of the vehicle is available. On the basis of this collected, specific context information, a further semantic analysis can be generated.

In particular, context information can also be recorded that is stored, for example, from past interactions of the user with the speech recognition system. For example, with multiple voice inputs, the information that the user has entered or retrieved may be stored. This allows the resolution of back references, especially deictic expressions. In this way, for example, the term "there" can be interpreted by deducing from past interaction steps which location is meant. For example, this can be determined from past requests in the navigation system. Furthermore, user behavior can be learned in different contexts. For example, when an imminent deadline is recorded in the calendar, a small distance may be weighted higher than a lower price when choosing a gas station to approach. Further, for example, depending on the weather conditions, certain restaurants may be rated higher than others.

Furthermore, ambiguities can be resolved, for example when calling a Phone book entry, if several persons with the same name are stored. The ambiguity can be resolved in this case, for example, based on the geographical position, such as at work or at the place of residence.

In this method, as complete a semantic analysis of the speech input is made, so that in subsequent evaluation steps, for example by a dialog function, all essential semantic information is already recorded and the following analysis effort is reduced.

In a further development, semantic analyzes of speech input are stored. This advantageously allows the inclusion of past speech interactions in future semantic analysis and speech recognition. In particular, a learning function is made possible.

In a further embodiment, the data on the context of the speech input further comprises semantic analyzes of previous speech inputs. This advantageously includes earlier speech inputs in the semantic analysis. For example, in short successive speech inputs, the context of the spoken text can be recognized. This further facilitates the above resolution of deictic expressions and back references.

In a further development of the method according to the invention, a search function is carried out on the basis of the semantic analysis. The user can thus advantageously search for information by voice input.

In an education, search criteria are determined based on the semantic analysis and data on the context of the speech input. On the basis of the specific search criteria, a subset of data of a database is determined and the data of the subset is output. This advantageously implements a search function.

It searches a given database provided by, for example, in-vehicle facilities, user facilities, or facilities connected through a computer network. The search is intended to determine a subset of data from this database which are relevant for answering the speech input. The speech input is a search query in this case. To select the subset of data from the database, search criteria are determined. For example, keywords can be determined by which a selection of data from the database is made. The determination of the search criteria takes place in particular depending on the data on the context. As a result, for example, a search query can be specified.

In particular, such an incomplete, underspecified user's search query is completed. For example, the system may add information that the user implies as they type. For example, it can be recognized that in the search for a gas station only those places can be meant that can reach the vehicle with the current energy reserves.

In a further development, priorities are generated for the data of the subset and the output of the data of the subset is based on their assigned priorities. As a result, it can be advantageously determined and output for the user, which is about the most relevant data for the search query. In particular, the more relevant the data, the higher priorities are assigned.

In an education, the search criteria are further determined by previous calls of the search function. This advantageously implements a learning function of the search.

The learning can be done for example by means of neural networks or Bayesian networks. This can be used to extract correlations between certain information about the context and behavior of the user. The system can thus detect, for example, whether the user avoids a road type under specific weather conditions or approaches restaurants with certain characteristics. This can be used to prioritize results that are more relevant to the user when assigning priorities to the database data than less relevant data.

When learning from previous interactions of the user with the search function, different strategies are applied individually or in combination. The learning function can adjust the search function for the future based on positive, neutral or negative feedback.

For example, can be considered as positive feedback, if in a navigation system, the user selects a suggested address and, for example, the navigation to this location starts. As a further positive input can be considered if the user z. B. actually approaches the proposed location. If the user does not respond to the results of the extended search, he formulates the search query again with more specific keywords or if he does not start navigation to the proposed location, this indicates the failure of the search performed. Furthermore, a negative feedback can be recorded when the user discards a search result, such as by entering a rejection. By learning from the past user behavior, the extension of the search by the context can be improved.

In a further embodiment, at least two voice inputs of the user are recorded. For both speech inputs, respectively recognized texts are generated which consist of ordered text fragments, whereby the text recognized for the first speech input is the original text and the text recognized for the second speech input is the correction text. a corrected text is generated based on the original text and the correction text. As a result, an input can advantageously be improved.

This concerns the correction of a wrong speech input, which was correctly recognized by the system, but where the user made an incorrect input. The basis for the correction of the entered text are the machine-readable texts of the two speech inputs generated by the speech recognition. A comparison of the original text and the correction text determines which part of the original text should be improved. By means of the correction text, the original text is changed so that a corrected text is generated.

