Methods and apparatus for ASR with embedded noise reduction
Methods and an apparatus for performing feature extraction on speech in a microphone signal with embedded noise processing to reduce the amount of processing are provided. In embodiments, feature extraction and the noise estimate use an output of the same Fourier Transform, such that the noise filtering of the speech is embedded with the feature extraction of the speech.
1. A method, comprising:
receiving a microphone signal;
determining whether a noise level of the microphone signal exceeds a noise threshold;
selecting between a first signal processing procedure and a second signal processing procedure, different from the first signal processing procedure, based on whether the noise level exceeds the threshold;
performing the first signal processing procedure when the noise level of the microphone signal exceeds the noise threshold, the first signal processing procedure including:
determining whether the microphone signal contains speech;
determining a noise estimate for the microphone signal when speech is found not to be present;
performing noise filtering using the noise estimate on the microphone signal when speech is found to be present; and
performing feature extraction on the microphone signal when speech is found to be present, such that the noise filtering of the speech is embedded with the feature extraction of the speech; and
performing the second signal processing procedure when the noise level of the microphone signal does not exceed the noise threshold, the second signal processing procedure including performing feature extraction on the microphone signal.
2. The method according to claim 1 , wherein the noise estimate and the noise filtering are not performed on a same frame of the microphone signal.
3. The method according to claim 1 , further including determining whether the microphone signal contains a wake-up phrase after the noise filtering.
4. The method according to claim 1 , wherein the feature extraction is performed while a device containing the microphone is in a sleep state.
5. The method according to claim 1 , wherein the feature extraction includes the use of mel-frequency cepstral coefficients (MFCCs).
6. The method according to claim 1 , further including using a main processor and a lower power processor to provide processing of a wakeup phrase for a device.
7. The method according to claim 6 , wherein the lower power processor performs the feature extraction and noise processing to identify the wakeup phrase.
8. A system comprising:
a processor and memory configured to:
determine whether a noise level of a microphone signal exceeds a noise threshold;
select between a first signal processing procedure and a second signal processing procedure, different from the first signal processing procedure, based on whether the noise level exceeds the threshold;
perform the first signal processing procedure when the noise level of the microphone signal exceeds the noise threshold, the first signal processing procedure including:
determining whether the microphone signal contains speech;
determining a noise estimate for the microphone signal when speech is found not to be present;
performing noise filtering using the noise estimate on the microphone signal when speech is found to be present; and
performing feature extraction on the microphone signal when speech is found to be present, such that the noise filtering of the speech is embedded with the feature extraction of the speech; and
perform the second signal processing procedure when the noise level of the microphone signal does not exceed the noise threshold, the second signal processing procedure including performing feature extraction on the microphone signal.
9. The system according to claim 8 , wherein the noise estimate and the noise filtering are not performed on a same frame of the microphone signal.
10. The system according to claim 8 , further including determining whether the microphone signal contains a wake-up phrase after the noise filtering.
11. The system according to claim 8 , wherein the feature extraction is performed while a device containing the microphone is in a sleep state.
12. The system according to claim 8 , wherein the feature extraction includes the use of mel-frequency cepstral coefficients (MFCCs).
13. The system according to claim 8 , further including using a main processor and a lower power processor to provide processing of a wakeup phrase for a device.
14. The system according to claim 13 , wherein the lower power processor performs the feature extraction and noise processing to identify the wakeup phrase.
15. An article, comprising: a non-transitory computer readable medium having stored instructions that enable a machine to:
determine whether a microphone signal has a noise level above a noise threshold;
select between a first signal processing procedure and a second signal processing procedure, different from the first signal processing procedure, based on whether the noise level is above the threshold;
perform the first signal processing procedure when the noise level of the microphone signal is above the noise threshold, the first signal processing procedure including:
determining whether the microphone signal contains speech;
determining a noise estimate for the microphone signal when speech is found not to be present; perform noise filtering using the noise estimate on the microphone signal when speech is found to be present; and
performing feature extraction on the microphone signal when speech is found to be present, such that the noise filtering of the speech is embedded with the feature extraction of the speech; and
perform a second signal processing procedure when the noise level of the microphone signal is not above the noise threshold, the second signal processing procedure including performing feature extraction on the microphone signal.
16. The article according to claim 15 , wherein the noise estimate and the noise filtering are not performed on a same frame of the microphone signal.
17. The article according to claim 15 , further including instructions for determining whether the microphone signal contains a wake-up phrase after the noise filtering.
18. The article according to claim 15 , wherein the feature extraction is performed while a device containing the microphone is in a sleep state.
19. The article according to claim 15 , wherein the feature extraction includes the use of mel-frequency cepstral coefficients (MFCCs).
20. The article according to claim 15 , further including using a main processor and a lower power processor to provide processing of a wakeup phrase for a device.
21. The method of claim 6 , wherein the lower power processor performs denoising and feature extraction for an automatic speech recognition operation performed by the main processor.