WO2015044097A1 - Method and system for creating or augmenting a user-specific speech model in a local data memory that can be connected to a terminal - Google Patents

Abstract

The invention relates to a method for creating or augmenting a user-specific speech model (8) in a local data memory (7) that can be connected to a terminal (4, 5), wherein the speech model (8) is set up to assign control commands for controlling the terminal (4, 5) to natural-speech utterances from a user, comprising the steps of: – a natural-speech utterance being spoken by the user; – the natural-speech utterance being transmitted to an external data processing unit (2) and the natural-speech utterance being transcribed; – the speech model (8) being created or augmented on the external data processing unit (2) using the transcribed natural-speech utterance; – the speech model (8) being transmitted fully or at least partially from the external data processing unit (2) to the local data memory (7) and the speech model (8) being stored and/or augmented in the local data memory (7). The invention also relates to a system (100) that can be used to carry out the method.

Description

description

Method and system for creating or supplementing a user-specific language model in a local data memory connectable to a terminal

The invention relates to a method for creating or supplementing a user-specific language model in a local data memory which can be connected to a terminal, wherein the speech model is set up to associate natural language utterances of a user with control commands for controlling the terminal. The invention further relates to a system with which the method is feasible. From the prior art it is known terminals in ^¬ play as infotainment systems and navigation systems for car or truck to control by entering natural language utterances. Speech recognizers are usually supplemented with statistical language models (SLM) and semantic models (SEM). A natural language utterance is usually a text spoken by the user or sounds uttered by the user. The models mentioned are geared to one of the user z. B. to recognize entered into a corresponding acoustic input unit natural language utterance and the utterance one

Assign control command for controlling the terminal. Usually, a defined action or function of the terminal is then triggered by this control command. However, these models, which are commonly supplemented by known speech recognizers, have the drawback that they are very large and require a lot of memory space to be able to cover the vocabulary of a large number of users. The result is long latency in use. Furthermore, the use of these models requires the use of powerful and correspondingly expensive processors in the terminal. The present invention is therefore based on the object to develop a method for creating a powerful as possible and at the same time as inexpensively implementable language model, the model should be set to associate natural language utterances of a user control commands for controlling a terminal. Furthermore, a system is to be specified, with which this method is feasible. This object is achieved by a method and a system according to the independent claims. Special embodiments of the method and the system according to the invention are described in the subclaims. A method is therefore proposed for creating or supplementing a user-specific language model in a local data memory which can be connected to at least one terminal, wherein the language model is set up to assign natural language utterances of a user to control commands for the terminal, comprising the steps:

 - Speaking of a natural language utterance by the user;

 - Transmission of the natural language utterance to an external data processing unit and transcribing the natural language utterance, preferably transcribing into one

Writing;

 - creating or supplementing the user-specific language model on the external data processing unit based on the transcribed natural language utterance;

completely or at least partially transferring the

 Language model of the external data processing unit to the connectable to the terminal local data storage and storing and / or supplementing the language model in the local data memory.

By creating and / or supplementing the user-specific language model on the external data processing processing unit is carried out and that the language model is supplemented transmitted to ^¬ closing completely or at least partially by the ex ^¬-internal data processing unit to the local data store and stored in the local data store and / or a useful for the control of the terminal language model is provided, which tailored to the respective user. With a comparatively small memory requirement, the language model provided in the local data memory in this way can cover the vocabulary used by this user for controlling the terminal. The local data memory can thus be designed to be smaller than is the case with similarly high-performance speech recognizers which are known from the prior art. At the same time, good results and short latencies can be achieved even with comparatively low-performance and therefore low-cost processors.

The process of transcribing may include, in particular, the creation of an orthographic representation of the natural language utterance, e.g. In the form of a sequence of words and / or characters. Any fonts and alphabets can be used. Preferably, the representation after transcribing is in digital form, e.g. B. in the form of a text file.

