Systems and processes for providing user-specific acoustic models are provided. In accordance with one example, a method includes, at an electronic device having one or more processors, receiving a plurality of speech inputs, each of the speech inputs associated with a same user of the electronic device; providing each of the plurality of speech inputs to a user-independent acoustic model, the user-independent acoustic model providing a plurality of speech results based on the plurality of speech inputs; initiating a user-specific acoustic model on the electronic device; and adjusting the user-specific acoustic model based on the plurality of speech inputs and the plurality of speech results.

Implementations can reduce the time required to obtain responses from an automated assistant by, for example, obviating the need to provide an explicit invocation to the automated assistant, such as by saying a hot-word/phrase or performing a specific user input, prior to speaking a command or query. In addition, the automated assistant can optionally receive, understand, and/or respond to the command or query without communicating with a server, thereby further reducing the time in which a response can be provided. Implementations only selectively initiate on-device speech recognition responsive to determining one or more condition(s) are satisfied. Further, in some implementations, on-device NLU, on-device fulfillment, and/or resulting execution occur only responsive to determining, based on recognized text form the on-device speech recognition, that such further processing should occur. Thus, through selective activation of on-device speech processing, and/or selective activation of on-device NLU and/or on-device fulfillment, various client device resources are conserved.

A method for automatic speech recognition using joint acoustic echo cancellation, speech enhancement, and voice separation includes receiving, at a contextual frontend processing model, input speech features corresponding to a target utterance. The method also includes receiving, at the contextual frontend processing model, at least one of a reference audio signal, a contextual noise signal including noise prior to the target utterance, or a speaker embedding including voice characteristics of a target speaker that spoke the target utterance. The method further includes processing, using the contextual frontend processing model, the input speech features and the at least one of the reference audio signal, the contextual noise signal, or the speaker embedding vector to generate enhanced speech features.

A method includes training a set of teacher models. Training the set of teacher models includes, for each individual teacher model of the set of teacher models, training the individual teacher model to transcribe unlabeled audio samples and predict a pseudo labeled dataset having multiple labels. At least some of the unlabeled audio samples contain named entity (NE) audio data. At least some of the labels include transcribed NE labels corresponding to the NE audio data. The method also includes correcting at least some of the transcribed NE labels using user-specific NE textual data. The method further includes retraining the set of teacher models based on the pseudo labeled dataset from a selected one of the teacher models, where the selected one of the teacher models predicts the pseudo labeled dataset more accurately than other teacher models of the set of teacher models.

Estimation accuracies of a conversation satisfaction and a speech satisfaction are improved. A learning data storage unit (10) stores learning data including a conversation voice containing a conversation including a plurality of speeches, a correct answer value of a conversation satisfaction for the conversation, and a correct answer value of a speech satisfaction for each speech included in the conversation. A model learning unit (13) learns a satisfaction estimation model using a feature quantity of each speech extracted from the conversation voice, the correct answer value of the speech satisfaction, and the correct answer value of the conversation satisfaction, the satisfaction estimation model configured by connecting a speech satisfaction estimation model part that receives a feature quantity of each speech and estimates the speech satisfaction of each speech with a conversation satisfaction estimation model part that receives at least the speech satisfaction of each speech and estimates the conversation satisfaction.

A method for detecting a hotword includes receiving a sequence of input frames that characterize streaming audio captured by a user device and generating a probability score indicating a presence of a hotword in the streaming audio using a memorized neural network. The network includes sequentially-stacked single value decomposition filter (SVDF) layers and each SVDF layer includes at least one neuron. Each neuron includes a respective memory component, a first stage configured to perform filtering on audio features of each input frame individually and output to the memory component, and a second stage configured to perform filtering on all the filtered audio features residing in the respective memory component. The method also includes determining whether the probability score satisfies a hotword detection threshold and initiating a wake-up process on the user device for processing additional terms.

A method (800) of training a text-to-speech (TTS) model (108) includes obtaining training data (150) including reference input text (104) that includes a sequence of characters, a sequence of reference audio features (402) representative of the sequence of characters, and a sequence of reference phone labels (502) representative of distinct speech sounds of the reference audio features. For each of a plurality of time steps, the method includes generating a corresponding predicted audio feature (120) based on a respective portion of the reference input text for the time step and generating, using a phone label mapping network (510), a corresponding predicted phone label (520) associated with the predicted audio feature. The method also includes aligning the predicted phone label with the reference phone label to determine a corresponding predicted phone label loss (622) and updating the TTS model based on the corresponding predicted phone label loss.

Embodiments of the present application disclose an audio matching method and a related device. The audio matching method includes: obtaining audio data and video data; extracting to-be-recognized audio information from the audio data; extracting lip movement information of N users from the video data, where N is an integer greater than 1; inputting the to-be-recognized audio information and the lip movement information of the N users into a target feature matching model, to obtain a matching degree between each of the lip movement information of the N users and the to-be-recognized audio information; and determining a user corresponding to the lip movement information of the user with the highest matching degree as the target user to which the to-be-recognized audio information belongs.

Examples are disclosed that relate to methods, computing devices, and systems for detecting an in-field event. One example provides a method comprising, during a training phase, receiving one or more training data streams. The training data stream(s) include an audio input comprising a semantic indicator. The audio input is processed to recognize the semantic indicator. A subset of data is selected and used to train a machine learning model to detect the in-field event, and the method further comprises outputting the trained machine learning model. During a run-time phase, the method comprises receiving one or more run-time input data streams. The trained machine learning model is used to detect a second instance of the in-field event in the one or more run-time input data streams. The method further comprises outputting an indication of the second instance of the in-field event.

Devices and techniques are generally described for a speech processing routing architecture. In various examples, first data comprising a first feature definition is received. The first feature definition may include a first indication of first source data and first instructions for generating feature data using the first source data. In various examples, the feature data may be generated according to the first feature definition. In some examples, a speech processing system may receive a first request to process a first utterance. The feature data may be retrieved from a non-transitory computer-readable memory. The speech processing system may determine a first skill for processing the first utterance based at least in part on the feature data.