Phrase extraction for ASR models
A method of phrase extraction for ASR models includes obtaining audio data characterizing an utterance and a corresponding ground-truth transcription of the utterance and modifying the audio data to obfuscate a particular phrase recited in the utterance. The method also includes processing, using a trained ASR model, the modified audio data to generate a predicted transcription of the utterance, and determining whether the predicted transcription includes the particular phrase by comparing the predicted transcription of the utterance to the ground-truth transcription of the utterance. When the predicted transcription includes the particular phrase, the method includes generating an output indicating that the trained ASR model leaked the particular phrase from a training data set used to train the ASR model.
1. A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:
obtaining audio data characterizing an utterance and a corresponding ground-truth transcription of the utterance;
processing the ground-truth transcription to identify a particular phrase included in the ground-truth transcription that is associated with sensitive data;
modifying the audio data to obfuscate the particular phrase recited in the utterance;
processing, using a trained automated speech recognition (ASR) model, the modified audio data to generate a predicted transcription of the utterance;
determining whether the predicted transcription includes the particular phrase or another phrase substituted for the particular phrase from the ground-truth transcription that is associated with a same category of information as the particular phrase by comparing the predicted transcription of the utterance to the ground-truth transcription of the utterance; and
when the predicted transcription includes the other phrase substituted for the particular phrase, generating an output indicating that the trained ASR model leaked the other phrase from a training data set used to train the ASR model.
2. The computer-implemented method of claim 1 , wherein the operations further comprise,
when the predicted transcription includes the particular phrase, generating an output indicating that the trained ASR model leaked the particular phrase from the training data set used to train the ASR model.
3. The computer-implemented method of claim 2 , wherein the operations further comprise, when the predicted transcription does not include the particular phrase or the other phrase substituted for the particular phrase from the ground-truth transcription that is associated with the same category of information as the particular phrase, generating an output indicating that the trained ASR model has not leaked any information from the training data set used to train the ASR model.
4. The computer-implemented method of claim 1 , wherein the audio data comprises an audio waveform.
5. The computer-implemented method of claim 4 , wherein the audio waveform corresponds to human speech.
6. The computer-implemented method of claim 4 , wherein the audio waveform corresponds to synthesized speech.
7. The computer-implemented method of claim 1 , wherein modifying the audio data comprises:
based on the ground-truth transcription, identifying a segment of the audio data that aligns with the particular phrase in the ground-truth transcription; and
performing data augmentation on the identified segment of the audio data to obfuscate the particular phrase recited in the utterance.
8. The computer-implemented method of claim 7 , wherein performing data augmentation on the identified segment of the audio data comprises adding noise to the identified segment of the audio data.
9. The computer-implemented method of claim 7 , wherein performing data augmentation on the identified segment of the audio data comprises removing the identified segment of the audio data.
10. A system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:
obtaining audio data characterizing an utterance and a corresponding ground-truth transcription of the utterance;
processing the ground-truth transcription to identify a particular phrase included in the ground-truth transcription that is associated with sensitive data;
modifying the audio data to obfuscate the particular phrase recited in the utterance;
processing, using a trained automated speech recognition (ASR) model, the modified audio data to generate a predicted transcription of the utterance;
determining whether the predicted transcription includes the particular phrase or another phrase substituted for the particular phrase from the ground-truth transcription that is associated with a same category of information as the particular phrase by comparing the predicted transcription of the utterance to the ground-truth transcription of the utterance; and
when the predicted transcription includes the other phrase substituted for the particular phrase, generating an output indicating that the trained ASR model leaked the other phrase from a training data set used to train the ASR model.
11. The system of claim 10 , wherein the operations further comprise,
when the predicted transcription includes the particular phrase, generating an output indicating that the trained ASR model leaked the particular phrase from the training data set used to train the ASR model.
12. The system of claim 11 , wherein the operations further comprise, when the predicted transcription does not include the particular phrase or the other phrase substituted for the particular phrase from the ground-truth transcription that is associated with the same category of information as the particular phrase, generating an output indicating that the trained ASR model has not leaked any information from the training data set used to train the ASR model.
13. The system of claim 10 , wherein the audio data comprises an audio waveform.
14. The system of claim 13 , wherein the audio waveform corresponds to human speech.
15. The system of claim 13 , wherein the audio waveform corresponds to synthesized speech.
16. The system of claim 10 , wherein modifying the audio data comprises:
based on the ground-truth transcription, identifying a segment of the audio data that aligns with the particular phrase in the ground-truth transcription; and
performing data augmentation on the identified segment of the audio data to obfuscate the particular phrase recited in the utterance.
17. The system of claim 16 , wherein performing data augmentation on the identified segment of the audio data comprises adding noise to the identified segment of the audio data.
18. The system of claim 16 , wherein performing data augmentation on the identified segment of the audio data comprises removing the identified segment of the audio data.