Diagnostic service in speech recognition

1 . A computing system configured to identify targeted datasets configured to facilitate an improvement in an accuracy of an acoustic model included in an automatic speech recognition system, whereby an overall accuracy of the automatic speech recognition system is improved, the computing system comprising:

one or more processors; and

one or more hardware storage devices storing one or more computer-executable instructions that are executable by the one or more processors to configure the computing system to:

obtain a test dataset comprising (i) audio data having a plurality of natural speech utterances and (ii) a transcription of the plurality of natural speech utterances;

generate a text-to-speech dataset comprising audio data having a plurality of synthesized speech utterances based on the transcription of the plurality of natural speech utterances;

apply the test dataset and the text-to-speech dataset to the automatic speech recognition system to obtain a first model output corresponding to the test dataset and a second model output corresponding to the text-to-speech dataset;

identify a first set of errors in the first model output and a second set of errors in the second model output; and

based on comparing the first set of errors and the second set of errors, generate an acoustic model error ratio; and

use the acoustic model error ratio to identify a set of characteristics for one or more additional datasets to be applied to the automatic speech recognition system to facilitate an improvement in the acoustic model error ratio when applying at least one of the one or more identified additional datasets to the automatic speech recognition system.

2 . The computing system of claim 1 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

distinguish between the one or more identified additional datasets configured to be applied to the automatic speech recognition system, based on the set of attributes and corresponding acoustic model error ratio, to identify and select a targeted dataset from the one or more identified additional datasets to apply as training data to the automatic speech recognition system to facilitate an improvement in the acoustic model error ratio of the automatic speech recognition system.

3 . The computing system of claim 2 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

train the automatic speech recognition system with the targeted dataset to modify the automatic speech recognition system to generate refined output that is associated with fewer errors than first model output; and

determine that the acoustic model has an improved acoustic model error ratio based on analyzing the refined output.

4 . The computing system of claim 2 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

identify one or more characteristics associated with the test dataset, wherein the targeted dataset is selected from the one or more additional datasets based on the targeted dataset corresponding to the one or more characteristics.

5 . The computing system of claim 4 , wherein the one or more characteristics are identified from: a particular speaking scenario or a particular set of background noises.

6 . The computing system of claim 4 , wherein the one or more characteristics are associated with a particular enterprise domain.

7 . A computing system configured to identify targeted datasets to be used to facilitate an improvement in an accuracy of a language model and a lexicon included in an automatic speech recognition system, the computing system comprising:

one or more processors; and

one or more hardware storage devices storing one or more computer-executable instructions that are executable by the one or more processors to configure the computing system to:

apply as input a test dataset comprising (i) a text transcription and (ii) audio data having a plurality of synthesized speech utterances generated using a text-to-speech system based on the text transcription;

obtain a first output corresponding to the lexicon based on the test dataset;

identify a first set of errors in the first output and distinguish between one or more errors included in the first set of errors according to different types of errors associated with the lexicon;

obtain a second output corresponding to the language model based on the test dataset;

identify a second set of errors in the second output and distinguish between one or more errors in the second set of errors according to different types of errors associated with the language model;

generate a language model error ratio for the language model and a lexicon error ratio for the lexicon; and

based on the first set of errors and second set of errors, identify and distinguish between one or more additional datasets configured to be applied to the language model and the lexicon to facilitate an improvement in the automatic speech recognition system.

8 . The computing system of claim 7 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

distinguish between the one or more errors at least based on at least on the generated text-to-speech dataset.

9 . The computing system of claim 7 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

prompt a user to obtain the one or more additional datasets;

receive the one or more additional datasets obtained by the user; and

apply the one or more additional datasets to the automatic speech recognition system in order to facilitate an improvement in the lexicon error ratio and/or the language model error ratio.

10 . The computing system of claim 9 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

identify one or more words or phrases that correspond to the second set of errors, wherein at least one of the one or more additional datasets comprises a plurality of sentences including a particular word or phrase for each of the one or more words or phrases.

11 . The computing system of claim 9 , the one or more computer-executable instructions being further executable by the one or more processors to further configured the computing system to:

identify one or more words that correspond to the first set of errors; and

prompt a user to verify a pronunciation of each of the one or more words, wherein at least one of the one or more additional datasets comprises a new pronunciation for each of the one or more words that the user has determined does not have a correct pronunciation included within the lexicon.

12 . A computing system configured to identify targeted datasets to be used to facilitate an improvement in an accuracy of an automatic speech recognition system, the computing system comprising:

one or more processors; and

one or more hardware storage devices storing one or more computer-executable instructions that are executable by the one or more processors to configure the computing system to:

apply as input a test dataset to the automatic speech recognition system to obtain an initial output, the automatic speech recognition system comprising an acoustic model, a language model, and a lexicon, the test dataset comprising (i) a text transcription, (ii) audio data having a plurality of natural speech utterances and a plurality of synthesized natural speech utterances;

identify one or more types of errors included in the initial output, the initial output;

based on the one or more types of errors included in the initial output, generate an error score for each of the acoustic model, the language model, and the lexicon;

identify and distinguish between one or more additional datasets based on the generated error scores for the acoustic model, the language model, and the lexicon, the one or more additional datasets configured to be applied to the automatic speech recognition system to facilitate an improvement in the accuracy of the automatic speech recognition system by correcting the one or more types of errors identified in the initial output;

prompt a user to obtain the one or more additional datasets; and

apply the one or more additional datasets to the automatic speech recognition system to generate a refined output having an improved error score corresponding to at least one of: the acoustic model, the language model, or the lexicon.

13 . The computing system of claim 12 , wherein the initial output comprises a first output from the acoustic model, wherein the first output comprises a plurality of speech utterances recognized from the audio data, a second output from the language model, and a third output from the lexicon, wherein the first output is applied as input to the language model and the lexicon to obtain the second output and the third output.

14 . The computing system of claim 13 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

identify a set of low recall words included in the second output;

determine that one or more low recall words of the set of low recall words included in the second output are out-of-grammar with respect to the language model; and

present the one or more low recall words to the user, such that the one or more additional datasets correspond to the one or more low recall words that are determined to be out-of-grammar.

15 . The computing system of claim 13 , the one or more computer-executable instructions being further executable by the one or more processors to further configure the computing system to:

identify a low recall word included in the second output;

determine that the low recall word included in the second output is not out-of-grammar with respect to the language model;

determine whether the low recall word is associated with a substitution error; and

upon determining that the low recall word is associated with a substitution error, determine that whether a pronunciation similarity exists between the low recall word included in the plurality of speech utterances recognized by the acoustic model and the low recall word included in a transcription of the plurality of speech utterances.