Signal processor for signal enhancement and associated methods
A signal processor for performing signal enhancement, the signal processor comprising: an input-terminal, configured to receive an input-signaling; an output-terminal; an interference-cancellation-block configured to receive the input-signaling and to provide an interference-estimate-signaling and an interference-cancelled-signal based on the input-signaling. The signal processor further comprises a feature-block configured to provide a combination-feature-signal based on the interference-cancelled-signal and the interference-estimate-signaling; and a neural-network-block configured to apply model parameters to the combination-feature-signal to provide a neural-network-output-signal to the output-terminal.
1. A speech signal processor for performing signal enhancement, the signal processor comprising:
an input-terminal, configured to receive an input-signaling;
an output-terminal;
an interference-cancellation-block configured to receive the input-signaling and to provide an interference-estimate-signaling and an interference-cancelled-signal based on the input-signaling;
a feature-block configured to provide a combination-feature-signal based on the interference-cancelled-signal and the interference-estimate-signaling;
a neural-network-block configured to apply model parameters to the combination-feature-signal to provide a neural-network-output-signal to the output-terminal;
a synthesis block configured to receive the neural-network-output-signal and the interference-cancelled-signal and to provide a reduced-interference-signal to the output-terminal based on magnitudes of the neural-network-output-signal and on phases of the interference-cancelled-signal.
2. The signal processor of claim 1 , wherein the feature-block is configured to determine the combination-feature-signal by concatenating:
a noise-signal representative of the interference-estimate-signaling; and
a clean-signal representative of the interference-cancelled-signal.
3. The signal processor of claim 2 , wherein:
the noise-signal comprises spectral coefficients of the interference-estimate-signaling; and
the clean-signal comprises spectral coefficients of the interference-cancelled-signal.
4. The signal processor of claim 1 , wherein the neural-network-output-signal comprises an estimated-clean-signal or may represent a gain function or underlying clean signal features.
5. The signal processor of claim 1 , wherein the interference-cancellation-block comprises an echo cancellation block or a beamformer.
6. The signal processor of claim 1 , wherein the interference-estimate-signaling and the interference-cancelled-signal are both time-domain signals.
7. The signal processor of claim 1 , wherein the input-signaling comprises a signal-frame, and the feature-block is configured to determine the combination-feature-signal for the input-signaling based on information representative of:
one or more earlier-signal-frames; and
one or more later-signal-frames.
8. The signal processor of claim 1 , further comprising a second-input-terminal configured to receive a second-input-signal, wherein the interference-cancellation-block is configured to provide both the interference-estimate-signaling and the interference-cancelled-signal based on both the input-signaling and the second-input-signal.
9. The signal processor of claim 1 , further comprising a second-input-terminal configured to receive a second-input-signal, wherein:
the interference-cancellation-block is configured to determine a second-interference-estimate-signaling and a second-interference-cancelled-signal based on the second-input-signal; and
the feature-block is configured to provide the combination-feature-signal by combining signaling representative of:
the interference-estimate-signaling;
the interference-cancelled-signal;
the second-interference-estimate-signaling; and
the second-interference-cancelled-signal.
10. The signal processor of claim 1 , wherein the model parameters are pre-determined based on a comparison of a plurality of signal pairs, each signal pair comprising:
an input-feature-signal, comprising training-interference-estimate-signaling and a training-interference-cancelled-signal; and
an output-feature-signal representative of a desired-neural-network-output-signal.
11. The signal processor of claim 10 , wherein the plurality of signal pairs comprise signal pairs representative of one or more of:
speech signals provided by one or more talkers;
one or more types of interference signals;
one or more ratios between clean-signals and interference-signals;
one or more electronic audio devices; and
one or more acoustic environments.
12. The signal processor of claim 1 , wherein the input-signaling comprises a speech signal, and the signal processor further comprises a speech-block configured to provide a desired speech-feature-signal based on the neural-network-output-signal.
13. The signal processor of claim 1 , wherein the neural-network-block comprises one or more of:
a feedforward neural network;
a recurrent neural network;
a long short-time memory neural network; and
a convolutional neural network.
14. The signal processor of claim 1 , further comprising a feature pre-processing block that is configured to apply normalization per feature dimension of the combination-feature-signal.
15. The signal processor of claim 1 , wherein the input-signaling comprises a signal-frame, and the feature-block is configured to determine the combination-feature-signal for the input-signaling based on information representative of:
one or more earlier-signal-frames.
16. The signal processor of claim 1 , wherein the input-signaling comprises a signal-frame, and the feature-block is configured to determine the combination-feature-signal for the input-signaling based on information representative of:
one or more later-signal-frames.