In a further development, the correction text comprises at least one text fragment which signals the beginning and / or the end of the correction text. This advantageously makes it possible to recognize the correction text as such. For example, the user can mark the beginning of the correction text by a defined expression when the voice is input. For example, he can say that a correction text now follows. For example, he can also operate a key or a control element. Furthermore, the end of the correction text can be marked analogously.

In a further development, the position of at least one text fragment in the original text, which is encompassed by both the original text and the correction text, is determined on the basis of a comparison of the original text with the correction text. The corrected text is generated based on the specific position. Thus it can advantageously be determined which part of the original text should be corrected.

This happens because individual text fragments, for example individual terms, occur both in the original text and in the correction text. Thus, the user repeats part of the original input to identify the position of the correction. For example, the beginning or the end of the text to be corrected can be marked in this way. Furthermore, both the beginning and the end can be defined.

In one embodiment, at least one legacy text fragment of the original text is further determined and at least the new text fragment of the correction text is determined, wherein the legacy text fragment is replaced by the new text fragment when the corrected text is generated. This can advantageously be done a defined correction of the original text.

The correction is made by replacing at least one text fragment of the original text with at least one text fragment of the correction text. The user thus improves the original input by speaking the text to be corrected and overwriting the original text at a defined location.

In a further embodiment, the at least one text fragment, which is encompassed by both the original text and the correction text, comprises the first and / or the last text fragment of the correction text. As a result, advantageously the text fragment to be exchanged can be specified particularly easily.

The user can thus make a correction by repeating a part of the original text that is to be retained. In this case, the part of the correction text that follows the common text fragment may be used to override a part of the original text. Conversely, at the end of the correction text, the user may repeat part of the original text to mark the end of the part to be corrected.

In a further development, a first and a second text fragment are encompassed by both the original text and the correction text. The first text fragment is arranged in the original text as well as in the correction text in front of the second text fragment. Based on the first and the second text fragment, the legacy text fragment and the new text fragment are determined. The text to be exchanged is thereby advantageously determined precisely.

In this way, for example, frame words are specified, which frame the part of the original text to be exchanged. If the user wishes to exchange a certain part of the original text with the correction text, he first repeats a part of the text lying directly in front of the old text fragment to be exchanged, then enters the new text fragment to be used here and ends the input with another text fragment that should be left in the original wording. So it can be a frame through before and after the old Text fragment words are defined to perform a particular replacement.

The speech recognition system according to the invention in a vehicle comprises a detection unit, by means of which at least one speech input by a user can be detected, and a context detection unit, by means of which data relating to the context of the speech input can be detected. It also comprises a speech recognition unit, by means of which a recognized text of the speech input can be generated, as well as a semantic recognizer, by means of which a semantic analysis of the recognized text can be generated. It also comprises an execution unit to which the semantic analysis can be transferred and by means of which a function can be executed on the basis of the semantic analysis.

The speech recognition system according to the invention is in particular designed to implement the method according to the invention described above. It thus has the same advantages as the method according to the invention.

The speech recognition system according to the invention is arranged in a vehicle. As a result, the voice recognition system can advantageously be used to control facilities of the vehicle and context information of the vehicle is available.

The invention will now be explained by means of embodiments with reference to the drawings.

1 shows the general structure of a speech recognition system,

2 shows a first embodiment of the speech recognition system according to the invention,

3 shows a further embodiment of the speech recognition system according to the invention,

4 shows an embodiment of the speech recognition system according to the invention in a vehicle,

5 shows an embodiment of the method according to the invention, in which a search request is recorded and processed,

6 shows the learning process in the context-sensitive search and

7A to 7D show embodiments of the inventive method, wherein a voice input is corrected.

In reference to 1 the general structure of a speech recognition system is shown.

The speech recognition system 10 comprises a detection unit 1 which in turn is a microphone 1a includes. This detection unit is equipped with a voice recognition unit 2 coupled. Here can be based on the with the microphone 1a recorded speech input machine-readable text. This text is sent to the semantic expert 3 where a semantic analysis is generated and a signal to an execution unit 5 is transmitted to perform a function.

The user can signal the beginning of a voice input, for example by pressing a push-to-talk button (PTT). Alternatively or additionally, it can be provided that the system prompts the user for a voice input and the text spoken after the request is treated as a voice input.