Also proposed is a system with which the method proposed here can be carried out. It is thus a system for creating or supplementing a user-specific language model in a local data memory which can be connected to a terminal, wherein the language model is set up to associate natural language utterances of a user with control commands for controlling the terminal. The system includes:

- at least one terminal;

 - A connectable to the terminal local data storage for storing the language model;

- An acoustic input unit, via the natural language utterances for controlling the terminal can be entered; - an external data processing unit to which the natuer ^¬ lichsprachlichen statements can be transmitted, wherein the external data processing unit is configured to create a user-specific language model based on the natural language expressions or supplement and the language model completely or at least partly for storage in the local data memory to the local To transfer data memory and

- A connectable to the acoustic input unit and the local data storage local transmitting and receiving unit for transmitting the natural language utterance to the external data processing unit and for receiving the language model from the external data processing unit. Preferably, the terminal, the local data memory, the acoustic input unit and the local transmitting and receiving ^¬ unit in a motor vehicle can be arranged, arranged or installed, for. B. in a car or truck. The local data memory for storing the language model can be integrated in the terminal. However, the terminal and the local data store may also be implemented as separate units. The acoustic input unit and the local transmitting and receiving unit can also be integrated into the terminal or designed as separate units. The terminal and the local data memory may be connectable or connected via a wired and / or wireless data link. The acoustic input unit typically comprises at least one microphone.

Usually, the system also includes at least one local processor wirelessly with the terminal device, the local data memory, the acoustic input unit and the local transmission and Emp ^¬ capturing unit or wired connectable or connected. Also, the local processor is preferably located in the car or truck. In particular, the local processor may be configured by executing a corresponding one

Speech recognition software to transcribe the entered natural language utterance. Based on the im stored local data store user-specific language model, the local processor of the utterance then assign at least one control command from a set of control commands for controlling the terminal and send this control command to the terminal. The terminal may then perform an action or function initiated by the control command.

The terminal can z. At least

 - an audio device for playing audio files and / or - a video device for playing back image and / or video files and / or

 a telecommunication device for receiving and / or sending text, voice, video or video messages and / or

- a navigation device and / or

- An actuator for adjusting the position of a window, a mirror, a seat or an air ^¬ feeder and / or

 - A system for controlling a temperature and / or humidity and / or

a controller of application software units downloaded from the Internet

include. The terminal may also be an on-board computer or a portable infotainment device, such as a smartphone.

The external data processing unit can, for. B. be a central server of a provider of speech recognizers. In particular, the external data processing unit can be set up to create, supplement, store and send user-specific language models for a multiplicity of different users to various local data memories for storage on these local data memories. Of course, various user-specific language models of different users can also be stored on the same local data store. This can be useful if several people use the same terminal alternately, eg. B. when they use the same vehicle alternately in which the terminal arranged or installed is. For the sake of simplicity only and in the following, only the creation and supplementation of a user-specific language model for only one given user will be described. The external data processing unit at least normally comprises an external transmitting and receiving unit, at least one external processor to create and / or supplement the user-specific language model and at least one ex ^¬-internal data memory in which the user voice is model permanently or at least temporarily stored.

Preferably, the local transmitting and receiving unit and the external transmitting and receiving unit via a wireless Da ^¬ tenverbindung, z. B. via a wireless Internet connection or via any other wireless telecommunications connection with each other for exchanging data connectable.

In a particular embodiment of the proposed method to create or supplement the be ^¬ user-specific language model to the external data processing unit comprises a creation or a complete one be ^¬ user-specific statistical language model (US-SLM), the US-SLM natural language from the transcribed utterance of the user is formed or is added to this and wherein the transcribed natural language utterance transcribed words and / or transcribed sounds and / or trans ^¬ kribierte word sequences and / or transcribed includes sound sequences. Alternatively or additionally, the creation or supplementation of the user-specific language model on the external data processing unit may comprise creating or supplementing a semantic language model (US-SEM), wherein keywords and / or word stems of keywords and / or word sequences are identified in the transcribed natural language utterance of the user and the US SEM from the

Key words and / or the word stems and / or the word sequences is formed or supplemented. Keywords of the semantic model can be trained by the user performing an action and assigning a keyword to it. In both cases, the created user-specific language model is therefore based primarily on the vocabulary, the language and the language habits of each user. The user-specific language model therefore does not require vocabulary and expressions that the user does not even use to enter the control commands. This advantageously contributes to reducing the memory requirements of the user-specific language model, so that the local data memory and the local processor can be made smaller and more cost-effective.

