Automatic speech recognition correction

1. A method comprising:

receiving a transcription of speech created using an automatic speech recognition system;

identifying one or more domain-specific contexts associated with the speech;

detecting a text span from the transcription that includes a mistranscribed entry of the speech based on a user-specific context included in the one or more domain-specific contexts associated with the speech;

extracting features from the mistranscribed entry by applying both a user-specific extractor associated with the user-specific context and a domain-specific sequence to sequence grapheme-to-phoneme model, wherein the features include one or more user-specific contextual features of the mistranscribed entry associated with the user-specific context and one or more phonetic features extracted through application of the sequence to sequence grapheme-to-phoneme model;

matching the extracted features against an index of domain-specific entries to identify a correct entry of the mistranscribed entry, based at least in part on both a user-specific contextual feature of the mistranscribed entry and a phonetic feature; and

correcting the transcription by replacing the mistranscribed entry with the correct entry.

2. The method of claim 1 , wherein the text span is detected by applying one or more domain-specific extractors trained based on text spans in the one or more domain-specific contexts and corresponding entries of the text spans in the one or more domain-specific contexts to the transcription.

3. The method of claim 2 , further comprising applying the one or more domain-specific extractors to the text span to identify the mistranscribed entry of the speech from the text span.

4. The method of claim 1 , wherein the mistranscribed entry is matched to the correct entry based on the one or more phonetic features of the mistranscribed entry and phonetic features of the correct entry using term frequency-inverse document frequency-based information retrieval.

5. The method of claim 1 , wherein the domain-specific grapheme-to-phoneme model is trained using phonetic features of the correct entry.

6. The method of claim 1 , further comprising:

identifying textual features of the mistranscribed entry, as part of extracting the features from the mistranscribed entry, based on the one or more domain-specific contexts by applying a domain-specific textual model to the mistranscribed entry; and

matching the mistranscribed entry to the correct entry of the mistranscribed entry based on similarities between the textual features of the mistranscribed entry and textual features of the correct entry in the index of domain-specific entries.

7. The method of claim 6 , wherein the textual features include one or a combination of character n-grams, word n-grams, edge n-grams, and normalized keywords identified based on the one or more domain-specific contexts by applying the domain-specific textual model to the mistranscribed entry.

8. The method of claim 6 , wherein the mistranscribed entry is matched to the correct entry based on the textual features of the mistranscribed entry and the textual features of the correct entry using term frequency-inverse document frequency-based information retrieval.

9. The method of claim 1 , further comprising:

identifying contextual features of the mistranscribed entry, as part of extracting the features from the mistranscribed entry, based on the one or more domain-specific contexts by applying a domain-specific contextual model to the mistranscribed entry; and

matching the mistranscribed entry to the correct entry based on similarities between the contextual features of the mistranscribed entry and contextual features of the correct entry in the index of domain-specific entries.

10. The method of claim 9 , wherein the contextual features of the mistranscribed entry include contextual features of a speaker who uttered the text span of the speech that includes the mistranscribed entry.

11. The method of claim 9 , wherein the mistranscribed entry is matched to the correct entry based on the contextual features of the mistranscribed entry and the contextual features of the correct entry using term frequency-inverse document frequency-based information retrieval.

12. The method of claim 1 , wherein the index of domain-specific entries is selected based on the one or more domain-specific contexts associated with the speech.

13. A system comprising:

one or more processors; and

at least one computer-readable storage medium having stored therein instructions which, when executed by the one or more processors, cause the one or more processors to perform operations comprising:

receiving a transcription of speech created using an automatic speech recognition system;

identifying one or more domain-specific contexts associated with the speech;

applying one or more domain-specific extractors to the transcription to detect a text span from the transcription that includes a mistranscribed entry of the speech based on a user-specific context included in the one or more domain-specific contexts associated with the speech;

extracting features from the mistranscribed entry by applying a user-specific extractor associated with the user-specific context and a domain-specific sequence to sequence grapheme-to-phoneme model, wherein the features include one or more user-specific contextual features of the mistranscribed entry associated with the user-specific context and one or more phonetic features extracted through application of the sequence to sequence grapheme-to-phoneme model;

matching the extracted features against an index of domain-specific entries to identify a correct entry of the mistranscribed entry, based at least in part on both a user-specific contextual feature of the mistranscribed entry and a phonetic feature; and

correcting the transcription by replacing the mistranscribed entry with the correct entry.

14. The system of claim 13 , wherein the mistranscribed entry is matched to the correct entry based on the one or more phonetic features of the mistranscribed entry and phonetic features of the correct entry using term frequency-inverse document frequency-based information retrieval.

15. The system of claim 13 , wherein the instructions which, when executed by the one or more processors, further cause the one or more processors to perform operations comprising:

identifying textual features of the mistranscribed entry, as part of extracting the features from the mistranscribed entry, based on the one or more domain-specific contexts by applying a domain-specific textual model to the mistranscribed entry; and

matching the mistranscribed entry to the correct entry of the mistranscribed entry based on similarities between the textual features of the mistranscribed entry and textual features of the correct entry in the index of domain-specific entries.

16. The system of claim 13 , wherein the instructions which, when executed by the one or more processors, further cause the one or more processors to perform operations comprising:

identifying contextual features of the mistranscribed entry, as part of extracting the features from the mistranscribed entry, based on the one or more domain-specific contexts by applying a domain-specific contextual model to the mistranscribed entry; and

matching the mistranscribed entry to the correct entry based on similarities between the contextual features of the mistranscribed entry and contextual features of the correct entry in the index of domain-specific entries.

17. The system of claim 13 , wherein the index of domain-specific entries is selected based on the one or more domain-specific contexts associated with the speech.

18. A non-transitory computer-readable storage medium having stored therein instructions which, when executed by a processor, cause the processor to perform operations comprising:

receiving a transcription of speech created using an automatic speech recognition system;

identifying one or more domain-specific contexts associated with the speech;

detecting a text span from the transcription that includes a mistranscribed entry of the speech based on a user-specific context included in the one or more domain-specific contexts associated with the speech;

extracting features from the mistranscribed entry by applying a user-specific extractor associated with the user-specific context and a domain-specific sequence to sequence grapheme-to-phoneme model, wherein the features include one or more user-specific contextual features of the mistranscribed entry associated with the user-specific context and one or more phonetic features extracted through application of the sequence to sequence grapheme-to-phoneme model;

matching the extracted features against an index of domain-specific entries to identify a correct entry of the mistranscribed entry using term frequency-inverse document frequency-based information retrieval and based at least in part on both a user-specific contextual feature of the mistranscribed entry and a phonetic feature; and

correcting the transcription by replacing the mistranscribed entry with the correct entry.