DE102023132751A1 - Method for operating a speech dialogue system and speech dialogue system - Google Patents

Abstract

The invention relates to a method for operating a speech dialogue system. First, a request (1) from a user (6) is detected via an interface of the speech dialogue system. Then, a domain recognition unit (8) is activated by a control logic (7) of the speech dialogue system, which is configured to assign the detected request (1) to at least one domain (24), wherein the request (1) is assigned to at least one domain (24) by means of domain recognition. Subsequently, the at least one domain (24) is associated with at least one language model (5) within the speech dialogue system using a decision matrix (3) of the control logic (7), wherein each domain (24) is associated with at least one language model (5). The request (1) is next forwarded to the at least one corresponding language model (5), wherein a feedback message (9) regarding the request (1) is generated based on information stored in the at least one selected language model (5). This feedback message (9) is then output to the user (6).

Description

The invention relates to a method for operating a speech dialogue system. Furthermore, the invention relates to a speech dialogue system configured to implement the method.

In current vehicles from a wide range of manufacturers, the vehicle infotainment system often only offers a rule-based speech dialogue system (SDS) or a voice assistant for voice control. The speech dialogue system is responsible for implementing commands via voice input. Such a voice assistant is often very limited in terms of functionality and knowledge. In order to retrieve new content or knowledge, the speech dialogue system must be extensively developed and released in a new vehicle generation or as an update. This requires speech recognition (natural language understanding) and information for conducting dialogues or interactions with a user to be newly implemented or adapted.

Such a speech dialogue system can therefore only execute a specific type of query or evaluate and answer it online. For many queries, no or only inadequate feedback can be provided. A complex query or a query consisting of multiple questions or instructions thus remains unanswered. For a query that cannot be clearly assigned to a domain, correct feedback is often not generated, even online. Complex questions or tasks, such as logic questions or summaries, or queries consisting of multiple domains, can thus only be answered inadequately or not at all.

The WO 2022/115839 A1 reveals a context-based selection of machine learning models. Wearable devices, such as mobile phones or bedside lamps, can be equipped with a language model so that corresponding functions can be controlled by voice. Hardware features of these devices, such as memory, are taken into account.

The WO 2021/252181 A1 reveals an orchestration of multiple artificial intelligence models. The method is based on training the models to make decisions.

The disadvantage of these methods is that the quality of the answers and decisions of these models can vary greatly in different situations, which can lead to inconsistent or unpredictable results.

The invention is based on the object of reducing the unpredictability of the quality when outputting feedback from a speech dialogue system.

The object is achieved by the subject matter of the independent patent claims. Advantageous developments of the invention are described by the dependent patent claims, the following description, and the figures.

The invention provides a method for operating a speech dialogue system. The method is implemented by the following steps.

In step a), a user request is recorded via an interface of the speech dialog system. Such an interface can be a microphone that captures the user's speech and forwards it to the speech dialog system. A "request" can refer to a request for information or an action. For example, the user can ask the speech dialog system what the weather will be like tomorrow.

In step b), a control logic of the speech dialog system activates a domain recognition unit that is configured to assign the detected request to at least one domain, wherein the request is assigned to the at least one domain by means of domain recognition. This domain recognition can include clustering, e.g., using a k-means algorithm or topic spotting, in particular, e.g., using Latent Dirichlet Allocation (LDA), in order to determine a respective domain. The control logic includes, in particular, decision structures, e.g., if-else statements, and/or loop control and/or error handling logic. The speech dialog system can, for example, be a program module of an operating device of an infotainment system. In the above example, the request can, for example, be assigned the domain "weather."

In step c), at least one language model within the speech dialog system is assigned to the at least one determined domain using a decision matrix of the control logic. The at least one domain is thus associated with at least one language model. Such a language model is, in particular, a pre-trained, i.e., finished language model, such as Chat-GPT and/or Bing and/or Llama.

The request is then routed or forwarded to at least one associated language model.

In step e), feedback on the query is generated based on information stored in at least one associated language model, and this feedback is provided to the user. The information relates to the knowledge and/or data and/or capabilities of the language model. With reference to the above example, the feedback may include information for a weather forecast.

A primary advantage of the invention lies in the independence of the speech dialogue system. By using different (domain-specific) language models, feedback based on the query is more likely to contain relevant information. Furthermore, this can reduce content or semantic ambiguities and/or errors in the output of the feedback. Furthermore, semantic routing of query data is achieved.

