Robust start-end point detection algorithm using neural network

1. A method comprising:

receiving an audio input stream including a plurality of audio frames;

extracting from each audio frame features representative of the audio frame;

detecting an end point in the extracted features using a first neural network;

providing an indication to a second neural network when an end point is detected; and

detecting a start point in the extracted features from audio frames preceding the detected end point in the audio input stream using the second neural network.

2. The method of claim 1 , wherein extracting from each audio frame features representative of the audio frame includes generating Mel-frequency cepstral coefficients (MFCCs) for each frame of the audio input stream.

3. The method of claim 1 , further comprising training the first neural network for end point detection.

4. The method of claim 1 , further comprising training the second neural network for start point detection.

5. The method of claim 1 , further comprising receiving, at a computing device, a stream of training data including a plurality of input samples having segmented labeled data;

computing, by the first neural network, a network output for each input sample in a forward pass through the training data; and

updating, by the first neural network, weights and biases through a backward pass through the first neural network, including determining whether an input frame is in a Region of Target (ROT).

6. The method of claim 5 , further comprising training the first neural network, weights and biases based on a cross-entropy cost function for non-ROT frames and a One-Spike Connectionist Temporal Classification (OSCTC) cost function for ROT frames;

wherein the first neural network is trained using the OSCTC cost function to cause the neural network to spike one time during the ROT.

7. The method of claim 6 , wherein updating, by the first neural network, weights and biases through the backward pass further comprises computing a signal error for all output nodes using the cross-entropy cost function for non-ROT regions.

8. The method of claim 5 , further comprising:

calculating a signal output error for each network output based on a determination of whether the input frame is in the ROT; and

updating the weights and biases during a backward pass based on the calculated signal output error.

9. The method of claim 8 , wherein updating the weights and biases during the backward pass further comprises applying a cross-entropy cost function if the input frame is not in the ROT and applying OSCTC cost function if the input frame is in the ROT.

10. The method of claim 1 , wherein the first neural network is trained for keyword and end point detection, and wherein the first neural network is trained to generate a spike when the keyword is detected.

11. A system comprising:

a feature extractor configured to receive an audio input stream comprising a sequence of frames and generate corresponding feature data for each frame;

an end detector comprising a first neural network trained to generate a detection probability for an end point of an acoustic event, the end detector configured to receive the feature data and detect the end point of the acoustic event when the detected probability is greater than a pre-determined end-point threshold;

a start detector configured to receive an indication of the detected end point and process the feature data associated with corresponding input frames to detect a start point of the acoustic event;

wherein the frames in the audio input stream from the start point to the end point comprise a region of interest of the acoustic event; and

wherein the first neural network is trained using a backward pass process that comprises determining whether an input frame is in a Region of Target (ROT), and applying a One-Spike Connectionist Temporal Classification (OSCTC) cost function for ROT frames.

12. The system of claim 11 , wherein the feature extractor is configured to compute Mel-frequency cepstral coefficients (MFCCs) for each frame of the audio input stream.

13. The system of claim 11 , further comprising a buffer configured to receive the feature and provide the corresponding feature data associated with corresponding input frames to the start detector.

14. The system of claim 11 , wherein the start detector comprises a second neural network trained to generate a detection probability for the start point of the acoustic event; and

wherein the start point is detected when the detection probability is greater than a pre-determined start point threshold.

15. A system comprising:

a feature extractor configured to receive an audio input stream comprising a sequence of frames and generate corresponding feature data for each frame;

an end detector configured to receive the feature data and detect an end point of an acoustic event; and

a start detector configured to receive an indication of the detected end point and process the feature data associated with corresponding input frames to detect a start point of the acoustic event;

wherein the frames in the audio input stream from the start point to the end point comprise a region of interest of the acoustic event;

wherein the start detector comprises a first neural network trained to generate a detection probability for the start point of the acoustic event;

wherein the start point is detected when the detection probability is greater than a pre-determined start point threshold;

wherein the end detected comprises a second neural network trained to generate a detection probability for the end point of the acoustic event; and

wherein the end point is detected when the detection probability is greater than a pre-determined end point threshold.

16. The system of claim 15 , wherein the second neural network is trained through a process comprising:

receiving, at a computing device, a stream of training data including a plurality of input samples having segmented labeled data;

computing, by the second neural network, a network output for each input sample in a forward pass through the training data; and

updating, by the second neural network, weights and biases through a backward pass through the first neural network, including determining whether an input frame is in a Region of Target (ROT).

17. The system of claim 16 , wherein the weights and biases are updated based on a cross-entropy cost function for non-ROT frames and a One-Spike Connectionist Temporal Classification (OSCTC) cost function for ROT frames;

wherein the first neural network is trained using the OSCTC cost function to cause the neural network to spike one time during the ROT.

18. The system of claim 17 , wherein the weights and biases are updated through the backward pass by computing a signal error for all output nodes using the cross-entropy cost function for non-ROT regions.

19. The system of claim 15 , wherein the end detector comprises a neural network trained to generate a detection probability for the end point of the acoustic event; and

wherein the end point is detected when the detection probability is greater than a pre-determined end point threshold.

20. The system of claim 15 , wherein the start detector comprises a first neural network trained to generate a detection probability for the start point of the acoustic event; and

wherein the start point is detected when the detection probability is greater than a pre-determined start point threshold.