Feature extraction for anonymized speech recognition

摘要

Various of the disclosed embodiments relate to systems and methods for extracting audio information, e.g. a textual description of speech, from a speech recording while retaining the anonymity of the speaker. In certain embodiments, a third party may perform various aspects of the anonymization and speech processing. Certain embodiments facilitate anonymization in compliance with various legislative requirements even when third parties are involved.

主权项

1. A computer-implemented method comprising:
receiving a raw audio waveform from a user device, where the raw audio waveform is recorded by the user device and contains speech data of a user; providing metadata and the raw audio waveform to a third-party processing system,
where the metadata identifies one or more frequency components to be redacted from a frequency representation of the raw audio waveform; directing the third-party processing system to perform an operation on the raw audio waveform, wherein the operation includes:
generating the frequency representation of the raw audio waveform,removing the one or more frequency components from the frequency representation that identify the user to produce a modified frequency representation, andgenerating an anonymized audio waveform from the modified frequency representation,where the third-party processing system is only allowed to perform speech processing on anonymized portions of the raw audio waveform; allowing the third-party processing system to copy elements of the anonymized audio waveform to improve a model employed by the third-party processing system to identify speech within audio waveforms; receiving a textual depiction of speech in the anonymized audio waveform from the third-party processing system; and performing a system operation based on the textual depiction.