METHOD AND DEVICE FOR CONTROLLING A FUNCTIONAL UNIT OF A VEHICLE

Description

[0001] The present invention relates to a method and a device for determining a multi-part keyword in a sound character sequence in a speech utterance of a user.

[0002] US 2018 / 0342237 A1 discloses a method for recognizing a keyword. A device receives a sound sequence and determines the keyword. If a keyword can be determined, the sound sequence is sent to an external server. In a further step, a text derived from the sound sequence is received from the external server and checked for a match with the keyword.

[0003] EP 3 078 021 B1 discloses a method for determining a two-part keyword. Initial activities, such as establishing a network connection, are triggered upon receipt and recognition of a first part of the keyword from a voice input. Upon receipt and recognition of a second part of the keyword from the voice input, a check is performed to determine whether the first and second parts match the keyword. If they match, semantic recognition of further voice inputs is activated.

[0004] EP 3 559 944 B1 discloses a voice control system for the two-stage determination of a keyword, wherein the device comprises a recognition module configured to determine a probability of match with the keyword for at least part of a voice input and to compare this probability with a first threshold. If the probability of match is above the first threshold, the audio signal is transmitted to a further recognition module of a client server for further testing. The further recognition module is configured to determine a further probability of match with the keyword for the voice input and to compare it with a second, more restrictive threshold. If the further probability of match is above the second threshold, the voice control system wakes up and / or triggers an action.

[0005] EP 3 555 881 B1 discloses a method for determining a keyword of an automatic assistant, in which audio data is received and analyzed using a trained model with regard to the probability of a keyword occurring. If it is determined in a first step that the audio data does not contain any keywords, then this received audio data is used to adapt a noise reduction filter. If it is determined in a further step following the first step that further received audio data potentially contains keywords, then filtered data packets are generated from the further audio data using the adapted noise reduction filter. If the presence of keywords is determined when checking the data packets using the trained data model, the assistant is activated.The method increases the probability of recognizing keywords, especially in the presence of background noise.

[0006] US 2016 / 379635 A1 discloses a voice control device in which voice data is received and monitored. If a first part of a keyword is recognized, the voice control device is activated. If the second part of the keyword is recognized after the first keyword, the activation of the voice control device is maintained. If the second part of the keyword is not recognized, the voice control device is deactivated again. Activation in this case involves waking up the voice control device so that voice inputs can be received and evaluated.

[0007] The object of the present invention is to provide an alternative method and a device for recognizing, in particular, a two-part keyword.

[0008] The object is achieved by a method having the features of claim 1 and a device according to claim 7. The dependent claims define preferred and advantageous embodiments of the present invention.

[0009] In the method according to the invention, the received phonetic character sequence is checked for the presence of a first part of the keyword in a first step or a first search phase. A stream of sounds originating from a user is converted into a stream of phonetic characters, i.e., a phonetic character sequence, which is compared with a reference character sequence.

