Speech recognition system and method for recognising given speech patterns, specially for voice control

0001The invention relates to a speech recognition system and a method for speech recognition ung predetermined speech patterns, in particular for voice control of motor vehicle systems, according to the preamble of claim 1 and 5 respectively.

0002Speech recognition systems are already in use for voice control of motor vehicle functions, for example navigation systems and car telephones. In these speech recognition systems one with a recording device, for. B. compared with a hands-free microphone, detected speech utterance with predetermined speech patterns. Such speech patterns are explained using the example of a car phone, commands such as<img file="EP1069551A2_D0001.tif" />Dialing the phone number", <img file="EP1069551A2_D0002.tif" />Save name ", <img file="EP1069551A2_D0003.tif" />Correction", <img file="EP1069551A2_D0004.tif" />Delete phonebook "as well as individual numbers. In this case, the predetermined speech pattern, which is the most similar to the detected speech utterance, is considered recognized. To determine the similarity speech recognizers, z. B. HMM recognizers, known to detect a speech utterance and determine the probability with which a detected speech utterance match the given speech patterns. The default speech pattern with the highest probability is selected. For a more detailed explanation of the mode of operation of HMM recognizers, reference is made, for example, to EP 0 559 349 A1.

0003If an utterance can not be assigned at all because, for example, the same probability was determined for two speech patterns, the utterance or the command is rejected or the speech recognition system queries in dialogue with the user (<img file="EP1069551A2_D0005.tif" />because these speech recognizers strongly depend on the pronunciation of the user (dialect, sociolect, accent, etc.) can occur in strongly exposed speech in terms of training population to frequent rejections of detection .This leads to the reduction of customer acceptance of such voice controls that are currently being expanded for automotive functions in the future (eg heating / climate control, radio / CD control, turn signal control, lighting control, etc.).

0004It is an object of the invention to improve a speech recognition system and a method for speech recognition of the type mentioned in such a way that a greater independence of the speech is achieved.

0005This object is achieved by the features of claim 1 and 5 procedurally or device. Advantageous developments of the invention are the subject matters of the dependent claims.

0006According to the invention, in a method for speech recognition of given speech patterns, in particular for voice control of motor vehicle systems, in which, in a speech recognition system, a speech utterance recorded with a recording device is compared with the predetermined speech patterns and in which the predetermined speech pattern, that most closely resembles the recorded speech, is recognized as a recognized speech pattern and is selected when a defined uncertainty situation with regard to speech recognition by the speech recognition system, an automatic night training routine allows.

0007A defined uncertainty situation can be a defined uncertainty value. Z. B. For example, an uncertainty value exists if a maximum probability with which a given speech pattern was detected is below a predetermined threshold. There may also be an uncertainty situation if at least two predetermined speech patterns with the same or very similar probabilities (neighbors) are recognized and thus a selection is not possible. Likewise, there is an uncertainty situation if the recorded speech is rejected several times or if requests for a recorded speech are repeatedly output by the speech recognition system.

0008The operator or the user of the speech recognition system may preferentially accept or reject the night training routine. If it is accepted, a predetermined language pattern to be traced is manually selected by the operator or automatically by the voice recognition system in order to carry out the night training routine in the first step. In the second step, the subsequent speech is recorded. In the third text, this recorded speech is assigned to the selected predetermined speech pattern and stored as a new speech pattern in the speech recognition system.

0009In a first alternative, when a defined uncertainty situation for a currently recorded utterance occurs again, the currently recorded speech utterance is first compared with the speech pattern newly stored in the night training routine during a comparison and then selected with sufficient similarity.

0010Thus, the speech recognition system is basically not changed. But, for example, the primary user can retrain vocalizations, which are often rejected.

0011In a second alternative, the speech patterns newly stored in the night training routines are immediately used together with the given speech patterns in a comparison with a currently recorded speech, without waiting for a defined uncertainty situation in a first comparison with the given speech patterns. Although the speech recognition system is changed, in this case all users can retrain.

0012The night training routines can be performed as often as you like. Either the speech utterances detected in the night training routines can all be deposited as speech patterns for the comparison, or only the last utterance associated with a given predetermined speech pattern is deposited as the newly acquired speech pattern for the comparison.

0013In the drawing, an embodiment of the invention is shown. It shows a rough functional sequence within a speech recognition system according to a preferred alternative method.

0014In a microprocessor-controlled speech recognition system with microphone, not shown here, a speech is first detected and forwarded to an HMM speech recognizer 1. In the HMM speech recognizer 1 are also the default speech patterns, z. B. the possible commands of a voice-controlled car phone, stored. The predetermined speech patterns are compared with the detected speech. In the probability calculation block 2, the probabilities P with which the detected speech utterance agrees with each predetermined speech pattern are obtained. The speech pattern with the highest probability P_Max for a match is selected and forwarded to decision block 3. In decision block 3 it is queried whether the greatest probability P_Max is less than a predetermined threshold. In this case, an absolute threshold can be specified or a relative threshold which defines the probability distance to the next most probable speech pattern. Is the maximum probability P_Max greater than the predetermined threshold, the predetermined speech pattern assigned to this probability is selected in the selection block 4 and, if it is a command, executed. If the maximum probability is less than the predetermined threshold, as a result of which there is an uncertainty situation, the training routine block 5 is entered. If there is no retrained newly deposited speech pattern in the night training routine block 5, the user is asked if he wants to perform the night training routine. If yes, the execution of the night training routine begins. In this case, in the first step nachzutrainierendes a predetermined language pattern, for example, manually by the operator, eg. B. via the non-voice-operated controls, or automatically listing, starting with the most probable voice pattern, selected by the speech recognition system. In the second step, the subsequent speech utterance is detected. In the third step, this detected speech utterance is assigned to the selected predetermined speech pattern and stored as a new speech pattern in the night training routine block 5 of the speech recognition system.

0015If an after-trained newly stored speech pattern already exists in the night training routine block 5, the recorded speech utterance is compared with this speech pattern and selected with a high probability (eg also P> S) or with a high degree of similarity. Otherwise, a night training routine will start again.

0016In this embodiment, therefore, a comparison with a newly deposited speech pattern is made only if the probability for an originally given speech pattern is too low or if two probabilities have too small a distance.