In reference to 2 a first embodiment of the speech recognition system according to the invention and the method according to the invention will be explained.

As already in 1 represented, has the speech recognition system according to the invention 10 via the registration unit 1 with the microphone 1a , It also comprises the speech recognition unit 2 , the semantic expert 3 and the execution unit 5 , According to the invention, a context detection unit is also provided 4 intended. This is in the embodiment shown with a context processing unit 4a coupled, in turn, with the speech recognition unit 2 and the semantic expert 3 is coupled. In the embodiment shown, the speech recognition system is located 10 in a vehicle 6 ,

In the method according to the invention, a voice input of the user is first detected. In the embodiment shown, the user signals the beginning of the speech input by pressing a PTT and the detection unit 1 records the audio data spoken by the user. A caching is made so that the audio data is available for further use.

Further, by the context detection unit 4 Data about the context of the speech input is captured. The context data generated by the context capture unit 4 be detected affect the user, the vehicle 6 and the environment of the vehicle 6 , In particular, data from other facilities of the vehicle 6 detected. These include, among other things, the speed, acceleration and direction of movement of the vehicle 6 . its geographical position, the available fuel reserves, the type of vehicle 6 , the type of fuel needed, the dimensions of the vehicle 6 and for the vehicle 6 scheduled maintenance intervals. Among other things, the data about the user includes an identification of the user, so that different users can be recognized. Furthermore, the data of the calendar of the user is recorded, in particular the dates and geographical positions of planned appointments. Further, for example, the sensor of the vehicle 6 provide additional data, such as the number of vehicle occupants, data on the attention and fatigue of the user and other features. It can also be detected which other devices are activated and whether the user is about to operate a device. The context data also includes stored and / or learned data about the user. In this case, previous interactions of the user with the speech recognition system can be stored and learning methods can be used to learn the behavior of the user and his preferences in dependence on the further context data.

The data on the context about the environment of the user and the vehicle 6 include in particular data on the current traffic situation. These include the type and condition of the road being traveled, the position, speed and direction of other vehicles in the vicinity of the vehicle 6 , Traffic reports, such as traffic jams, the climatic conditions such as rain and wind, the lighting conditions and possibly obstacles on the traveled route. To capture this data can be facilities of the vehicle 6 contribute, such as a navigation system or the sensors of the vehicle 6 , or data may be collected from remote sources, such as radio, cellular, or Internet services. This context information may also relate to so-called Points of Interest (POI). These include petrol stations, restaurants, rest areas, sights and landmarks.

The context processing unit 4a has funds available from the context capture unit 4 provided data for the speech recognition system 10 prepare and provide via an interface.

At the in 2 In the illustrated embodiment, the context data is used to optimize speech recognition. In a first step, speech recognition is performed based on the speech input and a machine-readable text is recognized. On the basis of this recognized text, a first semantic analysis is performed. Speech recognition uses a common language model. This language model includes a defined vocabulary and statistical data about the individual words and their relationship to each other. This general language model generally does not allow optimal recognition of the speech input. The recognition quality is determined on the basis of the semantic analysis and the context. For example, contradictions between the context and the recognized text are evaluated or the grammatical structure of the recognized text is checked for plausibility. If the recognition quality does not satisfy the given conditions, a different language model is selected on the basis of the context and the first semantic analysis, and speech recognition is carried out on the basis of this language model and the speech input. The resulting recognized text serves as the basis for a further semantic analysis. In turn, the recognition quality for the further recognized text is determined and, if necessary, a new language model is determined.

In this method, a new, more specific language model is iteratively selected over and over again until the resulting speech recognition either satisfies the predetermined requirements for the recognition quality or an abort condition is met. In this way, the speech recognition is gradually improved. At the same time allows the selection of more specific language models z. B. a limitation of the vocabulary available for speech recognition. This reduces the computational effort while improving the recognition quality.

For example, the user may dictate to the system a shopping list. Using the general language model, it can be deduced from signal words, such as "shopping" or "shopping list", that the user intends to dictate a shopping list. On the basis of this context information, a corresponding language model of the speech recognition unit 2 which includes, for example, the names of various products. The spoken word can thus be better recognized than with the general language model. If sufficient recognition quality is not achieved, for example because the product names used originate from an area that is insufficiently covered by the language model, the new selection of the language model is repeated and a more suitable language model is loaded. The language models also include sentence structures which are typical for the respective context. In the mentioned example of a shopping list, the language model can also be selected so that further instructions, such as the creation of a calendar entry for shopping or a search for a shop, can be better recognized.

