We propose GE2E-KWS—a generalized end-to-end training and evaluation framework for customized keyword spotting. Specifically, enrollment utterances are separated and grouped by keywords from the training batch and their embedding centroids are compared to all other test utterance embeddings to compute the loss. This simulates runtime enrollment and verification stages, and improves convergence stability and training speed by optimizing matrix operations compared to SOTA triplet loss approaches. To benchmark different models reliably, we propose an evaluation process that mimics the production environment and compute metrics that directly measure keyword matching accuracy. Trained with GE2E loss, our 419KB quantized conformer model beats a 7.5 GB ASR encoder by 23.6% relative AUC, and beats a same size triplet loss model by 60.7% AUC. Our KWS models are natively streamable with low memory footprints, and designed to continuously run on-device with no retraining needed for new keywords (zero-shot).

Named groups are persistent groups of contacts with whom a user may wish to share the same information. We have engineered a test-bed for evaluating algorithms that recommend such groups. It assumes that these algorithms first generate social graphs from email logs and then mine these graphs to recommend the lists. It supports a variety of mechanisms to gather user data to support users with different privacy needs. It can be attached to a variety of algorithms for mining the data collected. It accommodates different kinds of models for using named groups and offers an evaluation mechanism that requires no effort from the user. The test-bed also provides visualizations of the generated social graphs and generated lists. Thus, it frees algorithm designers from gathering data and evaluating and understanding the output. It has been used to evaluate and compare multiple algorithms for recommending contact groups.

The sending of a message raises two important questions about its response: When will the first response arrive? When will the first acceptable response arrive? These questions can be partly or completely answered by identifying distributions of response times, correlating features with response tim

Users’ contacts often need to be grouped into equivalence classes for various purposes such as easily sending a message to all members of the group. Several approaches have been recently developed to make such predictions (a) for both ephemeral and persistent groups (b) in both email and social netw

In a social network, users can sort members of their social graph into friend lists to both understand the social structures within the graph and control the flow of incoming and outgoing information. To reduce the user-effort required to create these lists, previous work has developed techniques for generating friend-lists in a static social graph. This paper considers the user effort required to create friend lists in an evolving graph. We have developed several new initial quantitative metrics to capture this effort, and identified an initial technique for modeling graph growth. We have used these metrics and model to compare two techniques for evolving friend lists when the social graph grows: manual evolution - the user evolves friend lists using no external tools -- and full recommendation -- an existing state of the art tool recommends a whole new set of friend lists. In these comparisons, we used the friend lists of 12 individuals, and simulated the growth of their social graphs and friend lists using our graph-growth model. Intuitively, when the graph evolves by a small (large) amount, the manual (automatic) approach should perform better. Our experiments show that full recommendation performs better than manual when the social graph changes by more than 1%, and yields an almost complete reduction in effort in the best cases.

Previous email prediction algorithms generate individual predictions based on the past groupings of recipients or the contents of past emails. Our work builds on this research by (a) introducing new algorithms for extending and combining previous techniques and generating hierarchical recipient pred

A multi-domain collaboration toolkit hides heterogeneity of user-interface toolkits and associated domains from both programmers and end users of collaborative, widget-synchronizing, applications. We have developed such a system for the stand-alone, Eclipse, and web domains; and the AWT, Swing, SWT, and GWT single-user toolkits associated with these domains. Several new concepts are supported to meet these requirements including a widget server allowing a distributed widget client to manipulate widgets on an interactive device, flexible widget synchronization, flexible placement of widget listeners, "piping" centralized non-interactive replicas communicating with interactive user replicas, factory-based retargeting of the user-interface toolkit, and a new process architecture.