Abstract: Systems and processes for operating an intelligent automated assistant are provided. In one example, a method includes receiving mixed speech data representing utterances of a target speaker and utterances of one or more interfering audio sources. The method further includes obtaining a target speaker representation, which represents speech characteristics of the target speaker; and determining, using a learning network, probability distributions of phonetic elements directly from the mixed speech data. The inputs of the learning network include the mixed speech data and the target speaker representation. An output of the learning network includes the probability distributions of phonetic elements. The method further includes generating text corresponding to the utterances of the target speaker based on the probability distributions of the phonetic elements; and providing a response to the target speaker based on the text corresponding to the utterances of the target speaker.

Abstract: Systems and processes for operating an intelligent automated assistant are provided. In one example, a method includes receiving mixed speech data representing utterances of a target speaker and utterances of one or more interfering audio sources. The method further includes obtaining a target speaker representation, which represents speech characteristics of the target speaker; and determining, using a learning network, probability distributions of phonetic elements directly from the mixed speech data. The inputs of the learning network include the mixed speech data and the target speaker representation. An output of the learning network includes the probability distributions of phonetic elements. The method further includes generating text corresponding to the utterances of the target speaker based on the probability distributions of the phonetic elements; and providing a response to the target speaker based on the text corresponding to the utterances of the target speaker.

Abstract: Techniques for inverse text normalization are provided. In some examples, speech input is received and a spoken-form text representation of the speech input is generated. The spoken-form text representation includes a token sequence. A feature representation is determined for the spoken-form text representation and a sequence of labels is determined based on the feature representation. The sequence of labels is assigned to the token sequence and specifies a plurality of edit operations to perform on the token sequence. Each edit operation of the plurality of edit operations corresponds to one of a plurality of predetermined types of edit operations. A written-form text representation of the speech input is generated by applying the plurality of edit operations to the token sequence in accordance with the sequence of labels. A task responsive to the speech input is performed using the generated written-form text representation.

Abstract: Techniques for inverse text normalization are provided. In some examples, speech input is received and a spoken-form text representation of the speech input is generated. The spoken-form text representation includes a token sequence. A feature representation is determined for the spoken-form text representation and a sequence of labels is determined based on the feature representation. The sequence of labels is assigned to the token sequence and specifies a plurality of edit operations to perform on the token sequence. Each edit operation of the plurality of edit operations corresponds to one of a plurality of predetermined types of edit operations. A written-form text representation of the speech input is generated by applying the plurality of edit operations to the token sequence in accordance with the sequence of labels. A task responsive to the speech input is performed using the generated written-form text representation.

Abstract: A non-transitory computer-readable storage medium stores one or more programs including instructions, which when executed by an electronic device, cause the electronic device to receive natural-language speech input from one of a plurality of users, the natural-language speech input having a set of acoustic properties; and determine whether the natural-language speech input corresponds to both a user-customizable lexical trigger and a set of acoustic properties associated with the user; where in accordance with a determination that the natural language speech input corresponds to both a user-customizable lexical trigger and a set of acoustic properties associated with the user, invoke a virtual assistant; and in accordance with a determination that either the natural language speech input fails to correspond to a user-customizable lexical trigger or the natural-language speech input fails to have a set of acoustic properties associated with the user, forego invocation of a virtual assistant.

Abstract: A new word model is trained from synthetic word samples derived by Monte Carlo techniques from one or more prior word models. The prior word model can be a phonetic word model and the new word model can be a non-phonetic, whole-word, word model. The prior word model can be trained from data that has undergone a first channel normalization and the synthesized word samples from which the new word model is trained can undergo a different channel normalization similar to that to be used in a given speech recognition context. The prior word model can have a first model structure and the new word model can have a second, different, model structure. These differences in model structure can include, for example, differences of model topology; differences of model complexity; and differences in the type of basis function used in a description of such probability distributions.