The DIRHA Portuguese Corpus: A Comparison of Home Automation Command Detection and Recognition in Simulated and Real Data.

The DIRHA Portuguese Corpus: A Comparison of Home Automation Command Detection and Recognition in Simulated and Real Data Miguel Matos1,2, Alberto Abad1,2, Ant´onio Serralheiro1,3 1L2F - Spoken Language Systems Lab, 2IST - Instituto Superior T´ecnico, University of Lisbon, 3Academia Militar INESC-ID Lisbon, Lisbon - Portugal jmatos, alberto, ajs@l2f.inesc-id.pt Abstract In this paper, we describe a new corpus –named DIRHA-L2F RealCorpus– composed of typical home automation speech interactions in European Portuguese that has been recorded by the INESC-ID’s Spoken Language Systems Laboratory (L2F) to support the activities of the Distant-speech Interaction for Robust Home Applications (DIRHA) EU-funded project. The corpus is a multi-microphone and multi-room database of real continuous audio sequences containing read phonetically rich sentences, read and spontaneous keyword activation sentences, and read and spontaneous home automation commands. The background noise conditions are controlled and randomly recreated with noises typically found in home environments. Experimental validation on this corpus is reported in comparison with the results obtained on a simulated corpus using a fully automated speech processing pipeline for two fundamental automatic speech recognition tasks of typical “always-listening” home-automation scenarios: system activation and voice command recognition. Attending to results on both corpora, the presence of overlapping voice-like noise is shown as the main problem: simulated sequences contain concurrent speakers that result in general in a more challenging corpus, while real sequences performance drops drastically when TV or radio is on. Keywords: robust speech processing, real vs. simulated data, voice activity detection, speech recognition, keyword spotting, home automation 1. Introduction In domestic environments, the presence of reverberation, background noise and interfering sources, critically de- grades the performance of standard speech processing al- gorithms. A possible solution for improving the overall recognition robustness is the adoption of a network of dis- tributed microphones that can partially reduce the impact of these nuisance factors on spoken dialogue home automa- tion systems (Seltzer, 2003; Chu et al., 2006). This is the scenario addressed in the DIRHA project1, in which an “always-listening” monitoring system is foreseen to be able to analyse and understand the activities and the intentions of the users inside a domestic environment. In the most typical DIRHA scenario, the automation system can be op- erated based on a dialogue manager strategy in which spo- ken interactions are usually initiated by the user by means of a certain activation sentence. Then, a spoken dialogue session is initiated, in which the user can provide specific home automation commands or ask for status information. The availability of relevant corpora characterizing home environments is of fundamental importance to support re- search on the area of multi-channel processing for smart- home applications. Thus, some related data collection ef- forts have been conducted in the past. For instance, in the context of smart-room services for meeting and semi- nar assistance, in which typically a single room is simul- taneously characterized (Janin et al., 2003; McCowan et al., 2005; Moreau et al., 2008). Concerning multiple- room environments, in (Vacher et al., 2014) automation commands in French were recorded in …

Content truncated to 500 words