The invention also includes embodiments which provide additional advantages.

One embodiment provides that the query is subjected to keyword extraction by the domain recognition unit, for example, and that an assignment unit performs an assignment of predefined keywords to domains, whereby the query is assigned to at least one domain. This keyword extraction can be implemented, for example, using a term frequency and/or term frequency inverse document frequency (TF-IDF). Additionally or alternatively, Word2vec and/or Fast Text can be used for this purpose. The predefined keywords can therefore be assigned to at least one domain, whereby the query is then assigned to that at least one domain. For example, the predefined keywords "weather", "rain" and/or "cold" can be assigned to the domain "weather". This makes it more likely that feedback relevant to the user will be output.

One embodiment provides that the decision matrix controls the at least one language model for generating the feedback depending on at least one of the following parameters and/or settings, whereby the language models are prioritized based on the parameters and/or settings per domain. For this purpose, at least one of the following parameters and/or settings can be used:

The size or the amount of data and/or the typology of the training data of the language model and/or a weighting of the language model depending on at least one associated domain and/or an expected data consumption and/or an expected processing time when using a respective language model. Weighting of the language model depending on at least one associated domain means that a value can be specified, e.g., to what percentage a language model covers and/or contains a domain. Additionally or alternatively, it can be set whether a sequential and/or parallel query should be performed on a language model or a determined domain. These are described in more detail in the embodiments below. Furthermore, an expected data consumption can be set, i.e., the user can, for example, enter a limit value, whereby this limit value describes a maximum value of the data costs to be generated. The at least one language model can therefore also be connected outside the motor vehicle via a radio connection, such as WLAN (Wireless Local Area Network) and/or LTE (Long Term Evolution). This provides the advantage of a structured decision-making process. The decision matrix enables a systematic evaluation, making it easier to select the most relevant language model and/or at least one relevant language model for the user's query.

This decision matrix can additionally or alternatively serve as documentation of the decision-making process. This can be relevant for future reviews and/or for understanding a decision, i.e., why a particular language model was triggered for a specific query. This can ensure transparency of the decision-making process. The decision matrix thus defines the "rules of the game" of the control logic. The decision matrix defines which language model is queried for which domain, in which order, and/or how the responses are evaluated. The decision matrix is therefore a tabular tool that can be implemented as a table.

One embodiment provides that the decision matrix is dynamically expanded with new language models when new language models with their parameters are provided. The decision matrix can be stored in a centralized data structure, such as a database, which facilitates adding and/or updating and/or deleting entries of the decision matrix. The control logic can be configured to regularly check certain storage locations and/or directories and/or repositories in which new language models are provided. The checking can be performed by means of The "FileSystemWatcher" component enables real-time monitoring of file system activities. Additionally or alternatively, such a check can be performed via a backend system and/or a cloud server of the vehicle, e.g., upon startup. This offers the advantage that the speech dialog system can be expanded with new, additional language models.

One embodiment provides that if the confidence of the response to the query is below a predefined threshold and/or if multiple language models have the same prioritization, the query is forwarded in a query cascade, i.e., sequentially, to at least one further language model following the decision matrix until a completeness criterion is met. Confidence of the response means that an internal calculation of the decision matrix (see figure description) contains a value below a predefined threshold. This ensures that the user receives at least relevant feedback.

One embodiment provides that the completeness criterion includes the confidence of the feedback exceeding or exactly matching a predefined threshold and/or the query being forwarded to a predefined number of language models and/or a manual termination of the process. Such a threshold can be, for example, 80% in percentage points or 80 to 90%. A predefined number can be, for example, 5 or in a range from 5 to 10. A manual termination can mean that the user deactivates the process, for example, using a control button on the control device.

One embodiment provides that the request is forwarded to at least two language models in parallel if the at least one domain of the request cannot be clearly determined. This can mean that a domain to be determined cannot be assigned to any of the domains already present in the decision matrix. "Clearly" can therefore mean that the at least one domain of the request is determined with a confidence of 70% in percentage points or 70% to 90%. This can reduce the runtimes of the speech dialog system because multiple language models are queried simultaneously or in parallel. For this purpose, it can be set so that several responses are output one after the other or only the response with the highest confidence is output.