[0010] Only if the first part of the keyword is recognized does a further step, i.e., a second search phase, check the phonetic character string for the presence of a second part of the keyword. If the second part is also recognized, the voice control system is activated. The recognition of the two-part keyword is implemented by first searching the phonetic character string for a first part of the keyword, also called the differentiator. Once this part is recognized, only the second part of the keyword, also called the body, is searched for, within certain time and data volume limits.Activating the voice control system involves an action; for example, upon recognition of the keyword, a vehicle function is triggered or a dialog is opened in which the user is asked via the entertainment system which services or functions they wish to activate or which information they require. Executing an action therefore also includes sending commands from the voice control system to other control systems to operate vehicle functions. The frequent incorrect recognition of a keyword due to the use of part or all of the keyword in a context different from the intended activation is reduced by two-step recognition. The recognition of the first and second parts of the keyword always has a certain error rate. If both have to be recognized independently of one another, this error rate will be much lower than if the keyword is recognized as a whole.If the error rate of the first part of the keyword is, for example, 0.2 and the error rate of the second part of the keyword is 0.1, the overall error rate is 0.02. In contrast, the error rate for recognizing the keyword as a whole is a much higher error rate, for example 0.1. The individual recognition of parts of the keyword in successive phases therefore results in a lower error rate than recognizing the entire keyword. According to the invention, the second part of the keyword is always determined following the most recent first part of the keyword. In this second search phase, the ring buffer is searched for the second part of the keyword behind the most recent first part of the keyword. The most recent first part of the keyword is understood to be the first part in the ring buffer that is closest in time to the current point in time.If the number of phonetic symbols stored in the ring buffer has reached the limit and / or a timer has expired, i.e. a predefined time has expired, the system continues with the search for the first part of the keyword. Otherwise, the system remains in the phase after the search for the second part of the keyword and begins the search for the second part after the most recent first part of the keyword in the current ring buffer. In other words, if the first part is recognized while the search for the second part is carried out, the timer is reset and / or the number of phonetic symbols stored in the ring buffer is counted up again up to the limit and the search for the second part of the keyword is carried out in a further step, starting from the last found first part.By resetting the timer and / or the count of the stored phonetic symbols, an aborted search for the second part of the keyword is related to the nearest first part, ie the probability of an aborted search for the second part of the keyword during voice input is significantly reduced.

[0011] If, after the search, the first part of the keyword is found first, the process is repeated, i.e. the timer is reset again, the phonetic symbols to be stored in the ring buffer are recounted, and in a next step the search is carried out for the second part of the keyword. The process is repeated until either a second part of the keyword following the first part is recognized, or the timer expires and / or the number of phonetic symbols stored in the ring buffer reaches the specified limit. In the first search phase, the search is carried out for the first part; in the second search phase, both the first and second parts of the keyword are searched for.

[0012] This advantageously enables improved recognition of the multi-part keyword in the event that the user aborts or pauses after stating the first part of the keyword and shortly thereafter repeats entering the two-part keyword.

[0013] In a modified embodiment, the first and second parts of the keyword are confirmed as recognized as soon as a similarity measure of the phonetic character string with a first and second stored comparison character string exceeds a predetermined threshold. The similarity measure is determined using methods known from the prior art for string distance, for example, the minimum edit distance.

[0014] In a further embodiment, consecutive identical phonetic symbols in an incoming phonetic string are considered a single phonetic. This advantageously compensates for different pronunciation durations of individual sounds, such as vowels, in a word and reduces misidentification of keywords.

[0015] In a further development of the invention, after a predetermined time has elapsed following recognition of the first part of the keyword, the check for the second part is aborted and the search for the first part is continued. The time is to be calculated such that even a very slowly pronounced second part of the keyword is not aborted due to the expiration of the predetermined time. Advantageously, in the event that the user pauses after pronouncing the first part of the keyword and later accidentally pronounces a word corresponding to the second part of the keyword, interpretation as a keyword and unwanted activation of the voice control system is avoided. The abort enables recognition of a newly spoken multi-part keyword.

[0016] According to a further additional or alternative embodiment of the present invention, the incoming phonetic symbols are continuously buffered in a memory, wherein the number of phonetic symbols stored in the buffer is determined from the recognition of the first part of the keyword. If the number of stored phonetic symbols reaches a predetermined limit, i.e. an upper limit, for example as soon as the first part of the keyword migrates out of the memory buffer designed as a ring buffer, the search for the second part of the keyword is aborted and the search for the first part is continued. As soon as the first part is recognized again, the number of stored phonetic symbols is counted again until the limit is reached.

[0017] This advantageously prevents a situation where, after the first part of the keyword has been deliberately or accidentally mentioned, after further sounds have been stored in the buffer memory and caused by any utterances of the user and noises in the environment, a randomly mentioned word corresponding to the second part of the keyword is mistakenly recognized as belonging to the first part of the keyword, resulting in unwanted activation.