A further specific embodiment of the method is characterized in that the creation or supplementation of the US SEM comprises assigning statistical weights and / or attributes to the keywords and / or the word roots and / or the word sequences, wherein the statistical weights and / or by means of the attributes, an assignment of the respective keyword or word stem or the respective word sequence to a subset of a set of control commands can be established.

In particular, the natural language statement uttered can be assigned at least one control command from a set of control commands, the assignment of the

Control command is made to the natural language uttered expression by means of the user-specific language model stored in the data memory and / or wherein the assignment of the control command to the natural language uttered speech is made by the external data processing unit and the control command is transmitted to the terminal.

The assignment of the control command to the spoken natural-language utterance can then depend, in particular, on the keywords and / or word stems and / or word sequences identified in the natural language utterance and / or depending on the statistical weights and / or attributes assigned to these keywords and / or word stems and / or phrases in the US SEM. Thus, a keyword or a corresponding word root, or one of several terminals designated by means of a corresponding attribute, the amount of those control commands are assigned, which are used to control this terminal. If, for example, an audio device and an exterior mirror are arranged in or on a vehicle, it can be ascertained by means of the semantic language model that an utterance containing the key word "exterior mirror" is an expression for controlling the exterior mirror and not for controlling the exterior mirror The assignment of a control command to a given utterance can therefore be limited to one or more (real) subsets of the set of all possible control commands based on the attributes.

It is conceivable that one and the same keyword can refer to different control commands. For example, in the context of a vehicle exterior mirror, the keyword "misalign" may refer to shifting with respect to different axes, so a plurality of attributes may be associated with a given keyword, with each of the attributes reflecting the assignment to different subsets of all possible control commands. Key words can also be assigned statistical weights, which can identify the different probabilities associated with assigning a given keyword to various control commands.

For example, the keyword "search" may refer to the search in an address book or the search in a list of audio titles, in the case that a given user with the keyword "search" searches in a list in the great majority of cases to initiate audio tracks and not search in an address book, the keyword may be For example, in the semantic model, for this user, "search" will be associated with a first attribute "search in address book" and a second attribute "search in list of titles", the second attribute for this user having, for example, a greater statistical weight than the first attribute.

If an utterance contains several keywords, each of which keyword is assigned to another control command, different probabilities for the assignment of the utterance to the various possible control commands can be determined on the basis of the statistical weights assigned to the keywords. For example, a given utterance may be assigned the slightest control command for which the determined probability is greatest. It can also be provided that a final assignment is only made if the probability thus determined a predetermined

Threshold exceeds.

If a final assignment of the utterance to a particular control command can not be made, the user may be prompted to select a control command from a selection of a plurality of candidate control commands. For this purpose, the system may include an additional input and output unit, for. B. in the form of a touch screen. The user-specific language model can then be adjusted accordingly based on the Inter ^¬ action with the user, and taking into account the input or supplemented. The ^provision or supplementing of the user-specific language model can therefore in particular include an interaction with the user. For example, if the user first uses a term that is not yet part of the language model, then the user may be prompted to associate that term with a particular control command or with a particular subset of control commands. The new term can then be added to the US SLM and / or the US SEM. If appropriate, attributes and / or statistical weights can then be assigned to the term in the US-SEM. So can the custom Language model gradually expanded and cut to the user to ^¬ .