One further development involves the speech dialogue system summarizing the feedback, i.e., reproducing it using fewer written words than the spoken words. By summarizing the feedback, the speech dialogue system can provide a compact representation of the semantic content, such as the linguistic utterance. Instead of reproducing the entire feedback in detail, the speech dialogue system extracts the important key points and/or summarizes them, for example using NLP (natural language processing) and/or deep learning methods. This allows the feedback to be reproduced to be provided quickly. Summarizing the feedback helps to reduce irrelevant and/or redundant information. The speech dialogue system thus identifies the key aspects of the feedback and/or provides a concise representation that contains only the essential information. This can save the user time and/or attention.

According to a further development, the speech dialogue system can provide feedback via speech synthesis. The use of speech synthesis can ensure a natural and/or understandable reproduction of the feedback. The feedback can be converted into a human-like voice, creating a pleasant listening experience for the user. The voice and/or tempo can be configured to ensure optimal intelligibility and/or listening comfort. The speech synthesis can therefore, for example, provide the option of selecting different voices and/or speech styles. The user can choose the one they like best and/or that matches their individual preferences, resulting in a pleasant and/or personally tailored communication experience.

The speech dialogue system and/or the at least one language model can be operated in the motor vehicle and/or in a stationary Internet server.

The speech dialogue system can be an application and/or, for example, part of a smartphone and/or an infotainment system, e.g., of a motor vehicle. The speech dialogue system according to the invention is therefore applicable to any hardware and/or software system.

For use cases or application situations that may arise during the procedure and which are not explicitly described here, it may be provided that, in accordance with the procedure, an error message and/or a request for user feedback is issued and/or which sets a default setting and/or a predetermined initial state.

The invention also includes the control device for the motor vehicle and/or for the speech dialogue system. The control device can have a data processing device or a processor device configured to carry out an embodiment of the method according to the invention. For this purpose, the processor device can have at least one microprocessor and/or at least one microcontroller and/or at least one FPGA (Field Programmable Gate Array) and/or at least one DSP (Digital Signal Processor). In particular, a CPU (Central Processing Unit), a GPU (Graphical Processing Unit), or an NPU (Neural Processing Unit) can be used as the microprocessor. Furthermore, the processor device can have program code configured to carry out the embodiment of the method according to the invention when executed by the processor device. The program code can be stored in a data memory of the processor device. The processor device can, for example, be based on at least a circuit board and/or at least an SoC (System on Chip).

The invention also includes further developments of the speech dialogue system and/or motor vehicle according to the invention, which have features as already described in connection with the further developments of the method according to the invention. For this reason, the corresponding further developments are not described again here.

The motor vehicle according to the invention is preferably designed as a motor vehicle, in particular as a passenger car or truck, or as a passenger bus or motorcycle.

As a further solution, the invention also encompasses a computer-readable storage medium comprising program code which, when executed by a computer or computer network, causes the computer or computer network to execute an embodiment of the method according to the invention. The storage medium can be provided at least partially as a non-volatile data memory (e.g., as a flash memory and/or as an SSD - solid state drive) and/or at least partially as a volatile data memory (e.g., as a RAM - random access memory). The storage medium can be arranged in the computer or computer network. However, the storage medium can also be operated, for example, as a so-called app store server and/or cloud server on the Internet. The computer or computer network can provide a processor circuit with, for example, at least one microprocessor. The program code can be provided as binary code and/or as assembly code and/or as source code of a programming language (e.g., C) and/or as a program script (e.g., Python).

The invention also includes combinations of features from the described embodiments. Thus, the invention also encompasses implementations that each have a combination of features from several of the described embodiments, unless these embodiments have been described as incompatible with one another.

• 1 ein Diagramm einer Ausführungsform des erfindungsgemäßen Verfahrens;
• 2 eine beispielhafte Darstellung einer internen Verarbeitung einer Datenarchitektur gemäß des erfindungsgemäßen Verfahrens; und
• 3 eine schematische Darstellung einer Entscheidungsmatrix gemäß des erfindungsgemäßen Verfahrens.

Exemplary embodiments of the invention are described below. Shown are:

• 1 a diagram of an embodiment of the method according to the invention;
• 2 an exemplary representation of an internal processing of a data architecture according to the method according to the invention; and
• 3 a schematic representation of a decision matrix according to the method according to the invention.