[0018] In an advantageous refinement, the search for the second part of the keyword is aborted as soon as the specified time has elapsed and as soon as the number of stored phonetic symbols reaches the limit, depending on which criterion is met first. Using both termination criteria allows for flexibility in adapting the procedure to speaker habits, thus achieving the best possible recognition of the keyword.

[0019] According to a further embodiment of the method, the incoming sound character sequence is smoothed by placing a continuous observation window over the sound signals, and only those sound signals in the observation window are used for the test if their number within the observation window exceeds a predetermined frequency threshold. The windows are larger than the sound signals cyclically detected by a control unit. For example, sound signals are detected cyclically every 10 ms, i.e., each sound lasting for a certain period of time is, ideally with perfect recognition quality, stored several times in succession in the buffer. The observation window has a value of 50 ms, for example. Now it is determined whether the signals determined in the observation window are stable or fluctuate greatly. If the signals are stable, i.e.If the number of identical sound signals detected in the observation window exceeds the specified frequency threshold, reliable recognition of the sound signals can be assumed. If the frequency threshold is not reached, the signals in the observation window are discarded due to insufficient recognition quality. Recognition errors in the first and second parts of the keyword are advantageously reduced.

[0020] The device according to the invention comprises means for receiving a sound character sequence, wherein the means check the sound character sequence for the presence of a first part of the keyword and only if the first part is recognized the means check the sound character sequence for the presence of a second part of the keyword and only if the second part of the keyword following the first is recognized the means activate the voice control system, where the second part of the keyword is determined following the most recent first part of the keyword.

[0021] The device according to the invention enables reliable recognition of a multi-part keyword.

[0022] The present invention is explained in more detail below using exemplary embodiments with reference to the accompanying drawings.

[0023] It shows: Fig. 1 is a flowchart of a method sequence and Fig. 2 is a voice control system for carrying out the method according to the invention.

[0024] The Fig. 1 The flowchart shown illustrates the process for recognizing a multi-part keyword. The process is executed in a voice control system that cyclically analyzes sounds from the environment. Upon recognition of a multi-part keyword, the voice control system changes its state, leaving a wake-up position, and executes an action. An action can include a dialogue with a user or the execution of a vehicle function such as opening the windows.

[0025] A sequence of sounds from the environment comprising sound signals 1 is fed to the voice control system 3 via a microphone (not shown). In a first step 5, the sound signals are converted into a stream of characters as a sound character sequence and stored in a ring buffer. The search for the keyword or parts of it takes place in the ring buffer. New sound characters are added to the ring buffer, with the same number of sound characters migrating out of the ring buffer.

[0026] In the following step 7, a check is made to determine whether a first part of the keyword (hereinafter referred to as a differentiator) is recognized. If no differentiator is detected, the language system remains in the alert state. If a differentiator is present in the sound string, the system proceeds from the first phase 9 via step 11 to a second phase 13, in which a second part of the keyword (hereinafter referred to as the body) is searched for. In step 11, the ring buffer is transmitted, and in step 15, a check is performed to determine whether the sound string contains a body.

[0027] In step 17, a check is made to determine whether a predefined time has elapsed since the differentiator was detected and whether the number of phonetic symbols stored in the ring buffer has reached a predefined limit. If one or both of these termination conditions are met, the program returns to step 5, and starts searching for the differentiator again. If the conditions are not met and the current check in step 19 shows that there is no body in the phonetic symbol sequence in the ring buffer, the program returns to step 15, from which point the current ring buffer, supplemented by newly accumulated phonetic symbols, is checked. If a body is found, the two-part keyword is deemed to have been recognized, and in step 21, the speech dialog system is switched from the wake-up mode to a mode for executing an action or to a dialog mode.Following step 21, the system is returned to the awake position in step 5.

