Cognitive audio classifier
A method, system, and computer program product are provided for classifying spoken audio content with a cognitive audio classifier by applying a set of distorted audio resources through a set of speech-to-text models STTi (STT 1 . . . STTn) to get a set of interference coherence scores based on the transcript for each speech-to-text model STTi, thereby generating a measured baseline Mi (M 1 . . . Mn) and a practical baseline Pi (P 1 . . . Pn) that is associated with a coherence matrix for the audio effects AEj (AE 1 . . . AEk) that were used to generate the distorted audio resources, thereby generating training data for use in training a cognitive audio classifier which classifies input spoken audio content to measure a quality of detected vocabulary elements from the spoken audio content under the set of audio distortion effects for each speech-to-text model STTi.
1. A method, in an information handling system comprising a processor and a memory, of classifying spoken audio content, the method comprising:
receiving, by the information handling system, a baseline audio resource and associated transcript;
applying, by the information handling system, a set of audio distortion effects AEj (AE 1 , AE 2 , . . . AEk) to the baseline audio resource to generate a set of distorted audio resources;
applying, by the information handling system, each distorted audio resource through a set of speech-to-text models STTi (STT 1 , STT 2 , . . . STTn) to get an interference coherence score for each speech-to-text model STTi based on the transcript as a practical baseline Pi (P 1 , P 2 , . . . Pn) that is associated with a coherence matrix for the audio effects AEj (AE 1 , AE 2 , . . . AEk) that were used to generate the distorted audio resources;
generating, by the information handling system, training data from the practical baseline Pi (P 1 , P 2 , . . . Pn) for use in training a cognitive audio classifier; and
applying, by the information handling system, spoken audio content to the cognitive audio classifier to measure a quality of detected vocabulary elements from the spoken audio content under the set of audio distortion effects for each speech-to-text model STTi.
2. The method of claim 1 , further comprising processing the baseline audio resource and spoken audio content to extract audio features and metadata.
3. The method of claim 1 , where applying each distorted audio resource comprises processing, by the information handling system, outputs from the set of speech-to-text models STTi (STT 1 , STT 2 , . . . STTn) with a discourse analysis model to get the interference coherence score for each speech-to-text model STTi as the practical baseline Pi (P 1 , P 2 , . . . Pn).
4. The method of claim 1 , where training the cognitive audio classifier comprises training a long short-term memory (LSTM) neural network classifier with the practical baseline Pi (P 1 , P 2 , . . . Pn) from the set of distorted audio resources which have a common quality level.
5. The method of claim 4 , where the common quality level is a human labeled noise level for each distorted audio resource.
6. The method of claim 4 , where training the cognitive audio classifier comprises training a deep learning service by ranking outcomes of speech-to-text models STTi using the interference coherence scores generated from the set of distorted audio resources.
7. The method of claim 1 , where applying spoken audio content to the cognitive audio classifier comprises applying the practical baseline Pi (P 1 , P 2 , . . . Pn) to measure a quality of the spoken audio.
8. An information handling system comprising:
one or more processors;
a memory coupled to at least one of the processors;
a set of instructions stored in the memory and executed by at least one of the processors to classify audio content, wherein the set of instructions are executable to perform actions of:
receiving, by the system, a baseline audio resource and associated transcript;
applying, by the system, a set of audio distortion effects AEj (AE 1 , AE 2 , . . . AEk) to the baseline audio resource to generate a set of distorted audio resources;
applying, by the system, each distorted audio resource through a set of speech-to-text models STTi (STT 1 , STT 2 , STTn) to get an interference coherence score for each speech-to-text model STTi based on the transcript as a practical baseline Pi (P 1 , P 2 , . . . Pn) that is associated with a coherence matrix for the audio effects AEj (AE 1 , AE 2 , . . . AEk) that were used to generate the distorted audio resources;
generating, by the system, training data from the practical baseline Pi (P 1 , P 2 , . . . Pn) for use in training a cognitive audio classifier; and
applying, by the system, audio content to the cognitive audio classifier to measure a quality of detected vocabulary elements from the audio content under the set of audio distortion effects for each speech-to-text model STTi.
