Amazon Technologies, Inc.

Yochai Zvik

Iftah Gamzu

Nissim Halabi

Natali Arieli

Marina Haikin

Ajay Kumar Soni

Adriano Devillaine

Sagi Bernstein

Yunqiang Zhu Zhu

Omer Shabtai Jakobinsky

Shlomi Chovel

Yossi Shasha

Colin Zhen De Kho

Kawshik Karur Rangaraju

Eran Fainman

Yaniv Ben-Yehuda

Moshe Peretz

Devices and techniques are generally described for adaptive targeting for voice notifications. In various examples, first data representing a predicted likelihood that a first user will interact with first content within a predefined amount of time may be received. A first set of features including features related to past voice notifications sent to the first user may be determined. A second set of features including features related to interaction with the first content when past voice notifications were sent may be received. A first machine learning model may generate a prediction that a voice notification will increase a probability that the first user interacts with the first content based on the first data, the first set of features, and the second set of features. Audio data comprising the voice notification may be sent to a first device associated with the first content.

Adaptive targeting for proactive voice notifications

Systems and methods are described for mitigating errors introduced during processing of user input such as voice input. A query may be derived from processed user input. A performance predictor analyzes the query and uses historical data to predict whether the query will return relevant results if executed. If the query's predicted performance is below a threshold, a query rewriter may identify potential alternatives to the query from a library of 'known good' queries. Different analyzers may be applied to identify different sets of alternatives, and machine learning models may be applied to rank the outputs of the analyzers. The best-matching alternatives from each analyzer may then be provided as inputs to a further machine learning model, which assesses the probability that each of the identified alternatives reflects the intent of the user. A most likely alternative may then be selected to execute in place of the original query.

Rewriting queries

A new approach to automatic speech recognition is disclosed. An example method include receiving a first text representing speech recognition of a phrase spoken by a user, isolating a candidate named entity from within the phrase, receiving a first phonetic representation of the candidate named entity, comparing the first phonetic representation to phonetic representations in a mapping database which map the phonetic representations to words to yield a comparison, based on the comparison, identifying a second phonetic representation in the mapping database that matches a second text in the mapping database to the second phonetic representation and replacing the candidate named entity with the second text. The approach can be used for new brands for which automatic speech recognition error rates are high.

System and method of providing recovery for automatic speech recognition errors for named entities

Techniques for speech-to-text hypothesis generation and hypothesis selection described. A text input representing at least part of a voice recording is received from a speech-to-text component. A first text alternative is generated using a finite state transducer based at least in part on the text input. A hypothesis from a hypothesis set is selected using a language model that includes probabilities for sequences of words, the hypothesis set including the text input and the first text alternative. A selected hypothesis text associated with the selected hypothesis is sent to a search engine.

Hypothesis generation and selection for inverse text normalization for search

Growing community of inventors

Modiin, Israel

Yochai Zvik