Universal dialogue-act tagging scheme, hybrid slot-tracking system promise to improve dialogue state tracking.

Techniques include weighting training samples from out-of-domain data sets and noise contrastive estimation, which turns the calculation of massive probability distributions into simple binary decisions.

Two Interspeech papers report a system that transfers prosody — inflection and rhythm — from a recorded speaker to a synthesized voice and a neural vocoder that works with any speaker.

Based on embeddings, system suggests named entities — or "slot values" — that developers might want their skills to recognize.

Two new papers explore techniques for increasing the computational efficiency and reducing the memory footprints of neural networks that process audio data.

Pooling the training data for related skills, and using it to train the skills simultaneously, improves performance for all of them.

Different Alexa services use different names for the same types of data, which makes it hard to track references across dialogues. By learning correlations between data types, a machine learning model can make better decisions about which references to track from one round of dialogue to the next.

Most question-answering systems rely on either text search or knowledge graphs. A hybrid approach, which knits together data from multiple textual sources to produce an ad hoc knowledge graph, yields better results on complex questions.

The new focus areas for the Amazon Research Awards, which provide up to $80,000 in funding and up to $20,000 in Amazon Web Services Promotional Credits to academic researchers investigating topics related to machine learning, were announced this month. The application period opens on September 10.

Treating news headlines and Wikipedia section headers as "search terms" and the associated texts as search results enables the training of neural search engines with less need for manually annotated data.

Student system demonstrates how "early exit" strategies could improve Alexa's efficiency, by letting neural networks break off computations when they have high confidence in their solutions.

Alexa scientists use machine learning to improve the performance of "multilabel classifiers", which classify data according to several categories at once — identifying multiple objects in an image, for instance, or multiple topics touched on by a single article.

New approach to "active learning", or automatically selecting training examples for machine learning, improves the performance of natural-language-understanding system by 7% to 9%.

Natural-language-understanding system that includes both a generic model for a language and several locale-specific models improves accuracy by an average of 59% over individual locale-specific models.

A new Alexa system adopts a simpler, more scalable solution to the problem of "reference resolution", or tracking references through several rounds of dialogue, by overwriting referring words with their referents.