An example in which the context of the current driving situation affects speech recognition is a congestion situation. Detects the system that the vehicle is 6 is in a traffic jam, for example on the basis of traffic reports or because the speed for the road type is too low, then a language model can be loaded in which, for example, the text of an e-mail informing about a delayed arrival is better recognized. In this case, for example, words such as "congestion" or "delay" may be prioritized higher, and sentence structures such as "I'll be delayed X minutes" will be provided.

The context information thus contributes to limiting the high complexity of all possible driving and speech situations to the specific situation. In particular, the vocabulary and the available sentence structures for speech recognition can be reduced.

In reference to 3 a further embodiment of the speech recognition system according to the invention and the method according to the invention will be explained.

The speech recognition system shown here 10 includes the same elements as the one in 2 shown. However, the context processing unit is here with the semantic recognizer 3 coupled. In this example, the context is used to optimize the semantic recognition. It is assumed in the embodiment shown here that the speech recognition has already been completed and an improved semantic recognition is to be performed.

Based on a first semantic analysis, a basic understanding of the user's intention is established. Typically, at least one class and type of user intention are recognized. For example, it is recognized that the user wants to search for a gas station. At the same time, all relevant contextual information about the driver, the vehicle and the environment is taken into account for the intention class thus determined. For example, information about the current driver status (eg, fatigue), information about upcoming appointments, the user's usual behavior in the same or similar situations, and information about the current weather conditions may be included. These contextual data are used to derive new semantic information from the spoken word and to enhance and supplement the previous results.

In addition to the context information mentioned above, the past interaction steps of the user are also stored. This makes it possible to resolve user references to objects that are part of the interaction history. This includes in particular deictic expressions, for example "there" or "this". For example, it is possible to ask for the prices of "this gas station" which was delivered in a previous interaction step as a result of a gas station search. It may also be supplemented for understanding a voice input relevant information, such as the type of fuel required by the vehicle.

Furthermore, a Disambiguierung be made, for example, if a name in the address book is not unique. If the user wants to call a person of this name, semantic analysis based on the context can determine which of the several persons of this name is meant to be more likely. For example, coworkers can be preferred spent when the user is at the workplace. Conversely, private contacts can be preferred in the private sector.

The context information may also include the usual user behavior in the same or similar situations. For example, in situations where an appointment is imminent, the user may rate the proximity of gas stations as more relevant than a low price. Conversely, the cost savings can have higher priority, if no urgent date is pending. Further, for example, in good weather, those restaurants may be preferred in which outdoor seating is possible. The properties of POIs can thus be weighted depending on the context of the speech input and contribute to the semantic analysis of speech input. This leads in particular to the result of the semantic analysis comprising a further understanding of the user intention than is known from other semantic experts. As a result, the analysis effort for other facilities, to which, for example, a request of the user is forwarded, is reduced.

The methods described above for optimizing speech recognition and semantic analysis may be implemented individually or in combination.

In reference to 4 an embodiment of the speech recognition system according to the invention is explained in a vehicle.

The speech recognition system 10 is in the vehicle 6 arranged. Via an interface are other facilities of the vehicle 6 with the speech recognition system 10 coupled: a navigation system 7 , a media player 8th and a vehicle information system 9 , The latter includes in particular sensors that detect properties of the vehicle, the user and the environment. The speech recognition system 10 is also data technology coupled with other facilities, in particular with the Internet server 11 and the mobile user device 12 , The speech recognition system 10 collects data about the context of a voice input from the devices coupled to it and conversely can transmit signals to these other devices. For example, information of the navigation device, such as the route to be traveled, can be used as context information, but the speech recognition system 10 Can also be used as an input device for the navigation system 7 serve. Analogously, this also applies to the other devices shown.

The speech recognition system 10 Here is a combination of the in the 2 and 3 be shown systems. In particular, both the speech recognition and the semantic analysis are optimized by the method according to the invention.

In reference to 5 a further embodiment of the method according to the invention is explained, in which a search request is detected and processed.

