ASAPP, INC.

A speech processing model may be trained using pseudo tokens. Training a speech processing model with pseudo tokens may allow for training with a smaller amount of labeled training data and accordingly lower costs. A set of pseudo tokens may be determined by computing feature vectors from unlabeled training data, clustering the feature vectors, and performing token compression using the clustered feature vectors. A first speech processing model may be trained using unlabeled training data by determining sequences of pseudo tokens corresponding to the unlabeled training data. A second speech processing model may be initialized using the first speech processing model and then trained using labeled training data. The second speech processing model may then be deployed to a speech processing application.

For any application that processes speech, improving the quality of the feature vectors may improve the quality of the speech application. The quality of feature vectors may be improved by modifying a neural network architecture for computing feature vectors to allocate computational resources where they are more effective for learning and computing the feature vectors. Contextual feature vectors may be computed from feature vectors by using a parameterized downsampling operation that decreases a vector sequence rate, processing the downsampled vectors with a neural network, and using a parameterized upsampling operation that increases a vector sequence rate. For example, parameterized downsampling may decrease a vector sequence rate by a factor of two, a neural may require fewer computational resources since it operates with a lower vector sequence rate, and parameterized upsampling may then increase the vector sequence rate by a factor of two.

Audio signals of speech may be processed using an acoustic model. An acoustic model may be implemented with multiple streams of processing where different streams perform processing using different dilation rates. For example, a first stream may process features of the audio signal with one or more convolutional neural network layers having a first dilation rate, and a second stream may process features of the audio signal with one or more convolutional neural network layers having a second dilation rate. Each stream may compute a stream vector, and the stream vectors may be combined to a vector of speech unit scores, where the vector of speech unit scores provides information about the acoustic content of the audio signal. The vector of speech unit scores may be used for any appropriate application of speech, such as automatic speech recognition.

A speech processing model may be trained using pseudo tokens. Training a speech processing model with pseudo tokens may allow for training with a smaller amount of labeled training data and accordingly lower costs. A set of pseudo tokens may be determined by computing feature vectors from unlabeled training data, clustering the feature vectors, and performing token compression using the clustered feature vectors. A first speech processing model may be trained using unlabeled training data by determining sequences of pseudo tokens corresponding to the unlabeled training data. A second speech processing model may be initialized using the first speech processing model and then trained using labeled training data. The second speech processing model may then be deployed to a speech processing application.

For any application that processes speech, improving the quality of the feature vectors may improve the quality of the speech application. The quality of feature vectors may be improved by modifying a neural network architecture for computing feature vectors to allocate computational resources where they are more effective for learning and computing the feature vectors. Contextual feature vectors may be computed from feature vectors by using a parameterized downsampling operation that decreases a vector sequence rate, processing the downsampled vectors with a neural network, and using a parameterized upsampling operation that increases a vector sequence rate. For example, parameterized downsampling may decrease a vector sequence rate by a factor of two, a neural may require fewer computational resources since it operates with a lower vector sequence rate, and parameterized upsampling may then increase the vector sequence rate by a factor of two.

The present disclosure relates to a system, method, and computer program for creating a complete transcription of an audio recording from separately transcribed redacted and unredacted words. The system receives an original audio recording and redacts a plurality of words from the original audio recording to obtain a modified audio recording. The modified audio recording is outputted to a first transcription service. Audio clips of the redacted words from the original audio recording are extracted using word-level timestamps for the redacted words. The extracted audio clips are outputted to a second transcription service. The system receives a transcription of the modified audio recording from the first transcription service and transcriptions of the extracted audio clips from the second transcription service. The transcriptions of the extracted audio clips are combined with the transcription of the modified audio recording to obtain a complete transcription of the original audio recording using word-level timestamps.

Audio signals of speech may be processed using an acoustic model. An acoustic model may be implemented with multiple streams of processing where different streams perform processing using different dilation rates. For example, a first stream may process features of the audio signal with one or more convolutional neural network layers having a first dilation rate, and a second stream may process features of the audio signal with one or more convolutional neural network layers having a second dilation rate. Each stream may compute a stream vector, and the stream vectors may be combined to a vector of speech unit scores, where the vector of speech unit scores provides information about the acoustic content of the audio signal. The vector of speech unit scores may be used for any appropriate application of speech, such as automatic speech recognition.