Use of ASR confidence to improve reliability of automatic audio redaction
A speech redaction engine includes a natural language processing (NLP)-based content redaction module receives an automatic speech recognition (ASR) decoding of a decoded portion of said digitized speech signal and utilizes NLP techniques to determine whether it contains sensitive information that should be redacted, and an ASR confidence-based redaction module that receives a confidence indicator and utilizes said confidence indicator to determine, independent of said NLP-based content redaction module, whether said decoded portion contains one or more word(s) that were recognized with a confidence level that is below a threshold. The speech redaction engine includes means for redacting said decoded portion if the NLP-based content redaction module determines that said portion should be redacted, and means for redacting the one or more word(s) if the ASR confidence-based redaction module determines that the one or more word(s) have the confidence level that is below the threshold.
1. A method comprising:
receiving an automatic speech recognition (ASR) decoding of a decoded portion of a digitized speech signal;
determine determining whether said decoded portion of said digitized speech signal contains sensitive information that should be redacted by utilizing a national language processing (NLP);
receiving a confidence score and utilizing said confidence score to determine, independent of said NLP-based content redaction determination, whether said decoded portion of said digitized speech signal contains one or more word(s) within said digitized speech signal that were recognized with a confidence level that is below a threshold;
redacting said decoded portion of said digitized speech signal if the NLP-based content redaction determination determines that said portion should be redacted; and
redacting the one or more word(s) within said digitized speech signal if the ASR confidence-based redaction module determines that the one or more word(s) have the confidence level that is below the threshold.
2. The method of claim 1 further comprising means for encrypting and storing an encrypted version of said digitized speech signal prior to redaction.
3. The method of claim 1 further comprising storing a redacted version of said digitized speech signal and an encrypted version of the decoded portion of said digitized speech signal that is unredacted.
4. The method of claim 1 , wherein the ASR confidence score is derived, at least in part, from normalized likelihood scores.
5. The method of claim 1 , wherein the ASR confidence score is computed, at least in part, using an N-best homogeneity analysis.
6. The method of claim 1 , wherein the ASR confidence score is computed, at least in part, based an acoustic stability analysis.
7. The method of claim 1 , wherein the ASR confidence score is computed, at least in part, based on a word graph hypothesis density analysis.
8. The method of claim 1 , wherein the ASR confidence score is derived, at least in part, based on state, phoneme, or word durations.
9. The method of claim 1 , wherein the ASR confidence score is derived, at least in part, from language model (LM) scores or LM back-off behaviors.
10. The method of claim 1 , wherein the ASR confidence score is computed, at least in part, using a posterior probability analysis.
11. The method of claim 1 , wherein the ASR confidence score is computed, at least in part, using a log-likelihood-ratio analysis.
12. The method of claim 1 , wherein the ASR confidence score is computed, at least in part, using a neural net that includes word identity and aggregated words as predictors.
13. The method of claim 1 , wherein identifying whether the decoded portion contains sensitive information is based, at least, on Personally Identifiable Information (PII).
14. The method of claim 1 , wherein identifying whether the decoded portion contains sensitive information is based, at least, on Nonpublic Personal information (NPI).
15. The method of claim 1 , wherein identifying whether the decoded portion contains sensitive information is based, at least, on Personal Health Information (PHI).
16. The method of claim 1 , wherein identifying whether the decoded portion contains sensitive information is based, at least, on Sensitive Personal Information (SPI) or Personal Credit Information (PCI).
17. The method of claim 1 further comprising de-identifying a voice of a speaker in the digitized speech signal.
18. The method of claim 1 further comprising normalizing accent associated with the digitized speech signal.
19. The method of claim 1 further comprising de-identifying the redacted digitized speech signal.
20. The method of claim 1 , wherein said digitized speech signal is determined to contain sensitive information based on a redaction strictness factor that, depending upon its setting, affects the likelihood of a given ASR decoding being identified as containing sensitive information.