In this embodiment, it is assumed that the speech recognition and the semantic analysis are performed as shown above. In particular, from the in 4 represented speech recognition system 10 in the vehicle 6 went out. By means of the semantic analysis, a function is carried out, which is a search function in the case shown. The first steps, for example, by systems such as in the 2 to 4 illustrated speech recognition systems 10 be performed. Based on the semantic analysis, a search function is now executed.

In particular, in the case shown, the result of the semantic analysis is a search query 13 , There is also data about the context of the speech input, the context data 14 , to disposal. Based on the search query 13 and the context data 14 become search criteria 15 certainly. Based on the search criteria thus determined becomes a subset 17 from the database 16 certainly. In particular, in the case shown, the subset 17 ordered, ie the individual data are assigned priorities.

The database 16 includes the set of all possible search results, for example provided by the navigation system 7 of the vehicle 6 with data about POIs, possible routes and entries in an address book. The database may further include all data entries, for example, by a computer system, such as the Internet server 11 , to be provided.

When determining the search criteria 15 also becomes the learned knowledge 18 considered. In the process, the system has learned from past search queries using artificial intelligence methods, for example neural or Bayesian networks, the preferences of the user in certain situations and contexts. In particular, this learned knowledge 18 over time adapted to the individual user.

The main task in determining the search criteria 15 is an incomplete search 13 to complete, so the resulting subset 17 is as relevant as possible and meets the expectations of the user. This will be the search query 13 , the context data 14 and the learned knowledge 18 connected. This will now be illustrated by examples.

In a first example, the user's input is "Search Gas Station". To the collected context data 14 is an urgent appointment in the user's calendar. Further, from the learned knowledge 18 It is known that the user rates a short distance of the gas station more important than a low fuel price in the case of an imminent appointment. Wanted, therefore, gas stations in the shortest possible distance to the traveled route. Conversely, from the learned knowledge 18 known that if no urgent date is imminent, a cheap gas station should be sought. The prioritization of the data of the subset 17 takes place in this case not after a short distance, but after low price. In learning user preferences, context therefore plays a crucial role in allowing the user to have different preferences in different situations and contexts.

In another example, the user's voice input is "Search Restaurant". From the context data is known that the calendar contains no urgent appointment and the weather is good. In combination of this search query 13 and the context data 14 is from the learned knowledge 18 known that the user in good weather and without deadline preferred restaurants that are well rated and where an outdoor sitting is possible. This information is combined to search criteria by which such restaurants from the database 16 selected and in the subset 17 with high priorities. Conversely, for example, the learned knowledge 18 Information about the fact that in bad weather the possibility to dine inside the restaurant is preferred and that in case of deadline The spatial proximity outweighs the quality of the reviews.

In reference to 6 explains the learning process in the context-sensitive search.

As already in 5 shown, becomes a determination of search criteria 15 carried out. This is done on the basis of the (incomplete) search query 13 , it will be contextual data 14 are queried and there are context-sensitive learning preferences from the learned knowledge 18 accessed. This is done in the example shown by a learning unit 18a , Based on the search criteria 15 becomes an advanced search query 15a generated. In the next step, a search is performed and determines whether the extended search 15 corresponds to the intention of the user. If so, this information flows through the lesson 18a into the learned knowledge 18 one. Otherwise, the extended search will be 15a discarded.

The decision, whether the extended search 15a can be accepted or rejected in various ways. For example, a search query may be considered accepted if the user enters on the basis of the extended search query 15a selects specific destination for a navigation. This is evaluated as positive feedback and learned. As a further positive confirmation of the user preference can for example apply that the driver actually approaches the destination. On the other hand, it is also possible that the driver starts the navigation, but later breaks off or changes. The driver goes to the search results based on the extended search 15a not one, but he specifies, for example, the search 13 So this may be a failure of the extension of the search query 13 mean. Furthermore, negative learning can also be implemented, for example if the user makes it clear in a certain way that he rejects the results. This can be a negative correlation for the learned knowledge 18 mean.

Related to the 7A to 7D Embodiments of the method according to the invention are explained, wherein a voice input is corrected.

In a first step, two voice inputs of the user are detected and a text recognition is performed for both voice inputs. The text recognized for the first voice input is the original text 20 and the text recognized for the second voice input is the correction text 21 , Both texts consist of ordered text fragments, in particular individual words. Based on the original text 20 and the correction text 21 becomes a corrected text 22 generated.