The exemplary embodiments explained below are preferred embodiments of the invention. In the exemplary embodiments, the described components of the embodiments each represent individual features of the invention that can be considered independently of one another, each of which also develops the invention independently of one another. Therefore, the disclosure is intended to encompass combinations of the features of the embodiments other than those shown. Furthermore, the described embodiments can also be supplemented by further features of the invention already described.

In the figures, the same reference symbols denote elements with the same function.

1 shows an embodiment according to the concept. Shown are a user 6, a request 1, a control logic 7 comprising a domain recognition unit 8, a decision matrix 3, and language models 5, wherein the language models 5 are symbolically requested once sequentially 2 and once in parallel 4. A feedback 9 is also shown. A speech dialogue system according to the concept can comprise the decision matrix 3, the control logic 7 together with the domain recognition unit 8, and the language models 5. The decision matrix 3 and/or the domain recognition unit 8 can be part of the control logic 7 or arranged independently of it.

First, a request 1 from the user 6 can be captured via an interface of the speech dialogue system. Then, a domain recognition unit 8 can be activated by a control logic 7 of the speech dialogue system, which is configured to assign the captured request 1 to at least one domain 24, wherein the request 1 is assigned to at least one domain 24 by means of domain recognition. The domain recognition unit 8 can additionally or alternatively perform an a priori recognition and/or have an a priori algorithm. Subsequently, the at least one domain 24 can be associated with at least one language model 5 within the speech dialogue system using a decision matrix 3 of the control logic 7, wherein each domain 24 is associated with at least one language model 5. The request 1 can next be forwarded to the at least one corresponding language model 5, wherein a feedback 9 regarding the request 1 is generated based on information stored in the at least one selected language model 5. This feedback 9 can then be output to user 6.

If a confidence of the feedback 9 to the query 1 is below a predetermined threshold and/or if several language models 5 have the same prioritization, the query 1 can be forwarded in a query cascade or sequentially 2 to at least one further language model 5 following according to the decision matrix 3 until a completeness criterion is reached.

The completeness criterion can mean that the confidence of the feedback 9 exceeds or exactly corresponds to a specified threshold and/or the request 1 has been forwarded to a specified number of language models 5 and/or the procedure is manually aborted.

The request 1 can additionally or alternatively be forwarded to at least two language models 5 in parallel 4 if the at least one domain 24 of the request 1 is not reliably determined.

2 shows an exemplary representation of internal processing of a data architecture according to one embodiment of the concept. This processing can be divided into two components, one component containing an indexing process 12 and the other component having query processing 13. As already mentioned, the speech dialogue system according to the concept can be enriched with additional language models 5. Additionally or alternatively, it can be provided that not only new, already trained language models 5 are embedded, but that additional data 19 is added overall. This data 19 can, for example, be further domain-specific information for already existing language models 5, e.g., taken from manuals and/or user guides and/or order catalogs and/or current information letters. The data 19 can then be subjected to data preprocessing, in particular tokenization 20. The tokenization 20 of data 19 is the process by which a text or sequence is divided into smaller units, so-called "tokens." These tokens can be words, letters, symbols, or other parts of the text, depending on the specific tokenization method. Tokenization 20 is a fundamental task in the field of natural language processing (NLP) and serves to convert text into a form more understandable by computers. Each token represents a discrete unit in the text, making it easier to analyze, process, and understand text data. The tokens can then be converted into numerical representations for use in NLP applications such as machine learning and/or text analysis. Through tokenization 20, embeddings 21 can be created, whereby the embeddings 21 can be considered a training dataset. The training dataset can be divided into several sections, i.e., into a number of training data chunks 22 or so-called "chunks." The training data chunks 22 can then be subjected to semantic indexing 23. In other words, information from the training data sections 22 can be categorized and/or assigned to at least one domain 24. These categorized training data sections 22 can then be stored in a knowledge database 16. The knowledge database 16 can be part of the control logic 7 and/or connected to the decision matrix 3.

Therefore, if a new language model 5 is available, e.g. from the same manufacturer, a knowledge database 16 can be built according to this figure.

