Described are techniques for noise-robust and speaker-independent keyword spotting (KWS) in an input audio signal that contains keywords used to activate voice-based human-computer interactions. A KWS system may combine the latent representation generated by a denoising autoencoder (DAE) with audio features extracted from the audio signal using a machine learning approach. The DAE may be a discriminative DAE trained with a quadruplet loss metric learning approach to create a highly-separable latent representation of the audio signal in the audio input feature space. In one aspect, spectral characteristics of the audio signal such as Log-Mel features are combined with the latent representation generated by a quadruplet loss variational DAE (QVDQE) as input to a DNN KWS classifier. The KWS system improves keyword classification accuracy versus using extracted spectral features alone, non-discriminative DAE latent representations alone, or the extracted spectral features combined with the non-discriminative DAE latent representations in a KWS classifier.

Described are techniques for noise-robust and speaker-independent keyword spotting (KWS) in an input audio signal that contains keywords used to activate voice-based human-computer interactions. A KWS system may combine the latent representation generated by a denoising autoencoder (DAE) with audio features extracted from the audio signal using a machine learning approach. The DAE may be a discriminative DAE trained with a quadruplet loss metric learning approach to create a highly-separable latent representation of the audio signal in the audio input feature space. In one aspect, spectral characteristics of the audio signal such as Log-Mel features are combined with the latent representation generated by a quadruplet loss variational DAE (QVDQE) as input to a DNN KWS classifier. The KWS system improves keyword classification accuracy versus using extracted spectral features alone, non-discriminative DAE latent representations alone, or the extracted spectral features combined with the non-discriminative DAE latent representations in a KWS classifier.

Automatic measurement of semantic textual similarity of conversations, by: receiving two conversation texts, each comprising a sequence of utterances; encoding each of the sequences of utterances into a corresponding sequence of semantic representations; computing a minimal edit distance between the sequences of semantic representations; and, based on the computation of the minimal edit distance, performing at least one of: quantifying a semantic similarity between the two conversation texts, and outputting an alignment of the two sequences of utterances with each other.

Automatic measurement of semantic textual similarity of conversations, by: receiving two conversation texts, each comprising a sequence of utterances; encoding each of the sequences of utterances into a corresponding sequence of semantic representations; computing a minimal edit distance between the sequences of semantic representations; and, based on the computation of the minimal edit distance, performing at least one of: quantifying a semantic similarity between the two conversation texts, and outputting an alignment of the two sequences of utterances with each other.

Disclosed herein are systems, computer-implemented methods, and tangible computer-readable storage media for speaker recognition personalization. The method recognizes speech received from a speaker interacting with a speech interface using a set of allocated resources, the set of allocated resources including bandwidth, processor time, memory, and storage. The method records metrics associated with the recognized speech, and after recording the metrics, modifies at least one of the allocated resources in the set of allocated resources commensurate with the recorded metrics. The method recognizes additional speech from the speaker using the modified set of allocated resources. Metrics can include a speech recognition confidence score, processing speed, dialog behavior, requests for repeats, negative responses to confirmations, and task completions. The method can further store a speaker personalization profile having information for the modified set of allocated resources and recognize speech associated with the speaker based on the speaker personalization profile.

Disclosed herein are systems, computer-implemented methods, and tangible computer-readable storage media for speaker recognition personalization. The method recognizes speech received from a speaker interacting with a speech interface using a set of allocated resources, the set of allocated resources including bandwidth, processor time, memory, and storage. The method records metrics associated with the recognized speech, and after recording the metrics, modifies at least one of the allocated resources in the set of allocated resources commensurate with the recorded metrics. The method recognizes additional speech from the speaker using the modified set of allocated resources. Metrics can include a speech recognition confidence score, processing speed, dialog behavior, requests for repeats, negative responses to confirmations, and task completions. The method can further store a speaker personalization profile having information for the modified set of allocated resources and recognize speech associated with the speaker based on the speaker personalization profile.

A system and method of automatically discovering unigrams in a speech data element may include receiving a language model that includes a plurality of n-grams, where each n-gram includes one or more unigrams; applying an acoustic machine-learning (ML) model on one or more speech data elements to obtain a character distribution function; applying a greedy decoder on the character distribution function, to predict an initial corpus of unigrams; filtering out one or more unigrams of the initial corpus to obtain a corpus of candidate unigrams, where the candidate unigrams are not included in the language model; analyzing the one or more first speech data elements, to extract at least one n-gram that comprises a candidate unigram; and updating the language model to include the extracted at least one n-gram.

A system and method of automatically discovering unigrams in a speech data element may include receiving a language model that includes a plurality of n-grams, where each n-gram includes one or more unigrams; applying an acoustic machine-learning (ML) model on one or more speech data elements to obtain a character distribution function; applying a greedy decoder on the character distribution function, to predict an initial corpus of unigrams; filtering out one or more unigrams of the initial corpus to obtain a corpus of candidate unigrams, where the candidate unigrams are not included in the language model; analyzing the one or more first speech data elements, to extract at least one n-gram that comprises a candidate unigram; and updating the language model to include the extracted at least one n-gram.

An information processing system includes: a first device configured to acquire a user's uttered voice, transfer the user's uttered voice to at least one of a second and a third devices each actualizing a voice interaction agent, when a control command is acquired, convert a control signal based on the acquired control command to a control signal that matches the second device, and transmit the converted control signal to the second device; a second device configured to recognize the uttered voice transferred from the first device, and output, to the first device, a control command regarding a recognition result obtained by recognizing the uttered voice and response data based on the control signal; and a third device configured to recognize the uttered voice transferred from the first device, and output, to the first device, a control command regarding a recognition result obtained by recognizing the uttered voice.

An information processing system includes: a first device configured to acquire a user's uttered voice, transfer the user's uttered voice to at least one of a second and a third devices each actualizing a voice interaction agent, when a control command is acquired, convert a control signal based on the acquired control command to a control signal that matches the second device, and transmit the converted control signal to the second device; a second device configured to recognize the uttered voice transferred from the first device, and output, to the first device, a control command regarding a recognition result obtained by recognizing the uttered voice and response data based on the control signal; and a third device configured to recognize the uttered voice transferred from the first device, and output, to the first device, a control command regarding a recognition result obtained by recognizing the uttered voice.