In a specific embodiment of the proposed method, the transcribing of the natural language

Utterance can be made on the external data processing unit and / or on a further external data processing unit that can be connected to the external data processing unit via a data connection. On the external data processing unit and / or on the further external Since ^¬ tenverarbeitungseinheit thus a speech recognition software can be executed or to be executed. In the other external data processing unit, it may be z. B. to act an inter ^¬ netbasierten dictation service. In this case, less Dik ^¬ animal resources have to be kept to transcribe in the (first) external data processing unit so that the first external data processing unit can be implemented cost-effectively. In particular, it may be provided that the transcription of the natural language utterance and / or creating or supplementing of the language model by means of a on the external Since ^¬ tenverarbeitungseinheit or by means of data stored on the further data processing unit further statistical language model and / or other semantic language model and / or a neural network is made. The further statistical language model and the further semantic language model are usually non-user-specific models, ie speaker-independent models. Since they are implemen ^¬ advantage on the external data processing unit and on the other external data processing unit, these models can be extensive without the local data store and / or the local processor must be so designed to be larger or more powerful. This can increase the likelihood that terms or expressions used by the user can be transcribed even if those terms are not yet part of the less extensive user-specific language model stored in the external data processing unit. An additional interaction with the user, in which the user is prompted for input via the additional input unit as described above, can thus be largely avoided.

A further special embodiment of the method is characterized in that is communicated with the natural language utterance a user ID for identifying the user to the data processing unit, wherein the Benut ^¬ zerkennung

 - is entered by the user and / or

 - includes an acoustic voiceprint determined on the basis of the natural language utterance spoken by the user, and / or

 - Is determined when inserting a key into an insertion opening on the basis of a shape of the key and / or

- Is determined by facial recognition and / or

 - Is transmitted by a mobile communication device of the user, in particular from a mobile phone.

This advantageously allows an identification of the user. Thus, the eingesprochene by the user natuer ^¬ lichsprachliche utterance can be used in the external data processing unit to create or complete of that user-specific language model, which is exactly assigned to this user in the external data processing unit. This is particularly advantageous if in the external data processing unit a plurality of benut zerspezifi- see language models each different users are stored, created or supplemented. This is also advantageous for the above-described case in which different users use the same terminal alternately, e.g. B. when they use the same vehicle alternately in which the terminal is arranged or installed. Normally, it is provided that the natural language utterance and / or the user ID via a wireless data connection to the external data processing unit via ^¬ averages and / or that the user specific language model is transmitted via a wireless data link from the Datenverarbei ^¬ processing unit to the terminal. This is preferably the already described wireless Internet connection or other wireless Telekommunikati ^¬ onsverbindung.

An embodiment of the invention is illustrated in the drawings and will be explained in more detail with reference to the following description. In the drawings: Fig. 1 schematically shows an inventive system to create or supplement a user-specific language model in a local data storage, wherein the language model is adapted, of course said ^¬ utterances of a user control commands for controlling a first terminal and / or of a second

Assign terminal,

2 schematically shows steps of a method for creating or supplementing the user-specific language model in the local data memory from FIG. 1 as well as

Fig. 3 shows schematically further steps of the presented in Fig. 2 Darge ^¬ presented method, wherein a control command for controlling one of the terminals is assigned to a user-specific language model stored in the local data memory to a user-specific language model stored in the local data store, the control command is transmitted to this terminal and this terminal performs an action initiated by the control command. 1 shows a system 100 according to the invention. In the exemplary embodiment shown, the system 100 comprises a motor vehicle 1, a first external data processing unit 2 and a second external data processing unit 3. In an interior of the motor vehicle 1, a first terminal 4 and a second terminal 5 are connected ^¬ ordered. The first terminal 4 is an audio device with a plurality of audio files stored therein, which are playable by means of the audio device. Each audio file is z. For example, a title, an artist, a year of recording, and a file size are assigned. In the second terminal 5 is a rearview mirror in the interior of the vehicle 1, the z. B. by means of an electric actuator about a first and a second axis is pivotable. A position of the second terminal 5 is given for example by a first angle of rotation with respect to the first axis and by a second angle of rotation with respect to the second axis.

The terminals 4 and 5 are controllable via control commands.

Control commands concerning the first terminal 4 are, for example, a command for switching on and off the first terminal 4, for displaying all titles of a specific artist, for playing a specific title, for jumping within a given title, for adjusting a volume or for sorting the audio files based on predetermined criteria such as title, artist, Aufnähme ähr, file size, etc. Control commands that affect the second terminal 5, commands for adjusting the mirror, these commands z. B. may include one of the two axes as adjustment axis, an adjustment angle and an adjustment.