9. The information handling system of claim 8 , wherein the set of instructions are executable to process the baseline audio resource and audio content to extract audio features and metadata.
10. The information handling system of claim 8 , wherein the set of instructions are executable to apply each distorted audio resource by processing outputs from the set of speech-to-text models STTi (STT 1 , STT 2 , . . . STTn) with a discourse analysis model to get the interference coherence score for each speech-to-text model STTi as the practical baseline Pi (P 1 , P 2 , . . . Pn).
11. The information handling system of claim 8 , wherein the set of instructions are executable to train the cognitive audio classifier by training a long short-term memory (LSTM) neural network classifier with the practical baseline Pi (P 1 , P 2 , . . . Pn) from the set of distorted audio resources which have a common quality level.
12. The information handling system of claim 11 , where the common quality level is a human labeled noise level for each distorted audio resource.
13. The information handling system of claim 11 , wherein the set of instructions are executable to train the cognitive audio classifier by training a deep learning service by ranking outcomes of speech-to-text models STTi using the interference coherence scores generated from the set of distorted audio resources.
14. The information handling system of claim 11 , wherein the set of instructions are executable to apply audio content to the cognitive audio classifier by applying the practical baseline Pi (P 1 , P 2 , . . . Pn) to measure a quality of the audio content.
15. A non-transitory computer program product stored in a computer readable storage medium, comprising computer instructions that, when executed by a processor at an information handling system, causes the system to classify spoken audio content by:
receiving, by the system, a baseline audio resource and associated transcript;
applying, by the system, a set of audio distortion effects AEj (AE 1 , AE 2 , . . . AEk) to the baseline audio resource to generate a set of distorted audio resources;
applying, by the system, each distorted audio resource through a set of speech-to-text models STTi (STT 1 , STT 2 , . . . STTn) to get an interference coherence score for each speech-to-text model STTi based on the transcript as a practical baseline Pi (P 1 , P 2 , . . . Pn) that is associated with a coherence matrix for the audio effects AEj (AE 1 , AE 2 , . . . AEk) that were used to generate the distorted audio resources;
generating, by the system, training data from the practical baseline Pi (P 1 , P 2 , . . . Pn) for use in training a cognitive audio classifier; and
applying, by the system, spoken audio content to the cognitive audio classifier to measure a quality of detected vocabulary elements from the spoken audio content under the set of audio distortion effects for each speech-to-text model STTi.
16. The non-transitory computer program product of claim 15 , further comprising computer instructions that, when executed by the system, causes the system to process the baseline audio resource and spoken audio content to extract audio features and metadata.
17. The non-transitory computer program product of claim 15 , further comprising computer instructions that, when executed by the system, causes the system to apply each distorted audio resource by processing outputs from the set of speech-to-text models STTi (STT 1 , STT 2 , . . . STTn) with a discourse analysis model to get the interference coherence score for each speech-to-text model STTi as the practical baseline Pi (P 1 , P 2 , . . . Pn).
18. The non-transitory computer program product of claim 15 , further comprising computer instructions that, when executed by the system, causes the system to train the cognitive audio classifier by training a long short-term memory (LSTM) neural network classifier with the practical baseline Pi (P 1 , P 2 , . . . Pn) from the set of distorted audio resources which have a common quality level.
19. The non-transitory computer program product of claim 15 , further comprising computer instructions that, when executed by the system, causes the system to train the cognitive audio classifier by training a deep learning service by ranking outcomes of speech-to-text models STTi using the interference coherence scores generated from the set of distorted audio resources.
20. The non-transitory computer program product of claim 15 , further comprising computer instructions that, when executed by the system, causes the system to apply spoken audio content to the cognitive audio classifier by applying the practical baseline Pi (P 1 , P 2 , . . . Pn) to measure a quality of the spoken audio.