Single channel overlapped-speech detection and separation of spontaneous conversations

Abstract

In the thesis, spontaneous conversation containing both speech mixture and speech dialogue is considered. The speech mixture refers to speakers speaking simultaneously (i.e. the overlapped-speech). The speech dialogue refers to only one speaker is actively speaking and the other is silent. That Input conversation is firstly processed by the overlapped-speech detection. Two output signals are then segregated into dialogue and mixture formats. The dialogue is processed by speaker diarization. Its outputs are the individual speech of each speaker. The mixture is processed by speech separation. Its outputs are independent separated speech signals of the speaker. When the separation input contains only the mixture, blind speech separation approach is used. When the separation is assisted by the outputs of the speaker diarization, it is informed speech separation. The research presents novel: overlapped-speech detection algorithm, and two speech separation algorithms. The proposed overlapped-speech detection is an algorithm to estimate the switching instants of the input. Optimization loop is adapted to adopt the best capsulated audio features and to avoid the worst. The optimization depends on principles of the pattern recognition, and k-means clustering. For of 300 simulated conversations, averages of: False-Alarm Error is 1.9%, Missed-Speech Error is 0.4%, and Overlap-Speaker Error is 1%. Approximately, these errors equal the errors of best recent reliable speaker diarization corpuses. The proposed blind speech separation algorithm consists of four sequential techniques: filter-bank analysis, Non-negative Matrix Factorization (NMF), speaker clustering and filter-bank synthesis. Instead of the required speaker segmentation, effective standard framing is contributed. Average obtained objective tests (SAR, SDR and SIR) of 51 simulated conversations are: 5.06dB, 4.87dB and 12.47dB respectively. For the proposed informed speech separation algorithm, outputs of the speaker diarization are a generated-database. The database associated the speech separation by creating virtual targeted-speech and mixture. The contributed virtual signals are trained to facilitate the separation by homogenising them with the NMF-matrix elements of the real mixture. Contributed masking optimized the resulting speech. Average obtained SAR, SDR and SIR of 341 simulated conversations are 9.55dB, 1.12dB, and 2.97dB respectively. Per the objective tests of the two speech separation algorithms, they are in the mid-range of the well-known NMF-based audio and speech separation methods.