Merci de votre visite. Cette page est disponible en anglais uniquement.
Alexa Blogs

Alexa Blogs

Want the latest?

alexa topics

Recent Posts

Archive

Showing posts tagged with Alexa science

May 21, 2019

Viktor Rozgic

The combination of an autoencoder, which is trained to output the same data it takes as input, and adversarial training, which pits two neural networks against each other, confers modest performance gains but opens the door to extensive training with unannotated data. 

[Read More]

May 16, 2019

Ming Sun

Text normalization is the process of converting particular words of a sentence into a standard format so that software can handle them. Breaking inputs into component parts and factoring in syntactic information reduces the error rate of a neural text normalization system by 98%.

[Read More]

May 13, 2019

Judith Gaspers

Cross-lingual transfer learning improves performance of named-entity recognizer by up to 7.4% versus a system trained from scratch in the new language.

[Read More]

May 03, 2019

Young-Bum Kim

In the past year, we’ve introduced name-free skill interaction for Alexa, which allows customers to invoke skills without mentioning them by name. Y. B. Kim explains how we add new skills to the skill selection model without causing "catastrophic forgetting."

[Read More]

May 02, 2019

Rahul Goel

Transfer learning, copy mechanism improve performance of "semantic parser".

[Read More]

April 25, 2019

Jakub Lachowicz

The ability to build new text-to-speech models with relatively little speaker-specific training data could enable a wide variety of customizable speaker styles.

[Read More]

April 22, 2019

Xing Fan

The wake word provides an acoustic profile that can be used to identify utterances from the same speaker.

[Read More]

April 18, 2019

Ming Sun

Alexa scientists use semi-supervised learning and "pyramidal" neural networks to address the problems of sound identification and media detection.

[Read More]

April 11, 2019

Jun Yang

Novel reconfigurable-filter-bank design enables more precise control of signal waveforms.

[Read More]

April 08, 2019

Quynh Do

Transfer of a model co-trained on intent classification and slot (variable) tagging halved the data required to achieve a given level of performance.

[Read More]

April 04, 2019

Hari Parthasarathi

To make it computationally feasible to train a speech recognizer on a million hours of speech, Alexa scientists used an array of techniques that could generalize to other large-scale machine learning projects.

[Read More]

April 01, 2019

Kenichi Kumatani

Echo devices use information about sound sources' directions to isolate speech signals. Turning speech isolation and automatic speech recognition into a single, large, machine learning problem improves speech recognition accuracy, even on devices with fewer microphones.

[Read More]

March 28, 2019

Yuan-yen Tai

Acoustic watermarking, which identifies audio signals through noise patterns imperceptible to humans, breaks down when signals are broadcast and re-captured by microphones. A new Alexa system is the first to solve this problem in real time, to prevent false device wakes and aid echo cancellation.

[Read More]

March 21, 2019

Rahul Gupta

To produce synthetic training data for a machine-learning application where real data is scarce, Alexa scientist Rahul Gupta uses generative adversarial learning, which pits two neural nets against each other -- one trying to generate convincing fakes, the other trying to discern fake from real.

[Read More]

March 20, 2019

Minhua Wu

Using one neural network to label speech data, adding synthetic environmental noise to that data, and then using it to train a second neural network improves speech recognition, particularly under noisy conditions.

[Read More]

Want the latest?

alexa topics

Recent Posts

Archive