How we update Alexa's skill selector without causing “catastrophic forgetting”developer.amazon.com·6 pts·georgecarlyle76·0
“Alexa, turn down the lights and play music”: How to handle compound requestsdeveloper.amazon.com·8 pts·georgecarlyle76·0
Multi-speaker TTS models outperform single-speaker ones but sound the samedeveloper.amazon.com·5 pts·georgecarlyle76·0
Using wake word acoustics to filter background speech improves speech recognizerdeveloper.amazon.com·7 pts·georgecarlyle76·0
Alexa uses tri-training, “pyramidal” NNs for sound recognition, media detectiondeveloper.amazon.com·5 pts·georgecarlyle76·0
Cross-Lingual Transfer Learning Halves Data Requirements to Bootstrap AI Systemsdeveloper.amazon.com·7 pts·georgecarlyle76·0
Alexa researchers train speech recognizer on 1M hours of unlabeled speechdeveloper.amazon.com·12 pts·georgecarlyle76·1
Joint training on beamforming/ASR improves Alexa's ASR, with fewer microphonesdeveloper.amazon.com·3 pts·georgecarlyle76·0
Audio Watermarking Algorithm First to Solve Second-Screen Problem in Real Timedeveloper.amazon.com·10 pts·georgecarlyle76·0
Adversarial Training Produces Synthetic Data for Machine Learningdeveloper.amazon.com·10 pts·georgecarlyle76·0
Machine-Labeled Data plus Artificial Noise = Better Speech Recognitiondeveloper.amazon.com·6 pts·georgecarlyle76·0
To Correct Imbalances in Training Data, Don’t Oversample: Clusterdeveloper.amazon.com·5 pts·georgecarlyle76·0
Conversational-AI Innovations from Contestants in the 2018 Alexa Prizedeveloper.amazon.com·7 pts·georgecarlyle76·0
The AI Tools That Amazon Built to Let Alexa Prize Participants Focus on Sciencedeveloper.amazon.com·1 pts·georgecarlyle76·0
Why Alexa Won't Wake Up When She Hears Her Name in Amazon's Super Bowl Addeveloper.amazon.com·6 pts·georgecarlyle76·0
Transfer learning trains NN to recognize new classes, w/o original training datadeveloper.amazon.com·6 pts·georgecarlyle76·0
With new kernel method, NLP system learns in four hours what used to take 7 daysdeveloper.amazon.com·17 pts·georgecarlyle76·0
Leveraging Unannotated Data to Bootstrap Alexa Functions More Quicklydeveloper.amazon.com·5 pts·georgecarlyle76·0
New Method for Compressing Neural Networks Better Preserves Accuracydeveloper.amazon.com·120 pts·georgecarlyle76·36
How to use the (non-differentiable) F1 score to train a machine learning modeldeveloper.amazon.com·4 pts·georgecarlyle76·0
How Alexa knows “peanut butter” is one shopping-list item, not twodeveloper.amazon.com·246 pts·georgecarlyle76·140
New data representation scheme reduces Alexa's skill selection errors by 40%developer.amazon.com·8 pts·georgecarlyle76·0
Building language models directly from grammars reduces ASR error up to 15%developer.amazon.com·3 pts·georgecarlyle76·0
How NLU domains can independently calibrate their decisions to field a querydeveloper.amazon.com·3 pts·georgecarlyle76·0
Amazon scientists: What NeurIPS (formerly NIPS) tells us about the state of AIblog.aboutamazon.com·1 pts·georgecarlyle76·0
New details about how Alexa decides which skill will handle a requestdeveloper.amazon.com·1 pts·georgecarlyle76·0
New details about the model that decides which Alexa skill will handle a requestdeveloper.amazon.com·7 pts·georgecarlyle76·0
Researchers at Amazon and U. of Sheffield team on project to detect fake newsdeveloper.amazon.com·4 pts·georgecarlyle76·0
Locating sound with an Echo's microphone array could enable clap-based controlsdeveloper.amazon.com·4 pts·georgecarlyle76·0