Methods, apparatus and computer programs for automatic speech recognition
View Patent ↗An automatic speech recognition (ASR) system includes a speech-responsive application and a recognition engine. The ASR system generates user prompts to elicit certain spoken inputs, and the speech-responsive application performs operations when the spoken inputs are recognized. The recognition engine compares sounds within an input audio signal with phones within an acoustic model, to identify candidate matching phones. A recognition confidence score is calculated for each candidate matching phone, and the confidence scores are used to help identify one or more likely sequences of matching phones that appear to match a word within the grammar of the speech-responsive application. The per-phone confidence scores are evaluated against predefined confidence score criteria (for example, identifying scores below a ‘low confidence’ threshold) and the results of the evaluation are used to influence subsequent selection of user prompts. One such system uses confidence scores to select prompts for targetted recognition training—encouraging input of sounds identified as having low confidence scores. Another system selects prompts to discourage input of sounds that were not easily recognized.
1. A method for controlling operation of an automatic speech recognition (ASR) system, comprising acts of:
comparing sounds within an input audio signal with phones of an acoustic model to identify candidate matching phones, the input audio signal comprising a first user input;
calculating recognition confidence scores for individual candidate matching phones;
evaluating the recognition confidence scores to identify at least one phone of the candidate matching phones having a particular recognition confidence characteristic, wherein the act of evaluating the recognition confidence scores comprises combining the calculated recognition confidence scores with a calculated inherent likelihood of confusion between a phone and at least one other shone to obtain a combined result and then comparing the combined result with the particular recognition confidence characteristic; and
selecting a user prompt for eliciting at least one subsequent user input that differs from the first user input, wherein the user prompt is selected based at least in part on a determination of whether the user is expected to speak the identified at least one phone in response to the user prompt.
2. The method of claim 1 , wherein:
the first user input is associated with a first operation of the ASR system;
the at least one subsequent user input is associated with a second operation of the ASR system;
the at least one phone is a candidate matching phone for a sound within the first user input; and
the user prompt for eliciting the at least one subsequent input is selected based on the recognition confidence score for the candidate matching phone as a match for the sound within the first user input.
3. The method of claim 1 , wherein the user prompt is selected to encourage input of phones identified as having low confidence recognition scores for training the acoustic model.
4. The method of claim 2 , further comprising acts of:
comparing sounds within a subsequently input audio signal with phones of the acoustic model to identify candidate matching phones;
calculating recognition confidence scores for the candidate matching phones; and
updating a recognition confidence score that relates a recognition confidence score for the first user input and a recognition confidence score for the subsequent user input.
5. The method of claim 1 , wherein the act of selecting the user prompt comprises comparing phones identified as having low recognition confidence scores with a plurality of optional user prompts and expected input phones associated with the optional user prompts, and wherein the selected user prompt is one of the plurality of optional user prompts associated with an expected input phone that is identified as having a likelihood of confusion with at least one other phone above a determined threshold.
6. The method of claim 1 , wherein the user prompt is selected to discourage input of phones identified as having low confidence recognition scores.
7. The method of claim 6 , wherein the act of selecting the user prompt comprises:
selecting, from a plurality of synonyms, a synonym that comprises a phone identified as having a low recognition confidence score; and
selecting a user prompt in response to which the user is expected to speak the selected synonym.
8. The method of claim 1 , wherein the act of calculating an inherent likelihood of confusion comprises calculating a distance between a first state of an acoustic model and at least one other state of the model nearest to the first state, the first state corresponding to a first sound.
9. The method of claim 1 , wherein an application grammar is modified in response to the calculated recognition confidence scores.
10. The method of claim 9 , wherein said modification of the application grammar comprises acts of:
identifying at least one word within the application grammar associated with a confidence recognition score below a threshold score; and
replacing said identified at least one word within the application grammar with a synonym.
11. The method of claim 10 , further comprising an act of checking that an inherent confusability between said synonym and at least one other word in the grammar is below a threshold before carrying out said act of replacing.