At the in 7A example shown is the original text 20 : "My little son is ill. I'm on my way. "The correction text reads:" little Tom is sick ". After these two texts have been recognized, frame words are searched, by which the text to be corrected is defined. In the example shown, a text comparison results in the text fragment 21a "Smaller" of the correction text 21 the text fragment 20b "Smaller" in the original text 20 equivalent. Furthermore, the text fragments correspond 21c and 21d "Is sick" of the correction text 21 the text fragments 20d and 20e "Is sick" of the original text 20 , In this way, the text fragment "smaller" is recognized as the frame preface and the text fragment "is sick" as the frame post-word. In between lies in the original text 20 the text fragment 20c "Son" and in the correction text 21 the text fragment 21b "Tom". The text fragment 20c So "son" should be through the text fragment 21b "Tom" to be replaced. The corrected text 22 is therefore: "My little Tom is sick. I am on the way."

Individual words can be defined as text fragments, but combinations of several consecutive words can also be considered as a text fragment. In particular, it is possible in this method that a plurality of consecutive words constitute a frame prefix and a plurality of consecutive words form a frame postword. Likewise, the correction part may consist of several consecutive words defined by the frame prefix and frame afterword. For example, in this way, a single word in the original text 20 by a succession of successive words in the correction text 21 be replaced. Likewise, conversely, a sequence of several consecutive words in the original text 20 by a single or a different number of words in the correction text 21 be replaced. That is, the number of text fragments in the original text 20 and in the corrected text 22 is not necessarily the same.

The corrected text 22 can be generated by only the text of the correction text between the frame words 21 the text of the original text between the corresponding frame words 20 overrides. Furthermore, the frame words themselves can also be overwritten. In the example shown this corresponds to a replacement of the word "son" by the word "Tom", or the replacement of the words "little son is sick" by the words "little Tom is sick". The result of both approaches is equivalent.

7B shows an example in which the correction text 21 only a frame foreword 21a Includes "smaller". A frame afterword is missing. Through the frame foreword 21a "Smaller" in the correction text 21 can the corresponding Rahmenvorwort 20b "Smaller" in the original text 20 be determined. Accordingly, the subsequent text fragment becomes 20c "Son" in the original text 20 through that on the frame foreword 21a following text fragment 21b "Tom" replaced.

If only the frame prefix is available, then either the subsequent text fragments in the correction text 21 the same number of text fragments in the original text 20 replace, or in the correction text 21 Text fragments following the framework foreword can be found in the original text at the appropriate place 20 be inserted.

7c represents the corresponding case, in contrast to 7b only the frame afterword is present. This consists in the case shown from the text fragments 21b to 21c "is sick". This will cause the text fragment in front of it 21a "Tom" defined as a correction part. Accordingly, in the original text 20 that before the text fragments 20d to 20e "Is sick" text fragment 20c "Son" is defined as a correction part. So "son" is replaced by "Tom". Analogous to that related to 7b Here too, either the same number of text fragments in the original text can be used 20 through the corresponding text fragments in the correction text 21 be overwritten or an insertion can be made.

At the in 7d the case shown contains the correction text 21 only a frame preface, here the text fragment 21a "Son". The following text fragments 21b to 21c "Has fever" make up the correction part. The in 7d illustrated corrected text 22 "My little son has a fever" is generated by the text fragments 20d to 20j ("I'm sick, I'm on my way.") Through the text fragments 21b to 21c ("Has fever.") To be replaced. In this case, the entire text that is arranged after the frame prefix is therefore replaced.

To determine the text to be corrected, other methods can be used alternatively or additionally. For example, a text distance method (such as the Levenshtein distance) can be used to find the position in the original text that has the smallest distance to the corrected text. Furthermore, the system can detect whether a text with high recognition quality or confidence was detected. If a text was not or only insufficiently recognized, this can be communicated to the user, for example graphically or by voice output. A new speech input from the user is interpreted as a correction of that word or phrase, and the corrected text is generated accordingly. Furthermore, the search for similar passages in the original text can be done by finding words with a distinctive number of syllables, ie a similarly long word, by finding words with similar syllables and phonemes, ie similar-sounding words, or by text passages whose words are the same Number of syllables have to be determined.

The above-described embodiments of the speech recognition system and method according to the invention can be combined with each other as desired. For example, only the speech recognition or only the semantic recognition can be enriched by the context data. The extension of underspecified searches can also be done independently. The same applies to the correction of voice inputs, which can also be performed in combination or in isolation.

