Learning to Rank Intents in Voice Assistants

Abstract

Voice Assistants aim to fulfill user requests by choosing the best intent from multiple options generated by its Automated Speech Recognition and Natural Language Understanding sub-systems. However, voice assistants do not always produce the expected results. This can happen because voice assistants choose from ambiguous intents -user-specific or domain-specific contextual information reduces the ambiguity of the user request. Additionally the user information-state can be leveraged to understand how relevant/executable a specific intent is for a user request. In this work, we propose a novel Energy-based model for the intent ranking task, where we learn an affinity metric and model the trade-off between extracted meaning from speech utterances and relevance/executability aspects of the intent. Furthermore we present a Multisource Denoising Autoencoder based pretraining that is capable of learning fused representations of data from multiple sources. We empirically show our approach outperforms existing state of the art methods by reducing the error-rate by 3.8%, which in turn reduces ambiguity and eliminates undesired dead-ends leading to better user experience. Finally, we evaluate the robustness of our algorithm on the intent ranking task and show our algorithm improves the robustness by 33.3%.

From User Perceptions to Technical Improvement: Enabling People Who Stutter to Better Use Speech Recognition

“…An initial cohort of 50 participants (Phase 1: 25 mild; 18 moderate; 7 severe) 4 was collected starting in Summer 2020 and 41 additional (Phase 2: 0 mild; 26 moderate; 15 severe) in 2021. 5 VA phrases were based on prompts where participants came up with their own utterance to accomplish a given task. This was designed to prevent participants from simply reading pre-defined phrases in a monotone manner, and some examples include "Find out when a place you might want to visit will open," "Find the lyrics to a song that you like," and "Ask a question about something that is not well known. "…”

“…A&H severities were computed using the same process as with the speech survey 5. 65 participants successfully recorded 131 utterances, 2 recorded 130 utterances, and 24 only recorded 121 VA commands.…”

Analysis and Tuning of a Voice Assistant System for Dysfluent Speech

“…Table 5 shows the improvement our tuned ASR system has on VA-Dysfluent using recent domain [24] and intent recognition [25] models. Overall, there is a 3.6% and 1.7% relative increase in domain and intent recognition performance, respectively (statistically significant), obtained from the text generated by the tuned ASR decoder as opposed to the default.…”

The Future of New Learning Way for The Visually Impaired Learners: Use Bionic Eye with Voice Assistant

“…As for Voice Assistant, it' must overcome complex problems and hence they typically are formed of a number of components: one that transcribes the user speech (Automated Speech Recognition -ASR), one that understands the transcribed utterances (Natural Language Understanding -NLU), one that makes decisions (Decision Making -DM), and one that produces the output speech (TTS) (R. Anantha, Chappidi, & Dawoodi, 2021;Raviteja Anantha, Chappidi, & Dawoodi, 2020). Many VAs have similar with the chatbot (Weber, Ritschel, Lingenfelser, & André, 2018), our discussion will focused on the how to use successful learning chatbot to apply on the VAs system, and also combined with the bionic eye technology to let the blind use the new way to learn.…”

“…We sorted it into three categories. All privatesensitive information stays on the user's device (R. Anantha et al, 2021 Bionic Eye Architecture: The main objective of bionic eye is to provide vision to those visually impaired people who are suffering from retinal diseases. By conducting bionic eye in human eye, this way will be able to help a person to recognize the object .…”