Abstract: Providing a framework for merging automatic speech recognition (ASR) systems having a shared deep neural network (DNN) feature transformation is provided. A received utterance may be evaluated to generate a DNN-derived feature from the top hidden layer of a DNN. The top hidden layer output may then be utilized to generate a network including a bottleneck layer and an output layer. Weights representing a feature dimension reduction may then be extracted between the top hidden layer and the bottleneck layer. Scores may then be generated and combined to merge the ASR systems which share the DNN feature transformation.

Abstract: The adaptation and personalization of a deep neural network (DNN) model for automatic speech recognition is provided. An utterance which includes speech features for one or more speakers may be received in ASR tasks such as voice search or short message dictation. A decomposition approach may then be applied to an original matrix in the DNN model. In response to applying the decomposition approach, the original matrix may be converted into multiple new matrices which are smaller than the original matrix. A square matrix may then be added to the new matrices. Speaker-specific parameters may then be stored in the square matrix. The DNN model may then be adapted by updating the square matrix. This process may be applied to all of a number of original matrices in the DNN model. The adapted DNN model may include a reduced number of parameters than those received in the original DNN model.

Abstract: A Deep Neural Network (DNN) model used in an Automatic Speech Recognition (ASR) system is restructured. A restructured DNN model may include fewer parameters compared to the original DNN model. The restructured DNN model may include a monophone state output layer in addition to the senone output layer of the original DNN model. Singular value decomposition (SVD) can be applied to one or more weight matrices of the DNN model to reduce the size of the DNN Model. The output layer of the DNN model may be restructured to include monophone states in addition to the senones (tied triphone states) which are included in the original DNN model. When the monophone states are included in the restructured DNN model, the posteriors of monophone states are used to select a small part of senones to be evaluated.

Abstract: A high-dimensional posterior-based feature with partial distance elimination may be utilized for speech recognition. The log likelihood values of a large number of Gaussians are needed to generate the high-dimensional posterior feature. Gaussians with very small log likelihoods are associated with zero posterior values. Log likelihoods for Gaussians for a speech frame may be evaluated with a partial distance elimination method. If the partial distance of a Gaussian is already too small, the Gaussian will have a zero posterior value. The partial distance may be calculated by sequentially adding individual dimensions in a group of dimensions. The partial distance elimination occurs when less than all of the dimensions in the group are sequentially added.

Abstract: Described herein are various technologies pertaining to a multilingual deep neural network (MDNN). The MDNN includes a plurality of hidden layers, wherein values for weight parameters of the plurality of hidden layers are learned during a training phase based upon training data in terms of acoustic raw features for multiple languages. The MDNN further includes softmax layers that are trained for each target language separately, making use of the hidden layer values trained jointly with multiple source languages. The MDNN is adaptable, such that a new softmax layer may be added on top of the existing hidden layers, where the new softmax layer corresponds to a new target language.

Abstract: Technologies pertaining to training a deep neural network (DNN) for use in a recognition system are described herein. The DNN is trained using heterogeneous data, the heterogeneous data including narrowband signals and wideband signals. The DNN, subsequent to being trained, receives an input signal that can be either a wideband signal or narrowband signal. The DNN estimates the class posterior probability of the input signal regardless of whether the input signal is the wideband signal or the narrowband signal.

Abstract: Personalization for Automatic Speech Recognition (ASR) is associated with a particular device. A generalized i-vector clustering method is used to train i-vector parameters on utterances received from a device and to classify test utterances from the same device. A sub-loading matrix and a residual noise term may be used when determining the personalization. A Universal Background Model (UBM) is trained using the utterances. The UBM is applied to obtain i-vectors of training utterances received from a device and a Gaussian Mixture Model (GMM) is trained using the i-vectors. During testing, the i-vector for each utterance received from the device is estimated using the device's UBM. The utterance is then assigned to the cluster with the closest centroid in the GMM. For each utterance, the i-vector and the residual noise estimation is performed. Hyperparameter estimation is also performed. The i-vector estimation and hyperparameter estimation are performed until convergence.

Abstract: A speech recognition system adaptively estimates a warping factor used to reduce speaker variability. The warping factor is estimated using a small window (e.g. 100 ms) of speech. The warping factor is adaptively adjusted as more speech is obtained until the warping factor converges or a pre-defined maximum number of adaptation is reached. The speaker may be placed into a group selected from two or more groups based on characteristics that are associated with the speaker's window of speech. Different step sizes may be used within the different groups when estimating the warping factor. VTLN is applied to the speech input using the estimated warping factor. A linear transformation, including a bias term, may also be computed to assist in normalizing the speech along with the application of the VTLN.

Abstract: Noise and channel distortion parameters in the vectorized logarithmic or the cepstral domain for an utterance may be estimated, and subsequently the distorted speech parameters in the same domain may be updated using an unscented transformation framework during online automatic speech recognition. An utterance, including speech generated from a transmission source for delivery to a receiver, may be received by a computing device. The computing device may execute instructions for applying the unscented transformation framework to speech feature vectors, representative of the speech, in order to estimate, in a sequential or online manner, static noise and channel distortion parameters and dynamic noise distortion parameters in the unscented transformation framework. The static and dynamic parameters for the distorted speech in the utterance may then be updated from clean speech parameters and the noise and channel distortion parameters using non-linear mapping.