LIST OF REFERENCE NUMBERS

1

 acquisition unit

1a

 microphone

2

 Voice recognition unit

3

 Semantikerkenner

4

 Context sensing unit

4a

 Context processing unit

5

 execution unit

6

 vehicle

7

 navigation system

8th

 media player

9

 Vehicle information system (with sensors)

10

 Voice recognition system

11

 Internet server

12

 Mobile user device

13

 search query

14

 context data

15

 search criteria

15a

 Extended search

16

 database

17

 Ordered subset

18

 Learned knowledge

18a

 lesson

19

 discard

20

 original

20a-20j

 Text fragments (original text)

21

 Korrigiertext

21a-21d

 Text fragments (correction text)

22

 Corrected text

22a-22j

 Text fragments (corrected text)

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 10329546 A1 [0004]
• EP 2273491 A1 [0005]
• US 2007/0073540 A1 [0006]

Claims (15)

Method for operating a speech recognition system ( 10 ) in a vehicle ( 6 ), in which a) at least one speech input of a user is detected, b) data on the context of the speech input are detected, c) a recognized text of the speech input is generated, d) a semantic analysis of the recognized text is generated, and e) on the basis of semantic analysis performs a function. Method according to claim 1, characterized in that the data relating to the context of the speech input characteristics of the vehicle ( 6 ), the user and / or the environment of the vehicle ( 6 ) affect. Method according to one of the preceding claims, characterized in that further after the generation of the semantic analysis i. depending on the semantic analysis and the data on the context of the speech input, a recognition quality of the recognized text is determined and a language model is selected from a plurality of language models, ii. Based on the selected language model, another recognized text of the speech input is generated, iii. a further semantic analysis of the further recognized text is generated and iv. the steps i. to iii. be repeated until the recognition quality exceeds a predetermined value. Method according to one of the preceding claims, characterized in that, based on the semantic analysis, at least one user intervention class is determined, the data relating to the context of the speech input is detected as a function of the user initiation class and at least one further semantic analysis depending on the data relating to the context of the Voice input is generated. A method according to claim 4, characterized in that semantic analyzes of speech inputs are stored. A method according to any one of claims 5, characterized in that the data on the context of the speech input further comprises semantic analyzes of previous speech inputs. Method according to one of the preceding claims, characterized in that based on the semantic analysis, a search function is performed. A method according to claim 7, characterized in that based on the semantic analysis and the data on the context of the voice input search criteria are determined and based on the specific search criteria, a subset of data of a database is determined and the data of the subset are output. A method according to claim 8, characterized in that the data of the subset of priorities are generated and the output of the data of the subset based on the priorities assigned to them. Method according to one of claims 8 and 9, characterized in that the search criteria are further determined by previous calls of the search function. Method according to one of the preceding claims, characterized in that at least two voice inputs of the user are detected, for each voice input respectively recognized texts ( 20 . 21 ), which consist of ordered text fragments ( 20a - 20j . 21a - 21d ), the text recognized for the first voice input being the original text ( 20 ) and the text recognized for the second voice input is the correction text ( 21 ), and from the original text ( 20 ) and the correction text ( 21 ) a corrected text ( 22 ) is produced. Method according to claim 11, characterized in that the correction text ( 21 ) comprises at least one text fragment indicating the beginning and / or the end of the correction text ( 21 ) signals. Method according to one of claims 11 and 12, characterized in that based on a comparison of the original text ( 20 ) with the correction text ( 21 ) the position of at least one text fragment in the original text ( 20 ), both from the original text ( 20 ) as well as the correction text ( 21 ) and the corrected text ( 22 ) is generated based on the determined position. Method according to one of claims 11 to 13, characterized in that further comprises at least one legacy text fragment of the original text ( 20 ) and at least one new text fragment of the correction text ( 21 ), wherein when generating the corrected text ( 22 ) the old text fragment is replaced by the new text fragment. Speech recognition system ( 10 ) in a vehicle ( 6 ) with - a registration unit ( 1 ), by which at least one user's voice input is detectable, - a context capture unit ( 4 ), by which data relating to the context of the speech input are detectable, - a speech recognition unit ( 2 ), by which a recognized text of the speech input can be generated, - a semantic expert ( 3 ), by means of which a semantic analysis of the recognized text can be generated, and - an execution unit ( 5 ), to which the semantic analysis is transferable and by means of which semantic analysis a function is executable.

Images