Alexa Speech Science Technology
The Amazon Alexa Speech Science and Machine Learning team employs cutting-edge research and technology to create magical customer experiences with Alexa, the voice service that powers Amazon’s family of Echo products, Amazon Fire TV, and more.
Alexa is a spoken language understanding system comprised of several data-driven audio and text processing components that process and respond to customer voice requests. These components include signal processing, wake word detection, automatic speech recognition (ASR), natural language understanding (NLU), question answering, dialog management, and text-to-speech synthesis.
For example, on a hands-free WiFi-connected audio device such as Echo, a customer may ask, “Alexa, what’s the weather in Boston?” The wake word engine then detects the word “Alexa” and begins streaming audio to the speech platform in the Amazon cloud. The platform passes the audio stream to the ASR module, which returns a list of the most likely transcriptions of the user’s speech. This result is then transferred to NLU, which analyzes the top ASR results to extract the most likely interpretation of the user’s request, consisting of an intent (e.g., “GetWeatherForecast”) and any associated slots (e.g., “Boston”). Based on the type of intent, the speech platform routes the intent to the appropriate skill, which then specifies what Alexa should say to the user. Finally, Text-to-Speech technology converts Alexa’s text response into audio, which is streamed to the device to play back to the user from the device’s speaker.
This seamless interaction requires relentless focus on the customer experience and customer feedback. The speech science team focuses on incorporating learning from every customer interaction to accelerate Alexa’s learning, using highly scalable deep learning techniques. The Amazon speech scientist team trains deep neural networks on large datasets using distributed processing, working at massive Amazon scale to optimize training for the AWS network. Learning at scale requires the right balance of invention and simplification to find the set of algorithms that maximize Alexa’s accuracy given the data. The challenge of interacting with the Amazon-scale catalog of shopping, music, and media requires world-class solutions for almost every known NLP task, from Anaphora Resolution to Semantic Parsing.
Past Interspeech publications and submissions from the Amazon team are listed below.
Meet the Team
Are you ready for your next opportunity? Check out our open positions on this page, and meet some of our speech scientists here. We have global opportunities available, and speech and machine learning scientists from the following locations will be available to meet:
If you would like to meet with a speech scientist in person at the conference, please contact firstname.lastname@example.org.
Sit back, relax and watch the 30-foot wolf and 1,000-pound gorilla wreck Chicago as the 1986 arcade game comes to life. (adsbygoog
In the week leading up to Thanksgiving, I talk to many patients about how to approach the big day in a way that balances their two competing prioritie
Senators Mark Warner (Va.) and Amy Klobuchar (Minn.) created a Facebook page for a fictional political group — Americans for Disclosure Solution
he newest, most outrageous, “super extreme” idea from Lamborghini is a box. Well, metaphorically. The Italian supercar giant unveiled a
Apple plans to give developers more freedom over the introductory pricing levels of in-app subscriptions. As referenced in the iOS 11.2 beta release n
After many years of fighting in the courts, Apple has finally claimed victory over Samsung to the count of $120 million. The case revolved aro