[0028] In a variation of the method described above, in step 19 the body behind the most recent differentiator in the ring buffer transmitted in step 11 is searched for. Upon entry into the second phase 13, in addition to searching for the body, the current ring buffer is continuously searched for a most recent differentiator. Once this is found, monitoring of the termination conditions is started, i.e. the predefined time is restarted and / or the number of phonetic symbols stored in the ring buffer is incremented again. After restarting the termination condition, a further step is taken to search only for the body. If no body is found after the most recent differentiator, then after the predefined time has elapsed and / or as soon as the number of stored phonetic symbols reaches the limit, in each case referenced to the time at which the most recent differentiator was detected, the process switches back to phase 9 in step 5.This ensures that if, for example, a user interrupts the input of a keyword "Hello Car" after "Hello.." and then continues with "Hello Car" after a pause, a time limit or a quantity limit of the stored sounds in the buffer memory does not refer to the first but to the second hello, so that the probability of interruption during input due to the pause is reduced.

[0029] Fig. 2 shows a voice control system 3, which comprises a means 23 for receiving and evaluating sound sequences. The means 23 contains, for example, a microphone 25 and a computing unit 27 connected to the microphone 25 for processing the transmitted signals. Furthermore, the computing unit 27 is configured to transmit signals 29 to other control devices based on the evaluation or to initiate a dialogue with a user. In the computing unit 27 of the means 23, the signals are processed as described in the Fig. 1carried out in phase 9 and subsequently processed step by step in phase 13.

Claims

1. Method for recognizing two-part keywords for activating a voice control system, the method comprising the following steps: - receiving a phonetic character sequence (5), - buffering the phonetic character sequence in a memory, - checking the phonetic character sequence for the presence of a first part of the keyword (7) and, only if the first part is recognized, wherein, when a limit of the number of phonetic characters stored in the memory is reached and / or after a predefined period of time has elapsed, the check for the first part of the keyword (7) is continued, - checking the phonetic character sequence for the presence of a second part of the keyword (19), wherein, additionally, a continuous search is carried out for a most recent first part of the keyword (7), characterized in that only if the second part of the keyword following the first is recognized, - activating the voice control system (21) to carry out an action, wherein the second part of the keyword is ascertained following the most recent first part of the keyword.

2. Method according to claim 1, characterized in that the first and the second part of the keyword are confirmed as recognized as soon as a similarity measure of the phonetic character sequence with a first and second stored comparison character sequence is above a predefined threshold value.

3. Method according to either of claims 1 or 2, characterized in that consecutive identical phonetic characters in an incoming phonetic character sequence are considered to be a single phonetic character.

4. Method according to any of claims 1 to 3, characterized in that after a predefined period of time after recognition of the first part of the keyword, the check for the second part is terminated (17) and the search for the first part is continued (7).

5. Method according to any of claims 1 to 4, characterized in that the incoming phonetic character sequence is continuously buffered in the memory, wherein the number of phonetic characters stored in the buffer is determined as from recognition of the first part of the keyword, wherein, when a predefined limit is reached, the search for the second part of the keyword is terminated (17) and the search for the first part is restarted (7).

6. Method according to any of claims 1 to 5, characterized in that the incoming phonetic character sequence is smoothed by placing a continuous observation window over the phonetic character sequence and only phonetic characters in the observation window whose number within the observation window exceeds a predefined frequency threshold are used for the check.

7. Voice control system for carrying out the method according to claims 1 to 6, comprising means (23) for receiving and buffering a phonetic character sequence, wherein - the means (23) check the phonetic character sequence for the presence of a first part of the keyword and, only if the first part is recognized, wherein, when a limit of the number of phonetic characters stored in the memory is reached and / or after a predefined period of time has elapsed, the check for the first part of the keyword is continued, - the means (23) check the phonetic character sequence for the presence of a second part of the keyword following the first part, wherein, additionally, a continuous search is carried out for a most recent first part of the keyword (7), characterized in that only if the second part of the keyword following the first is recognized, - the means (23) activate the voice control system to carry out an action, wherein the second part of the keyword is ascertained following the most recent first part of the keyword.