System and method using parameterized speech synthesis to train acoustic models
A method for removing private data from an acoustic model includes capturing speech from a large population of users, creating a text-to-speech voice from at least a portion of the large population of users, discarding speech data from a database of speech, creating text-to-speech waveforms from the text-to-speech voice and the new database of speech with the discarded speech data and generating an automatic speech recognition model using the text-to-speech waveforms.
1. A method comprising:
capturing speech from a large population of users;
generating noise data that identifies intermittent noise in the captured speech;
clustering the generated noise data and removing the clustered noise data from the captured speech;
creating a text-to-speech voice from at least a portion of the large population of users;
storing a first database of speech associated with the large population of users, the first database of speech including speech data for the large population of users and private speech data;
discarding the private speech data from the first database of speech to yield a second database of speech, wherein the second database of speech includes the speech data for the large population of users, and wherein the private speech data is associated with at least one protected speaker from the large population of users;
creating text-to-speech waveforms from the text-to-speech voice and the yielded second database of speech; and
generating an automatic speech recognition model using the created text-to-speech waveforms.
2. The method of claim 1 , wherein the capturing of the speech from the large population of users further comprises utilizing parameterized models.
3. The method of claim 2 , wherein the parameterized models utilize one or more of iVectors, D-Vectors or a particular type of embedding.
4. The method of claim 1 , further comprising:
implementing the automatic speech recognition model in a spoken dialogue system.
5. The method of claim 4 , further comprising:
interacting, via a spoken dialogue system implementing the automatic speech recognition model, with a user via speech.
6. The method of claim 1 , further comprising:
generating text for use in creating the text-to-speech voice from at least the portion of the large population of users.
7. The method of claim 6 , wherein the text for use in creating the text-to-speech voice comprises one or more of: (1) modified text from automatic speech recognition texts associated with the portion of the large population of users; and (2) generated text from users and separate from the large population of users.
8. The method of claim 6 , wherein generating the text for use in creating the text-to-speech voice is performed by a speaker parameterized text generator.
9. A system comprising:
at least a processor; and
a computer-readable storage medium having instructions stored which, when executed by at least the processor, cause at least the processor to perform operations comprising:
capturing speech from a large population of users;
generating noise data that identifies intermittent noise in the captured speech;
clustering the generated noise data and removing the clustered noise data from the captured speech;
creating a text-to-speech voice from at least a portion of the large population of users;
storing a first database of speech associated with the large population of users, the first database of speech including speech data for the large population of users and private speech data;
discarding the private speech data from the first database of speech to yield a second database of speech, wherein the second database of speech includes the speech data for the large population of users, and wherein the private speech data is associated with at least one protected speaker from the large population of users;
creating text-to-speech waveforms from the text-to-speech voice and the yielded second database of speech; and
generating an automatic speech recognition model using the created text-to-speech waveforms.
10. The system of claim 9 , wherein the capturing of the speech from the large population of users further comprises utilizing parameterized models.
11. The system of claim 10 , wherein the parameterized models utilize one or more of iVectors, D-Vectors or a particular type of embedding.
12. The system of claim 9 , wherein the computer-readable storage medium stores additional instructions stored which, when executed by the processor, cause the processor to perform operations further comprising implementing the automatic speech recognition model in a spoken dialogue system.
13. The system of claim 12 , wherein the computer-readable storage medium stores additional instructions stored which, when executed by the processor, cause the processor to perform operations further comprising:
interacting, via a spoken dialogue system implementing the automatic speech recognition model, with a user via speech.
14. The system of claim 9 , wherein the computer-readable storage medium stores additional instructions stored which, when executed by the processor, cause the processor to perform operations further comprising:
generating text for use in creating the text-to-speech voice from at least the portion of the large population of users.
15. The system of claim 14 , wherein the text for use in creating the text-to-speech voice comprises one or more of: (1) modified text from automatic speech recognition texts associated with the portion of the large population of users; and (2) generated text from users and separate from the large population of users.
16. The system of claim 14 , wherein generating the text for use in creating the text-to-speech voice is performed by a speaker parameterized text generator.
17. A computer-readable storage device having instructions stored which, when executed by a processor, cause the processor to perform operations comprising:
capturing speech from a large population of users;
generating noise data that identifies intermittent noise in the captured speech;
clustering the generated noise data and removing the clustered noise data from the captured speech;
creating a text-to-speech voice from at least a portion of the large population of users;
storing a first database of speech associated with the large population of users, the first database of speech including speech data for the large population of users and private speech data;
discarding the private speech data from the first database of speech to yield a second database of speech, wherein the second database of speech includes the speech data for the large population of users, and wherein the private speech data is associated with at least one protected speaker from the large population of users;
creating text-to-speech waveforms from the text-to-speech voice and the yielded second database of speech; and
generating an automatic speech recognition model using the created text-to-speech waveforms.
18. The computer-readable storage device of claim 17 , wherein the computer-readable storage device stores additional instructions stored which, when executed by the processor, cause the processor to perform operations further comprising:
generating text for use in creating the text-to-speech voice from at least the portion of the large population of users.