Jindrich Zdansky

Jiri Malek, Jakub Jansky, Zbynek Koldovsky et al.

This manuscript proposes a novel robust procedure for the extraction of a speaker of interest (SOI) from a mixture of audio sources. The estimation of the SOI is performed via independent vector extraction (IVE). Since the blind IVE cannot distinguish the target source by itself, it is guided towards the SOI via frame-wise speaker identification based on deep learning. Still, an incorrect speaker can be extracted due to guidance failings, especially when processing challenging data. To identify such cases, we propose a criterion for non-intrusively assessing the estimated speaker. It utilizes the same model as the speaker identification, so no additional training is required. When incorrect extraction is detected, we propose a ``deflation'' step in which the incorrect source is subtracted from the mixture and, subsequently, another attempt to extract the SOI is performed. The process is repeated until successful extraction is achieved. The proposed procedure is experimentally tested on artificial and real-world datasets containing challenging phenomena: source movements, reverberation, transient noise, or microphone failures. The method is compared with state-of-the-art blind algorithms as well as with current fully supervised deep learning-based methods.

Jan Nouza, Petr Cerva, Jindrich Zdansky

In the paper we present an architecture of a keyword spotting (KWS) system that is based on modern neural networks, yields good performance on various types of speech data and can run very fast. We focus mainly on the last aspect and propose optimizations for all the steps required in a KWS design: signal processing and likelihood computation, Viterbi decoding, spot candidate detection and confidence calculation. We present time and memory efficient modelling by bidirectional feedforward sequential memory networks (an alternative to recurrent nets) either by standard triphones or so called quasi-monophones, and an entirely forward decoding of speech frames (with minimal need for look back). Several variants of the proposed scheme are evaluated on 3 large Czech datasets (broadcast, internet and telephone, 17 hours in total) and their performance is compared by Detection Error Tradeoff (DET) diagrams and real-time (RT) factors. We demonstrate that the complete system can run in a single pass with a RT factor close to 0.001 if all optimizations (including a GPU for likelihood computation) are applied.