The system 100 further comprises a local processor 6, a local data memory 7 for storing at least one user-specific language model 8, which here comprises a user-specific statistical language model (US-SLM) 9 and a user-specific semantic language model (US-SEM) 10 first input unit 11, a second input unit 12 and a local transmitting and receiving unit 13. The local processor 6, the local data memory 7, the input units 8 and 9 and the local transmitting and receiving unit 10 are also arranged in the interior of the vehicle 1. Via data lines 14, the input units 11 and 12, the terminals 4 and 5, the local data memory 7 and the local transmitting and receiving unit 13 are each connected to the local processor 6.

The first input unit 11 comprises at least one acoustic input unit, for. In the form of a microphone through which a user can input natural language utterances. These can then be converted in the first input unit 11 itself or by the local processor 6 into a digital audio file. The second input unit 12 is embodied here as an input and output unit and comprises at least one sensor screen (touch screen), via which the user can interact with the system 100.

The first external data processing unit 2 comprises at least a first external transmitting and receiving unit 15, at least one first external processor 16 and at least one first external data memory 17 for permanent or at least temporary storage of the user-specific language model 8 with the US SLM 9 and the US SEM 10. the first external transmitting and receiving unit 15 is connected via a first wireless Internetver- connection 18 with the local transmitting and receiving unit 13 ver ^¬ prevented. The first external data processing unit 2 is z. B. a server of a provider of speech recognition software. The second external data processing unit 3 we ^¬ nigstens comprises a second external transmitting and receiving unit 19, at least one second external processor 20 and at least one second external data memory 21. In the data memory 21 z. For example, a non-user-specific, ie speaker-independent, further statistical language model 22 and also a non-user-specific further semantic language model 23 are stored. The second external transmit and Receiving unit 19 is connected via a second wireless Internet ^¬ connection 24 with the first external transmitting and receiving unit 15. In the second external data processing unit 3 is z. For example, a server of an Internet-based dictation service that can transcribe audio files into text files.

FIG. 2 shows steps of a method according to the invention for creating or supplementing the user-specific language model 8 in the local data memory 7. Here and in the following, recurring features are each provided with identical reference symbols. The user-specific language model 8 includes di ^¬ gitale data and / or an executable to the local processor 6, and on the first external processor 16 computer program. The language model 8 is set up to associate the natural language utterances spoken by the user via the first input unit 11 with one or more of a plurality of control commands for controlling the first terminal 4 and / or the second terminal 5.

In a first step 25, the user inputs via the touch screen of the second input unit 12 a user identifier that can be stored in digital form. This can be z. B. be a user name in conjunction with a password by which the user is uniquely identifiable. This tells the system 100 that the user speech model ^¬ 8 to be created or updated for this particular user in the following. In modified embodiments, the user identifier may also be given by an acoustic voice print, the z. B. is determined based on a voice spectrum of the natural language utterance spoken by the user. It is also conceivable that in the interior of the vehicle, a camera is arranged, which takes a picture of the user's face, wherein the user is identified by means of a face recognition software. The user ID can also be from a

Mobile phone of the user to the system 100 are transmitted. The user can also use an individual mechanical Key are identified, he pocketed for the purpose of iden ^¬ tification in an opening provided.

In the next step 26, the user enters a natural language utterance into the microphone of the first input unit 11. By entering the natural language utterance, the user intends to trigger a particular action or function of one of the terminals 4 or 5. In the present example, the user speaks the sentence "I want to hear X from Y now." This is to make the first terminal 4 play the audio file stored in the terminal 4, which is assigned the title X and the interpreter Y. Here are X and Y eg one-word sequences, ie in each case consist of only one word In step 27 the spoken utterance of the user is converted into a digital audio file which is transmitted together with the user identification by the local transmitting and receiving unit 13 via the wireless Internet connection 18 is transmitted to the first external transmitting and receiving unit 15 of the first external data processing unit 2. There the audio file with the natural language utterance is assigned on the basis of the user identification to the user-specific language model 8 stored in the first external data memory 17. In step 28 the audio file is included the natural language statement "I want to hear X from Y now "via the second wireless Internet connection 24 to the second external data processing unit 3. In step 29 it is transcribed there by means of the second external Pro ^¬ zessors 20 using the data stored in the data memory 21 non-user-specific language models SLM 22 and SEM 23rd This means here that the second external data processing unit 3 z. For example, a digital text file is created in which said natural language utterance of the user is stored in the form of a sequence of words given by a sequence of characters. For example, includes the transcribed natural language utterance is a sequence of letters of the Latin alphabet, punctuation and spaces. In step 30, the received speech utterance as naturally transcribed in turn in digital form from the second external data processing unit 3 to the first external Datenverar ^¬ beitungseinheit. 2 In alternative embodiments, it is also conceivable that the transcription described in step 29 is performed directly by means of the first external data processing unit 2. The SLM 22 and the SEM 23 can thus also be stored in the first external data processing unit 2. In this case, steps 28 and 30 need not be performed.

