Amazon Developer Blogs

Amazon Developer Blogs

Showing posts tagged with Alexa Science

September 04, 2018

Young-Bum Kim

Classifying "domains", or topics of conversation, is so central to natural-language understanding that some systems have modules that just look for out-of-domain utterances. Amazon researchers show that training out-of-domain classifiers together with domain recognizers improves their performance.

[Read More]

August 31, 2018

Viktor Rozgic

At Interspeech 2018, Amazon scientists report a new system that learns to distinguish sounds produced by media players, such as TVs and radios, from sounds of ordinary household activity, to reduce false positives from Alexa's speech recognizers still further.

[Read More]

August 29, 2018

Larry Hardesty

Amazon papers at Interspeech 2018 examine ways that interaction histories can improve customers' Alexa experiences, from "collaborative filtering" to infer customers' tastes to methods that use the immediate "context" of a customer's requests to increase the accuracy of speech understanding systems.

[Read More]

August 27, 2018

Arpit Gupta

Alexa AI machine-learning scientist Chetan Naik and speech scientist Arpit Gupta describe their Interspeech paper on "contextual slot carryover," a crucial element of Alexa's ability to conduct natural, multiround conversations.

[Read More]

August 24, 2018

Larry Hardesty

3 questions with Björn Hoffmeister, the senior manager for machine learning in the Alexa Automatic Speech Recognition group, about this year's Interspeech conference and research trends in the speech technology field.

[Read More]

August 23, 2018

Harish Mallidi

At this year's Interspeech, Amazon scientists report a neural network that learns to distinguish speech that's directed at an Echo device from speech that isn't, to enable Alexa to temporarily dispense with wake word repetition.

[Read More]

August 19, 2018

Larry Hardesty

Two months after Amazon released a dataset containing assertions of fact and Wikipedia data that either substantiate or refute them, preliminary results of a competition to use the dataset to train fact verification systems indicate a 134% improvement over baseline.

[Read More]

August 13, 2018

Grant Strimel

How a new algorithm for "perfect hashing" enables Amazon scientists to shrink the memory footprint of machine learning models by 94%.

[Read More]

August 08, 2018

Yuval Merhav

Amazon AI researchers have publicly released a new dataset with transliterations of 400,000 names, to aid the development of systems that can search for data across languages that use different scripts.

[Read More]

July 23, 2018

Anirudh Raju

An Alexa speech scientist explains how Alexa can dynamically modify the statistical language models underlying her speech recognition systems to improve performance for individual customers in particular contexts.

[Read More]