Amazon Technologies, Inc.

Viktor Rozgic

Chao Wang

Ming Sun

Qingming Tang

Chieh-Chi Kao

Daniel Kenneth Bone

Gustavo Alfonso Aguilar Alas

David Earl Heckerman

James G Droppo

Yang Liu

Spyridon Matsoukas

Siddhartha Reddy Jonnalagadda

Katherine M Ryan

Alexander Jonathan Pinkus

Qin Zhang

Alexandros Potamianos

Xiuye Chen

Avani Parakh

Srinivas Parthasarathy

Mohammad Taha Bahadori

Sripal Mehta

Andrew Thomas Bydlon

Steve Mark Lorusso

The present application provides techniques for implementing a skill component, configured to perform an assessment of a user, as part of a speech processing system. The system may receive a natural language user input requesting assistance. The skill component may, using one or more machine learning models, determine at least one characteristic of the natural language input (e.g., lexical embedding, acoustic embedding, topic, tone, etc.). The skill component may determine state data for a present session, where the state data indicates a topic of the natural language user input and/or a user state associated with the natural language user input. The skill component may determine past state data of one or more past sessions, and generate a question to the user based on the state data for the natural language user input and the past state data.

Conversation-based skill component for assessing a user's state

A system configured to detect custom acoustic events, where the system generates an acoustic event profile for the custom acoustic event based on a natural language description and without a sample of the sound. The system may generate a profile for a new custom sound based on a natural language description provided by the user; for example, a 'microwave beep.' If the system does not have an existing event profile for a microwave beep, the system may ask the user questions to determine whether any existing event profiles are close (e.g., is the sound similar to a “fan,” “alarm,” “appliance beep,” etc.). The system may detect an event that may be a possible match for the custom sound and ask the user to verify whether the detected event corresponds to the custom sound. The system may update the event profile based on the user's response.

Acoustic event detection

A system may be configured to detect custom acoustic events, where the system generates an acoustic event profile for the custom acoustic event based on a natural language description provided by a user and using an audio sample of the described acoustic event. For example, the user may describe the custom acoustic event as 'dog bark.' The system may ask the user questions to refine the description (e.g., dog breed, dog gender, age, etc.). Using an audio sample of the refined description, the system may then determine that audio captured in the user's environment is a potential sample of the custom acoustic event. Such captured audio may be presented to the user for confirmation, and then may be used to detect future occurrences of the custom acoustic event in the user's environment.

An acoustic event detection system may employ self-supervised federated learning to update encoder and/or classifier machine learning models. In an example operation, an encoder may be pre-trained to extract audio feature data from an audio signal. A decoder may be pre-trained to predict a subsequent portion of audio data (e.g., a subsequent frame of audio data represented by log filterbank energies). The encoder and decoder may be trained using self-supervised learning to improve the decoder's predictions and, by extension, the quality of the audio feature data generated by the encoder. The system may apply federated learning to share encoder updates across user devices. The system may fine-tune the classifier to improve inferences based on the improved audio feature data. The system may distribute classifier updates to the user device(s) to update the on-device classifier.

Self-supervised federated learning

Described is a system and method that determines character sequences from speech, without determining the words of the speech, and processes the character sequences to determine sentiment data indicative of emotional state of a user that output the speech. The emotional state may then be presented or provided as an output to the user.

Character-level emotion detection

Described herein is a system for sentiment detection in audio data. The system processes audio frame level features of input audio data using a machine learning algorithm to classify the input audio data into a particular sentiment category. The machine learning algorithm may be a neural network trained using an encoder-decoder method. The training of the machine learning algorithm may include normalization techniques to avoid potential bias in the training data that may occur when the training data is annotated for a perceived sentiment of the speaker.

Sentiment detection in audio data

Described herein is a system for sentiment detection in audio data. The system is trained using acoustic information and lexical information to determine a sentiment corresponding to an utterance. In some cases when lexical information is not available, the system (trained on acoustic and lexical information) is configured to determine a sentiment using only acoustic information.

Multiple classifications of audio data

Described herein is a system for emotion detection in audio data using a speaker's baseline. The baseline may represent a user's speaking style in a neutral emotional state. The system is configured to compare the user's baseline with input audio representing speech from the user to determine a emotion of the user. The system may store multiple baselines for the user, each associated with a different context (e.g., environment, activity, etc.), and select one of the baselines to compare with the input audio based on the contextual situation.

Emotion detection using speaker baseline

A device with a microphone acquires audio data of a user's speech. A neural network accepts audio data as input and provides sentiment data as output. The neural network is trained using training data based on input from raters who provide votes as to which sentiment descriptors they think are associated with a sample of speech. A vote by a rater assessing the sample for a particular semantic descriptor is distributed to a plurality of semantically similar semantic descriptors. Semantic descriptor similarity data indicates relative similarity between possible semantic descriptors in the semantic space. The distributed partial votes may be aggregated to produce training data comprising samples of speech and weights of corresponding semantic descriptors. The training data is then used to train the neural network. For example, the neural network may be trained with the training data using per-instance cosine similarity loss or correlational loss.

System to determine sentiment from audio data

Described herein is a system for media presence detection in audio. The system analyzes audio data to recognize whether a given audio segment contains sounds from a media source as a way of differentiating recorded media source sounds from other live sounds. In exemplary embodiments, the system includes a hierarchical model architecture for processing audio data segments, where individual audio data segments are processed by a trained machine learning model operating locally, and another trained machine learning model provides historical and contextual information to determine a score indicating the likelihood that the audio data segment contains sounds from a media source.

Growing community of inventors

Belmont, MA, United States of America

Viktor Rozgic