Abstract: Subspace speech adaptation may be utilized for facilitating the recognition of speech containing short utterances. Speech training data may be received in a speech model by a computer. A first matrix may be determined for preconditioning speech statistics based on the speech training data. A second matrix may be determined for representing a basis for the speech to be recognized. A set of basis matrices may then be determined from the first matrix and the second matrix. Speech test data including a short utterance may then be received by the computer. The computer may then apply the set of basis matrices to the speech test data to produce a transcription. The transcription may represent speech recognition of the short utterance.

Abstract: Scalar operations for model adaptation or feature enhancement may be utilized for recognizing an utterance during automatic speech recognition in a noisy environment. An utterance including distorted speech generated from a transmission source for delivery to a receiver, may be received by a computer. The distorted speech may be caused by the noisy environment and channel distortion. Computations using scalar operations in the form of an algorithm may then be performed for recognizing the utterance. As a result of performing all of the computations with scalar operations, computational complexity is very small in comparison to matrix and vector operations. Vector Taylor Series with diagonal Jacobian approximation may also be utilized as a distortion-model-based noise robust algorithm with scalar operations.

Abstract: Architecture for correcting incorrect recognition results in an Asian language speech recognition system. A spelling mode can be launched in response to receiving speech input, the spelling mode for correcting incorrect spelling of the recognition results or generating new words. Correction can be obtained using speech and/or manual selection and entry. The architecture facilitates correction in a single pass, rather than multiples times as in conventional systems. Words corrected using the spelling mode are corrected as a unit and treated as a word. The spelling mode applies to languages of at least the Asian continent, such as Simplified Chinese, Traditional Chinese, and/or other Asian languages such as Japanese.

Abstract: Online histogram recognition may be provided. Upon receiving a spoken phrase from a user, a histogram/frequency distribution may be estimated on the spoken phrase according to a prior distribution. The histogram distribution may be equalized and then provided to a spoken language understanding application.

Abstract: Techniques for enhanced automatic speech recognition are described. An enhanced ASR system may be operative to generate an error correction function. The error correction function may represent a mapping between a supervised set of parameters and an unsupervised training set of parameters generated using a same set of acoustic training data, and apply the error correction function to an unsupervised testing set of parameters to form a corrected set of parameters used to perform speaker adaptation. Other embodiments are described and claimed.

Abstract: A speech recognition system includes a receiver component that receives a distorted speech utterance. The speech recognition also includes an adaptor component that selectively adapts parameters of a compressed model used to recognize at least a portion of the distorted speech utterance, wherein the adaptor component selectively adapts the parameters of the compressed model based at least in part upon the received distorted speech utterance.

Abstract: Subspace speech adaptation may be utilized for facilitating the recognition of speech containing short utterances. Speech training data may be received in a speech model by a computer. A first matrix may be determined for preconditioning speech statistics based on the speech training data. A second matrix may be determined for representing a basis for the speech to be recognized. A set of basis matrices may then be determined from the first matrix and the second matrix. Speech test data including a short utterance may then be received by the computer. The computer may then apply the set of basis matrices to the speech test data to produce a transcription. The transcription may represent speech recognition of the short utterance.

Abstract: A speech recognition system described herein includes a receiver component that receives a distorted speech utterance. The speech recognition also includes an updater component that is in communication with a first model and a second model, wherein the updater component automatically updates parameters of the second model based at least in part upon joint estimates of additive and convolutive distortions output by the first model, wherein the joint estimates of additive and convolutive distortions are estimates of distortions based on a phase-sensitive model in the speech utterance received by the receiver component. Further, distortions other than additive and convolutive distortions, including other stationary and nonstationary sources, can also be estimated used to update the parameters of the second model.

Abstract: Noise and channel distortion parameters in the vectorized logarithmic or the cepstral domain for an utterance may be estimated, and subsequently the distorted speech parameters in the same domain may be updated using an unscented transformation framework during online automatic speech recognition. An utterance, including speech generated from a transmission source for delivery to a receiver, may be received by a computing device. The computing device may execute instructions for applying the unscented transformation framework to speech feature vectors, representative of the speech, in order to estimate, in a sequential or online manner, static noise and channel distortion parameters and dynamic noise distortion parameters in the unscented transformation framework. The static and dynamic parameters for the distorted speech in the utterance may then be updated from clean speech parameters and the noise and channel distortion parameters using non-linear mapping.

Abstract: The language of origin of a word is determined by analyzing non-uniform letter sequence portions of the word.

Abstract: Described is noise reduction technology generally for speech input in which a noise-suppression related gain value for the frame is determined based upon a noise level associated with that frame in addition to the signal to noise ratios (SNRs). In one implementation, a noise reduction mechanism is based upon minimum mean square error, Mel-frequency cepstra noise reduction technology. A high gain value (e.g., one) is set to accomplish little or no noise suppression when the noise level is below a threshold low level, and a low gain value set or computed to accomplish large noise suppression above a threshold high noise level. A noise-power dependent function, e.g., a log-linear interpolation, is used to compute the gain between the thresholds. Smoothing may be performed by modifying the gain value based upon a prior frame's gain value. Also described is learning parameters used in noise reduction via a step-adaptive discriminative learning algorithm.