Room-localized spoken command recognition in multi-room, multi-microphone environments

Room-localized speech activity detection in multi-microphone smart homes

“…In subsequent work [34], the first stage was replaced by a multi-channel room-independent SAD module, whereas the second stage adopted the use of specific features to discriminate room-inside vs. roomoutside speech by means of SVM-based classifiers. The approach was further refined in later work [16], as part of a modular pipeline of a smart home spoken command recognition system.…”

“…The experiments in Sections 8.1-8.4 are conducted on two databases: the Greek-language part of DIRHAsimcorpora II [61], hereafter referred to as "DIRHAsim", and the "DIRHA-real" Greek corpus [16]. 2 The datasets are either simulated or recorded inside a smart home apartment (with an average reverberation time of 0.72 s), developed for the purposes of the DIRHA research project [10].…”

“…In total, 150 1-min simulation sequences containing speech and noise are available. In contrast, the DIRHA-real set contains actual recordings of 5 subjects acquired by the 40 microphones inside the smart home under realistic noise conditions [16]. In total, 60 1-min recorded sequences of speech and noise are available.…”

“…Right: 1-min-long waveforms, captured by the red-colored microphones (one per room with an active speaker), shown together with the corresponding ground truth of room-localized SAD. The room-independent speech/non-speech segmentation is also depicted at the top such "room-localized SAD" variant can be useful in multiple ways: It can help disambiguate user commands for voice control of devices or appliances present in multiple rooms (e.g., light switches, windows, temperature control units, television sets); enable room-localized system feedback, for example, via a loudspeaker or visual display at the room where speech activity takes place; and allow parallel voice interaction sessions by multiple subjects inside different rooms, engaging separate system pipelines, one per room [16]; finally, ASR itself can benefit significantly from room localization [25].…”

“…It should be noted that two of the aforementioned extensions have already been proposed in our earlier work [16], namely the room-localized operation at the first stage, as well as the sliding window mode at the second. However, since our focus there has been on the entire pipeline of the smart home spoken command recognition system, a contrastive evaluation of these enhancements for roomlocalized SAD has not been investigated.…”