Speech separation has been studied widely for single-channel close-talk microphone recordings over the past few years; developed solutions are mostly in frequency-domain. Recently, a raw audio waveform separation network (TasNet) is introduced for single-channel data, with achieving high Si-SNR (scale-invariant source-to-noise ratio) and SDR (sourceto-distortion ratio) comparing against the state-of-the-art solution in frequency-domain. In this study, we incorporate effective components of the TasNet into a frequency-domain separation method. We compare both for alternative scenarios. We introduce a solution for directly optimizing the separation criterion in frequency-domain networks. In addition to speech separation objective and subjective measurements, we evaluate the separation performance on a speech recognition task as well. We study the speech separation problem for far-field data (more similar to naturalistic audio streams) and develop multi-channel solutions for both frequency and time-domain separators with utilizing spectral, spatial and speaker location information. For our experiments, we simulated multi-channel spatialized reverberate WSJ0-2mix dataset. Our experimental results show that spectrogram separation can achieve competitive performance with better network design. Multi-channel framework as well is shown to improve the single-channel performance relatively up to +35.5% and +46% in terms of WER and SDR, respectively.

Concealing speaker identity in speech signals refers to the task of speaker de-identification, which helps protect the privacy of a speaker. Although, both linguistic and paralinguistic features reveal personal information of a speaker and they both need to be addressed, in this study we only focus on speaker voice characteristics. In other words, our goal is to move away from the source speaker identity while preserving naturalness and quality. The proposed speaker de-identification system maps voice of a given speaker to an average (or gender-dependent average) voice; the mapping is modeled by a new convolutional neural network (CNN) encoder-decoder architecture. Here, the transformation of both spectral and excitation features are studied. The voice conversion challenge 2016 (VCC-2016) database is used to train the system and examine performance of the proposed method. We use two different approaches for evaluations:(1) objective evaluation: equal error rates (EERs) calculated by an i-vector/PLDA speaker recognition system range between 1.265 -3.46 % on average for all developed systems, and (2) subjective evaluation: achieved 2.8 naturalness mean opinion score (MOS). Both objective and subjective experiments confirm the effectiveness of our proposed de-identification method.

No abstract

No abstract

No abstract