If a query 1 is captured and/or received, the query 1 can be processed in the query processing component 13. For this purpose, the captured query 1 is subjected to a so-called query embedding 14. "Query embedding" 14 refers to the process of transforming a query 1 into a vector space, whereby this vector space represents a dense, often low-dimensional representation of the original query 1. From this, a semantic search 15 can be performed, whereby the knowledge database 16 is searched for semantically similar training data sections 22. This can then be used to create a ranking 17. The number of most probable search results 18, i.e., the m semantically most similar training data segments, can then be provided for one or more language models 5. Here, the decision matrix 3 can control one or more language models 5 using, for example, one or more recorded domains 24 from the search results 18, so that feedback 9 is output to the user 6.

3 shows a schematic representation of the decision matrix 3 according to the idea. Various elements are shown in the columns of the decision matrix 3, including a domain 24 and/or a language model type 27 and/or a language model 5 and/or a first weighting factor 31 and/or confidence calculator 26 and/or a second weighting factor 32 and/or a function 25 and/or a result 33. By way of example, in the cells below the column "Domain 24", a domain 24 "Weather" 28 and/or a domain 24 "Knowledge" 29 and/or a domain 24 "Unknown" 30 and/or further domains 24 are symbolically represented.

In the “Language Model Type” column 27, it can be set whether, when the query 1 is assigned to at least one domain 24, several language models 5 should be queried sequentially 2 (CA, cascade) or in parallel 4 (PA, parallel). The respective language models 5, such as ChatGPT and/or LLama, can be listed under the “Language Model” column 5. The first weighting factor 31 can, for example, have a weighting for the respective language models 5 depending on their previous relevance and/or use. In other words, the first weighting factor 31 can represent an a priori assessment that shows a relevance depending on a domain 24. The second weighting factor 32 can, for example, B. represent a factor based on the size and/or typology of the training data sections 22 and/or the training data 19, i.e., have a relevance depending on the data quality to be exhibited by the training data 19. Additionally or alternatively, a weighting factor 31, 32 can be set such that, for example, low data costs and/or a short processing time are generated by a respective language model 5, which is correspondingly represented by a respective weighting factor 31, 32. The function 25 can, for example, specify a number of embeddings 21 depending on the number n of training data used in the language model 5. The confidence calculator 26 can calculate the maximum of all confidences returned in the decision matrix 3. In other words, the confidence calculator 26 calculates the maximum confidence for all cells of the table, e.g., by addition or multiplication, which can then be displayed in the "Result" column 33.

The overall idea involves an extension and connection (to/into the existing speech dialogue system in the vehicle) of several generative AI integrations or language models 5 in order to be able to answer all queries 1, e.g., consisting of knowledge questions, tasks, conversations, as precisely as possible. As already mentioned, general AI-LLM models or language models 5 such as ChatGPT, Llama, and/or open source projects can be integrated, but AI models specialized for specific tasks can also be integrated.

The focus is also on language models 5 (generative AI embeddings), which can be expanded with custom information. As already described, these could be, for example, manuals, user guides, order catalogs, or current information letters. This allows the user 6 to not only ask a query 1 in the form of a general knowledge question, but also to be entertained and/or have tasks solved, and/or request details about vehicle functions, equipment, or information about events related to their vehicle.

The core of the idea is the development of the control logic 7, also called the "dialog conductor," which connects the speech dialog system to the various providers of AI and LLM models or language models 5. The dialog conductor decides which request 1 is distributed to which language models 5, evaluates (arbitrates) the feedback 9 from the language models 5, selects the best feedback 9, and passes it on to the speech dialog system for reading aloud or output. Figuratively speaking, it is a conductor who orchestrates an orchestra of diverse language models 5.

The current concept envisions using the voice assistant primarily for LLM leverage purposes, i.e., for intent recognition between different so-called semantic domains 24 (e.g., knowledge, board, book, weather 28, POI search). Each defined semantic domain 24 would then be fully supported by a generative AI model or a language model 5 with its own word embedding 21—its own semantic search 15. The first phase of the recognition process runs via the control logic 7, where it uses its a priori knowledge to orchestrate which LLM(s) 5 the request 1 should likely be forwarded to. The concept is based on the a priori knowledge of the control logic 7 regarding different semantic domains 24, where the control logic 7 must interact between different embedded LLMs 5. Each embedding 21 enables a semantic search 15 in its knowledge base 16 using the query 1 previously was extracted by control logic 7. Essentially, the knowledge graph of control logic 7 is inherently enhanced by connecting multiple language models 5, without the need for Natural Language Generation (NLG) dialog processing or additional vendor content. A very simple example of an application of this architecture in the automotive sector is trained language models 5 with OEM vehicle manuals, manufacturer brochures, current dealer promotions, or offers. In this case, the selection of a specific AI embedding, i.e., an embedding of a language model 5, is made by control logic 7 based on decision matrix 3.

