Crowd-supervised training of spoken language systems

Spoken language systems are often deployed with static speech recognizers. Only rarely are parameters in the underlying language, lexical, or acoustic models updated on-thefly. In the few instances where parameters are learned in an online fashion, developers traditionally resort to unsupervised training techniques, which are known to be inferior to their supervised counterparts. These realities make the development of spoken language interfaces a difficult and somewhat ad-hoc engineering task, since models for each new domain must be built from scratch or adapted from a previous domain. This thesis explores an alternative approach that makes use of human computation to provide crowd-supervised training for spoken language systems. We explore human-in-theloop algorithms that leverage the collective intelligence of crowds of non-expert individuals to provide valuable training data at a very low cost for actively deployed spoken language systems. We also show that in some domains the crowd can be incentivized to provide training data for free, as a byproduct of interacting with the system itself. Through the automation of crowdsourcing tasks, we construct and demonstrate organic spoken language systems that grow and improve without the aid of an expert. Techniques that rely on collecting data remotely from non-expert users, however, are subject to the problem of noise. This noise can sometimes be heard in audio collected from poor microphones or muddled acoustic environments. Alternatively, noise can take the form of corrupt data from a worker trying to game the system – for example, a paid worker tasked with transcribing audio may leave transcripts blank in hopes of receiving a speedy payment. We develop strategies to mitigate the effects of noise in crowd-collected data and analyze their efficacy. This research spans a number of different application domains of widely-deployed spoken language interfaces, but maintains the common thread of improving the speech recognizer’s underlying models with crowd-supervised training algorithms. We experiment with three central components of a speech recognizer: the language model, the lexicon, and the acoustic model. For each component, we demonstrate the utility of a crowd-supervised training framework. For the language model and lexicon, we explicitly show that this framework can be used hands-free, in two organic spoken language systems. Thesis Supervisor: Stephanie Seneff Title: Senior Research Scientist