In step 31, the user-specific language model 8 is first supplemented in the first external data memory 17 of the first external data processing unit 2 using the transcribed natural language utterance. If the language model 8 does not yet exist, it is created for the first time in step 31. Step 31 includes steps 31a and 31b. In step 31a, the US SLM 9 is supplemented or created in the data memory 17. In step 31b, the US SEM is added or created in the data memory 17.

In the present case, the US SLM 9 of the language model 8 in the data memory 17 already includes a plurality of transcribed word sequences, eg. One-word sequences, two-word sequences, three-word sequences or longer word sequences that the same user has entered or recorded in previous sessions. In step 31, the US SLM 9 will only be naturally transcribed ^¬ utterance "Now I want X heard of Y" adds, when the US-SLM 9 not include this. For example, US-SLM 9 includes the entire sequence of words "I do not want to hear X from Y" yet, so it will be added to US SLM 9 in data store 17. For example, only one-word sequences that are not yet included in the data memory of the US-SLM 9 are added 17 are included. Here, these are, for example, the one-word sequences "X" and "Y" which are added to the US SLM 9 in the data memory 17 in step 31a. Statistical weights of such words or phrases of the statement "I would now like to hear X of Y", which already includes the US SLM 9 in the data memory 17, can be adjusted accordingly, eg increased, on the basis of the newly-spoken utterance.

Similarly, the US SEM 10 already includes a plurality of keywords in the data memory 17, which are each assigned attributes and statistical weights. Here, for example, the US SEM 10 already includes the keyword "listen" in the data memory 17. To this end, the attributes "audio device" and "playback" with a specific statistical weight are already assigned in the US SEM 10. This means, for example ., That natural language ^¬ utterances containing the keyword "hearing", with a non-zero probability that Un ^¬ termenge of all possible control commands is assigned to the first terminal 4, ie the audio device concern.

In the present example, the title "X" and the interpreter "Y" are so far neither part of the US SLM 9 nor the US SEM 10 in the data memory 17. Because of the simultaneous occurrence of the US SLM 9 and the US SEM 10 already however, in step 31b the word "X" and the word "Y" are added to the US SEM 10 in the data memory 17 and there also receive the attributes "audio device" and "playback" assigned to the known keyword "listen". " However, at the same time a larger statistical weight is assigned to the attribute "audio device" than to the attribute "play." So every utterance contains the title "X" with great fortune ^¬ probability to assign a command that affects the audio device. However, not every utterance contains the title "X" focused on playing the title "X" It is conceivable, for example, that the user merely wishes to display the recording year associated with the title "X". Since the words "X" and "Y" are newly entered keywords in the US SEM 10 in the data memory 17, the user is prompted via the touch screen of the second input unit 12 to record the words "X" and "Y "as keywords in the US SEM 10 in the data memory 17 and the assignment of the attributes" audio device "and" play "to the new

Keywords "X" and "Y" to confirm. In addition, the user can be given the opportunity to add further attributes to the new keywords "X" and "Y". Here, for example, he assigns the new keyword "title" to the new keyword "X" and the further attribute "artist" to the new keyword "Y".

