To better model the contextual information and increase the generalization ability of a voice detection system, this paper leverages a multi-lingual Automatic Speech Recognition (ASR) system to perform Speech Activity Detection (SAD). Sequence-discriminative training of multi-lingual Acoustic Model (AM) using Lattice-Free Maximum Mutual Information (LF-MMI) loss function, effectively extracts the contextual information of the input acoustic frame. The index of maximum output posterior is considered as a frame-level speech/non-speech decision function. Majority voting and logistic regression are applied to fuse the language-dependent decisions. The leveraged multi-lingual ASR is trained on 18 languages of BABEL datasets and the built SAD is evaluated on 3 different languages. In out-of-domain datasets, the proposed SAD model shows significantly better performance w.r.t. baseline models. In the Ester2 dataset, without using any in-domain data, this model outperforms the WebRTC, phoneme recognizer based VAD (Phn\_Rec), and Pyannote baselines (respectively 7.1, 1.7, and 2.7% absolutely) in Detection Error Rate (DetER) metrics. Similarly, in the LiveATC dataset, this model outperforms the WebRTC, Phn\_Rec, and Pyannote baselines (respectively 6.4, 10.0, and 3.7% absolutely) in DetER metrics.
View on arXiv