Intelligent noise suppression for audio signals within a communication platform
Methods and systems provide users of a communication platform with intelligent, real-time noise suppression for audio signals broadcasted in a communication session. The system receives an input audio signal from an audio capture device; processes the input audio signal to provide a second version of the audio signal with noise suppression based on DSP techniques; transmits the second version of the audio signal to a communication platform for real-time streaming; classifies, via a machine learning algorithm, whether the second version of the audio signal contains noise beyond a noise threshold; based on a classification that the second version of the audio signal contains noise beyond the noise threshold, processes the second version of the audio signal to provide a third version of the audio signal with noise suppression based on AI techniques; and transmits the third version of the audio signal to the communication platform.
1 . A method for providing intelligent noise suppression for an audio signal within a communication platform, comprising:
classifying, using a first machine learning algorithm, whether the audio signal contains noise beyond a noise threshold based on a binary classification, comprising:
generating a binary output comprising the classification after a predefined time interval has expired; and
storing the binary output within a buffer, the buffer corresponding to a predefined window of time and configured to store a predefined number of binary outputs;
determining that the binary outputs within the buffer correspond to a likelihood that the audio signal contains noise beyond the noise threshold using a post-classification smoothing technique; and
processing the audio signal via a second machine learning algorithm trained to suppress noise to provide another version of the audio signal.
2 . The method of claim 1 , wherein the audio signal provides suppression of stationary noises, and wherein the another version of the audio signal provides suppression of both stationary and non-stationary noises.
3 . The method of claim 1 , wherein classifying, using the first machine learning algorithm, whether the audio signal contains noise beyond the noise threshold based on the binary classification comprises:
extracting a plurality of audio features from the audio signal, wherein the audio signal is a raw waveform;
transmitting the audio features to a neural network; and
analyzing the audio features via the neural network to provide a probability of whether the audio signal contains noise beyond the noise threshold.
4 . The method of claim 3 , further comprising:
generating a spectrogram based on the extracted audio features, wherein transmitting the audio features to the neural network comprises transmitting the spectrogram to the neural network, and
wherein analyzing the audio features via the neural network comprises analyzing the spectrogram.
5 . The method of claim 1 , wherein classifying whether the audio signal contains noise beyond the noise threshold further comprises performing one or more feature-based classification techniques.
6 . The method of claim 5 , wherein classifying whether the audio signal contains noise beyond the noise threshold further comprises:
determining that the predefined window of time has expired;
generating a confidence score for the predefined number of binary outputs stored within the buffer; and
based on the confidence score, determining whether the audio signal meets or exceeds the noise threshold.
7 . A non-transitory computer-readable medium containing instructions for providing intelligent noise suppression for an audio signal within a communication platform, comprising:
classifying, using a first machine learning algorithm, whether the audio signal contains noise beyond a noise threshold based on a binary classification, comprising:
generating a binary output comprising the classification after a predefined time interval has expired; and
storing the binary output within a buffer, the buffer corresponding to a predefined window of time and configured to store a predefined number of binary outputs;
determining that the binary outputs within the buffer correspond to a likelihood that the audio signal contains noise beyond the noise threshold using a post-classification smoothing technique; and
processing the audio signal via a second machine learning algorithm trained to suppress noise to provide another version of the audio signal.
8 . The non-transitory computer-readable medium of claim 7 , wherein the audio signal provides suppression of stationary noises, and wherein the another version of the audio signal provides suppression of both stationary and non-stationary noises.
9 . The non-transitory computer-readable medium of claim 7 , wherein classifying, using the first machine learning algorithm, whether the audio signal contains noise beyond the noise threshold based on the binary classification comprises:
extracting a plurality of audio features from the audio signal, wherein the audio signal is a raw waveform;
transmitting the audio features to a neural network; and
analyzing the audio features via the neural network to provide a probability of whether the audio signal contains noise beyond the noise threshold.
10 . The non-transitory computer-readable medium of claim 9 , further comprising instructions of:
generating a spectrogram based on the extracted audio features, wherein transmitting the audio features to the neural network comprises transmitting the spectrogram to the neural network; and
wherein analyzing the audio features via the neural network comprises analyzing the spectrogram.
11 . The non-transitory computer-readable medium of claim 9 , wherein classifying whether the audio signal contains noise beyond the noise threshold further comprises performing one or more feature-based classification techniques.
12 . The non-transitory computer-readable medium of claim 11 , wherein classifying whether the audio signal contains noise beyond the noise threshold further comprises:
determining that the predefined window of time has expired;
generating a confidence score for the plurality of predefined number of binary outputs stored within the buffer; and
based on the confidence score, determining whether the audio signal meets or exceeds the noise threshold.
13 . A communication system comprising one or more processors configured to perform the operations of:
classifying, using a first machine learning algorithm, whether an audio signal contains noise beyond a noise threshold based on a binary classification, comprising:
generating a binary output comprising the classification after a predefined time interval has expired; and
storing the binary output within a buffer, the buffer corresponding to a predefined window of time and configured to store a predefined number of binary outputs;
determining that the binary outputs within the buffer correspond to a likelihood that the audio signal contains noise beyond the noise threshold using a post-classification smoothing technique; and
processing the audio signal via a second machine learning algorithm trained to suppress noise to provide another version of the audio signal.
14 . The communication system of claim 13 , wherein the audio signal provides suppression of stationary noises, and wherein the another version of the audio signal provides suppression of both stationary and non-stationary noises.
15 . The communication system of claim 13 , wherein classifying, using the first machine learning algorithm, whether the audio signal contains noise beyond the noise threshold based on the binary classification comprises:
extracting a plurality of audio features from the audio signal, wherein the audio signal is a raw waveform;
transmitting the audio features to a neural network; and
analyzing the audio features via the neural network to provide a probability of whether the audio signal contains noise beyond the noise threshold.
16 . The communication system of claim 15 , further comprising the operations of:
generating a spectrogram based on the extracted audio features, wherein transmitting the audio features to the neural network comprises transmitting the spectrogram to the neural network; and
wherein analyzing the audio features via the neural network comprises analyzing the spectrogram.
17 . The communication system of claim 13 , wherein classifying whether the audio signal contains noise beyond the noise threshold further comprises performing one or more feature-based classification techniques.
18 . The method of claim 1 , wherein the post-classification smoothing technique involves:
determining one or more aggregate characteristics of the predefined number of binary outputs within the buffer, based on a distribution of binary outputs within the buffer;
determining a mistake likelihood for each of the predefined number of binary outputs within the buffer, based on the distribution of binary outputs within the buffer; and
determining that the audio signal contains noise beyond the noise threshold based on the aggregate characteristics and the mistake likelihood.
19 . The non-transitory computer-readable medium of claim 7 , wherein the post-classification smoothing technique involves:
determining one or more aggregate characteristics of the predefined number of binary outputs within the buffer, based on a distribution of binary outputs within the buffer;
determining a mistake likelihood for each of the predefined number of binary outputs within the buffer, based on the distribution of binary outputs within the buffer; and
determining that the audio signal contains noise beyond the noise threshold based on the aggregate characteristics and the mistake likelihood.
20 . The communication system of claim 13 , wherein the post-classification smoothing technique involves:
determining one or more aggregate characteristics of the predefined number of binary outputs within the buffer, based on a distribution of binary outputs within the buffer;
determining a mistake likelihood for each of the predefined number of binary outputs within the buffer, based on the distribution of binary outputs within the buffer; and
determining that the audio signal contains noise beyond the noise threshold based on the aggregate characteristics and the mistake likelihood.