12. At least one non-transitory computer-readable recording medium having encoded thereon executable program code that, when executed by at least one processor, performs a method for controlling operation of an automatic speech recognition (ASR) system, the method comprising acts of:
comparing sounds within an input audio signal with phones of an acoustic model to identify candidate matching phones;
calculating recognition confidence scores for each of the candidate matching phones;
evaluating the recognition confidence scores for the candidate matching phones to identify at least one phone having a particular recognition confidence characteristic, wherein the act of evaluating the recognition confidence scores comprises combining the calculated recognition confidence scores with a calculated inherent likelihood of confusion between a phone and at least one other shone to obtain a combined result and then com arin the combined result with the particular recognition confidence characteristic; and
selecting a user prompt to elicit at least one subsequent user input that differs from the first user input, wherein the user prompt is selected based at least in part on a determination of whether the user is expected to speak the identified at least one phone in response to the user prompt.
13. The at least one non-transitory computer-readable recording medium of claim 12 , wherein the user prompt is selected to encourage input of phones identified as having low confidence recognition scores for training the acoustic model.
14. The at least one non-transitory computer-readable recording medium of claim 12 , wherein the act of selecting the user prompt comprises comparing phones identified as having low recognition confidence scores with a plurality of optional user prompts and expected input phones associated with the optional user prompts, wherein the selected user prompt is one of the plurality of optional user prompts associated with an expected input phone that is identified as having a likelihood of confusion with at least one other phone above a determined threshold.
15. The at least one non-transitory computer-readable recording medium of claim 12 , wherein the user prompt is selected to discourage input of phones identified as having low confidence recognition scores.
16. The at least one non-transitory computer-readable recording medium of claim 15 , wherein the act of selecting the user prompt comprises:
selecting, from a plurality of synonyms, a synonym that comprises a phone identified as having a low recognition confidence score; and
selecting a user prompt in response to which the user is expected to speak the selected synonym.
17. The at least one non-transitory computer-readable recording medium of claim 12 , wherein the method further comprises acts of:
identifying at least one word within an application grammar associated with a confidence recognition score below a threshold score; and
replacing said identified at least one word within the application grammar with a synonym.
18. A system for controlling operation of an automatic speech recognition (ASP,) system, the system comprising:
at least one processor; and
at least one computer readable medium storing instructions that program the at least one processor to:
compare sounds within an input audio signal with phones of an acoustic model to identify candidate matching phones;
calculate recognition confidence scores for each of the candidate matching phones;
evaluate the recognition confidence scores for the candidate matching phones to identify at least one phone having a particular recognition confidence characteristic, wherein the act of evaluating the recognition confidence scores comprises combining the calculated recognition confidence scores with a calculated inherent likelihood of confusion between a phone and at least one other phone to obtain a combined result and then comparing the combined result with the particular recognition confidence characteristic; and
select a user prompt to elicit at least one subsequent user input that differs from the first user input, wherein the user prompt is selected based at least in part on a determination of whether the user is expected to speak the identified at least one phone in response to the user prompt.
19. The system of claim 18 , wherein the user prompt is selected to encourage input of phones identified as having low confidence recognition scores for training the acoustic model.
20. The system of claim 18 , wherein the instructions further program the at least one processor to select the user prompt at least in part by comparing phones identified as having low recognition confidence scores with a plurality of optional user prompts and expected input phones associated with the optional user prompts, and wherein the selected user prompt is one of the plurality of optional user prompts associated with an expected input phone that is identified as having a likelihood of confusion with at least one other phone above a determined threshold.
21. The system of claim 18 , wherein the user prompt is selected to discourage input of phones identified as having low confidence recognition scores.
22. The system of claim 21 , wherein the instructions further program the at least one processor to select the user prompt at least in part by:
selecting, from a plurality of synonyms, a synonym that comprises a phone identified as having a low recognition confidence score; and
selecting a user prompt in response to which the user is expected to speak the selected synonym.
23. The system of claim 18 , wherein the instructions further program the at least one processor to:
identify at least one word within an application grammar associated with a confidence recognition score below a threshold score; and
replace said identified at least one word within the application grammar with a synonym.