In the manner described here, the keywords and / or phrases of keywords of the US SEM 10 in the data memory 17 so statistic weights and attributes are ordered to ^¬ on the basis of which with a given probability assignment of the respective keyword or word sequence of keywords to a certain control command or at least a subset of the set of all possible control commands for controlling the terminals 4 and 5 can be produced. Although only the supplementing of the models US-SLM 9 and US-SLM 10 already present in the data memory 17 is described here, it will be directly apparent to the person skilled in the art how the method can be implemented if the language model 8 is completely different from a natural-language utterance New is created, z. At a very first session of the user. In step 32, the user-specific language model 8 with the US SLM 9 and the US SLM 10 is transmitted completely or at least partially from the first external data processing unit 2 via the wireless Internet connection 18 to the local data memory 7 and stored therein. Preferably, this is an update of one already in the local

Data storage 7 stored previous version of the language model 8. So it is preferably only changes and / or Supplements to the previous version. Such an update may be made at regular intervals automatically or z. B. be taken on a corresponding input from the user before ^¬ . This can be the user z. B. via the second input unit 12 make. It is also conceivable that the updating of the language model is 8 also conducted the local data memory 7 after each session of the user, so every time after the user has eingesprochen a na ^¬ türlichsprachliche statement.

FIG. 3 shows further method steps that can be carried out at the same time as the method steps illustrated in FIG. 2. These further method steps describe how the natural language utterance being spoken is assigned a control command on the basis of the model 8 already stored in the local data memory 7, and an action in one of the terminals 4 and 5 is triggered in accordance with this control command. The supplementing of the model 8 in the local data memory 7 described in FIG. 2 and the control of the terminals 4 and 5 on the basis of the spoken utterance and the version of the model 8 already stored in the data memory 7 can therefore be carried out in parallel.

After entering the user ID and the natural language utterance in steps 25 and 26, in a step 33, the natural language utterance transcription is performed by the local processor 6 using the US SLM 9 stored in the local data memory 7. In step 34, the transcribed utterance is then assigned a control command using the US SEM 10 stored in the local data memory 7. In step 35, this control command is then sent to the appropriate terminal. In step 36, this terminal then executes an action initiated by this control command. This action can z. B. the playback of an audio file by the first terminal 4 or the adjustment of the terminal 5 include.

In the case that the transcription in step 33 use of the US SLM 9 stored in the local data memory 7 or the assignment in step 34 using the US SEM 10 stored in the local data memory 7 can not be carried out because the versions of the US SLM 9 or of the US data stored in the local data memory 7 are not executable. SEM 10 are not yet powerful enough to process the currently uttered utterance of the user, the process steps shown in Fig. 2 are first performed, which is shown in Fig. 3 in the form of the process steps 37 and 38. In this case, therefore, an updating of the language model 8 stored in the local data memory 7 initially takes place.

In the event that the wireless Internet connection 18 between the local transmitting and receiving unit 13 and the first external transmitting and receiving unit 15 can not be produced because the vehicle 1 z. B. through a tunnel or is in an underground car park, only the steps 25, 26, 33, 34, 35 and 36 are performed.