Due to this architecture, it can happen that in a first request 1 (A), based on the rule in the decision matrix 3, only one LLM 5 is requested, but in another request 1 (B) a whole cascade of LLMs 5 are queried one after the other, i.e. sequentially 2, and in a further request 1 (C) several LLMs 5 could be queried in parallel 4.

The feedback 9 of the LLMs 5 will be provided with a confidence so that, based on the defined elements in the decision matrix 3, a formula for calculating the best result 33 can be calculated.

Overall, the examples show how generative AI integration can be deployed for an automotive voice assistant.

List of reference symbols

1

Inquiry

2

sequential

3

Decision matrix

4

parallel

5

Language model

6

Users

7

Control logic

8

Domain recognition unit

9

Return message

12

Indexing process

13

Request processing

14

Query Embedding

15

semantic search

16

Knowledge base

17

Ranking

18

Search results

19

Data

20

Tokenization

21

Embedding

22

Training data sections

23

semantic indexing

24

domain

25

function

26

Confidence calculator

27

Language model type

28

Weather

29

To know

30

Unknown

31

first weight factor

32

second weight factor

33

Result

QUOTES CONTAINED IN THE DESCRIPTION

This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This 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

• WO 2022/115839 A1 [0004]
• WO 2021/252181 A1 [0005]

Claims (11)

A method for operating a speech dialogue system, comprising the following steps: a) detecting a request (1) from a user (6) via an interface of the speech dialogue system, b) activating a domain recognition unit (8) by a control logic (7) of the speech dialogue system, wherein the domain recognition unit (8) is configured to assign the detected request (1) to at least one domain (24), and assigning the request (1) to the at least one domain (24) by means of a domain recognition of the domain recognition unit (8), c) associating the at least one domain (24) with at least one language model (5) within the speech dialogue system using a decision matrix (3) of the control logic (7), wherein each domain (24) is associated with at least one language model (5), d) routing the request (1) to the at least one associated language model (5), e) generating feedback (9) for the request (1) based on information that stored in at least one associated language model (5) and outputting this feedback (9) to the user (6). Procedure according to Claim 1 , wherein the query (1) is subjected to keyword extraction by the domain recognition unit (8), and an assignment unit carries out an assignment of predetermined keywords to domains (24), whereby the query (1) is assigned to at least one domain (24). Method according to one of the preceding claims, wherein the decision matrix (3) controls the at least one language model (5) for generating the feedback (9) depending on at least one of the following parameters and/or settings, wherein the language models (5) are preferably prioritized based on the parameters and/or settings per domain (24): i) data volume of the training data (19) of the language model (5) and/or ii) typology of the training data (19) of the language model (5) and/or iii) weighting of the language model (5) depending on at least one associated domain (24) and/or iv) expected data consumption and/or v) expected processing time. Method according to one of the preceding claims, wherein the decision matrix (3) is dynamically extended by new language models (5) when new language models (5) with their parameters are provided. Method according to one of the preceding claims, wherein in the event that a confidence of the feedback (9) for the query (1) is below a predetermined threshold value and/or several language models (5) have the same prioritization: forwarding the query (1) in a query cascade (2) to at least one further language model (5) following according to the decision matrix (3) until a completeness criterion is reached. Procedure according to Claim 5 , wherein the completeness criterion comprises that: - the confidence of the feedback (9) exceeds or exactly corresponds to a predetermined threshold value and/or - the request (1) has been forwarded to a predetermined number of language models (5) and/or - an abort of the process is initiated manually. Method according to one of the preceding claims, wherein the request (1) is forwarded to at least two language models (5) in parallel (4) if the at least one domain (24) of the request (1) is not clearly determined. Method according to one of the preceding claims, wherein the feedback (9) is reproduced in summary form. Method according to one of the preceding claims, wherein the feedback (9) is provided as a speech synthesis. A speech dialogue system comprising a processor device having program instructions which, when executed by the processor device, cause it to carry out a method according to one of the preceding method claims. Motor vehicle, comprising a speech dialogue system according to Claim 10 .

Images