Claims

claims A method for creating or supplementing a user-specific language model (8) in a local data memory (7) which can be connected to a terminal (4, 5), wherein the Language model (8) is arranged, natural language utterances of a user to assign control commands for controlling the terminal (4, 5), comprising the steps: - a natural language utterance by the user;  - transmitting the natural language statement to an external data processing unit (2) and transcribing the natural language utterance;  - creating or supplementing the language model (8) on the external data processing unit (2) based on the transcribed natural language utterance; - Completely or at least partially transmitting the language model (8) from the external data processing ^¬ unit (2) to the local data memory (7) and storing and / or supplementing the language model (8) in the local data memory (7). 2. The method according to claim 1, characterized in that the creation or supplementing of the language model (8) on the external data processing ^¬ unit (2) creating or supplementing a user-specific The US-SLM (9) is formed from the transcribed natürlichsprach- loan utterance or is added to this and wherein the transcribed natural language AEU ^¬ fication transcribed includes statistical language model (US-SLM) (9), words and / or transcribed sounds - and / or transcribed word sequences and / or transcribed sound sequences, and / or - Semantic language model (US-SEM) (10), wherein in the transcribed natural language utterance keywords and / or word stems of keywords and / or word sequences are identified and the US SEM is formed from or supplemented by the keywords and / or the word stems and / or the word sequences. The method of claim 1 or 2, characterized in that the creation or He ^¬ comprises complement of US-SEM (10) that the key words and / or word stems and / or the word sequences statistical weights and / or attributes are assigned, wherein by means of the statistical weights and / or by means of the attributes an assignment of the respective keyword or word stem or the respective word order to a subset of a set of control commands can be produced. Method according to one of the preceding claims, characterized in that a ^¬ spoken natural language utterance, a control command is assigned from a set of control commands, at least, wherein the assignment of the control command to be ^¬ sprochenen natural language utterance by means of the local data memory (7) stored user-specific language model ( 8) is carried out and / or wherein the assignment of the control command to the natural language uttered expression by means of the external data processing unit (2) is carried out and the STEU ^¬ erbefehl to the terminal (4, 5) is transmitted. Method according to claims 3 and 4, in that the assignment of the Control command for the natural language utterance in response to the identified in the natural language utterance keywords and / or word stems and / or phrases and / or depending on these keywords and / or word stems and / or phrases in the US SEM (10) respectively assigned statistical weights and / or attributes. Method according to one of the preceding claims, characterized in that the natural language utterance is transcribed on the external data processing unit (2) and / or on a further external data processing unit (3) connectable to the external data processing unit (2) via a data connection (24). Method according to claim 6, characterized in that the transcribing of the natural language utterance and / or the creation or supplementing of the speech model (8) are stored by means of one on the external data processing unit (2) or by means of one on another external data processing unit (3)  - another statistical language model (22) and / or - Additional semantic language model (23) and / or - neural network is made. Method according to one of the preceding claims, characterized in that the natural language utterance a user ID for identifying the user to the external data processing unit (2) is transmitted, wherein the Benut ^¬ zerkennung  - is entered by the user and / or  - includes an acoustic voiceprint determined on the basis of the natural language utterance spoken by the user, and / or  - Is determined when inserting a key into an insertion opening on the basis of a shape of the key and / or - Is determined by facial recognition and / or from a mobile communication device of the user, especially from a mobile phone. Method according to one of the preceding claims, characterized in that the na ^¬ türlichsprachliche utterance and / or the user ID via a wireless Internet connection (18) to the external data processing unit (2) is transmitted and / or that the user-specific language model (8) via a wireless Internet connection (18) from the external data processing unit (2) to the terminal (4, 5) over ^¬ wear. The system (100) to create or supplement a user-specific language model (8) in a with a terminal (4, 5) connectable local data memory (7), wherein the language model (8) is arranged, of course ^¬ spoken utterances of a user control commands for controlling the Assign terminal (4, 5), comprising: - At least one terminal (4, 5);  - One with the terminal (4, 5) connectable local  Data memory (7) for storing the language model (8); - An acoustic input unit (11), via the natural ^¬ lingual utterances for controlling the terminal (4, 5) can be entered; - an external data processing unit (2) to which the natural language statements can be transmitted, wherein the external data processing unit (2) a ^¬ is directed to create based on the natural language Äuße ^¬ Conclusions The user-specific language model (8) or to supplement and the language model (8 ) completely or at least partially for storage in the local data memory (7) to the local data memory (7) to transmit; and a local transmitting and receiving unit (13) connectable to the acoustic input unit (11) and to the local data memory (7) for transmitting the natural speech utterance to the external Datenverarbei ^¬ processing unit (2) and for receiving the language model (8) from the external data processing unit (2). System according to claim 10, characterized in that the terminal (4, 5)  - an audio device for playing audio files and / or - a video device for playing video and / or video files and / or  - A telecommunication device for receiving and / or  Sending text, voice, picture or video messages and / or  - a navigation device and / or  - An actuator for adjusting the position of a window, a mirror, a seat or an air supply device and / or  - A system for controlling a temperature and / or humidity The terminal is preferably arranged or installed in a vehicle (1), in particular in a car or truck.