Previous studies have proven that integrating video signals, as a complementary modality, can facilitate improved performance for speech enhancement (SE). However, video clips usually contain large amounts of data and pose a high cost in terms of computational resources and thus may complicate the SE system. As an alternative source, a bone-conducted speech signal has a moderate data size while manifesting speech-phoneme structures, and thus complements its air-conducted counterpart. In this study, we propose a novel multi-modal SE structure in the time domain that leverages bone- and air-conducted signals. In addition, we examine two ensemble-learning-based strategies, early fusion (EF) and late fusion (LF), to integrate the two types of speech signals, and adopt a deep learning-based fully convolutional network to conduct the enhancement. The experiment results on the Mandarin corpus indicate that this newly presented multi-modal (integrating bone- and air-conducted signals) SE structure significantly outperforms the single-source SE counterparts (with a bone- or air-conducted signal only) in various speech evaluation metrics. In addition, the adoption of an LF strategy other than an EF in this novel SE multi-modal structure achieves better results.

Deep learning techniques such as convolutional neural networks (CNN) have been successfully applied to identify pathological voices. However, the major disadvantage of using these advanced models is the lack of interpretability in explaining the predicted outcomes. This drawback further introduces a bottleneck for promoting the classification or detection of voice-disorder systems, especially in this pandemic period. In this paper, we proposed using a series of learnable sinc functions to replace the very first layer of a commonly used CNN to develop an explainable SincNet system for classifying or detecting pathological voices. The applied sinc filters, a front-end signal processor in SincNet, are critical for constructing the meaningful layer and are directly used to extract the acoustic features for following networks to generate high-level voice information. We conducted our tests on three different Far Eastern Memorial Hospital voice datasets. From our evaluations, the proposed approach achieves the highest 7%–accuracy and 9%–sensitivity improvements from conventional methods and thus demonstrates superior performance in predicting input pathological waveforms of the SincNet system. More importantly, we intended to give possible explanations between the system output and the first-layer extracted speech features based on our evaluated results.

Goal: Numerous studies had successfully differentiated normal and abnormal voice samples. Nevertheless, further classification had rarely been attempted. This study proposes a novel approach, using continuous Mandarin speech instead of a single vowel, to classify four common voice disorders (i.e. functional dysphonia, neoplasm, phonotrauma, and vocal palsy). Methods: In the proposed framework, acoustic signals are transformed into mel-frequency cepstral coefficients, and a bi-directional long-short term memory network (BiLSTM) is adopted to model the sequential features. The experiments were conducted on a large-scale database, wherein 1,045 continuous speech were collected by the speech clinic of a hospital from 2012 to 2019. Results: Experimental results demonstrated that the proposed framework yields significant accuracy and unweighted average recall improvements of 78.12–89.27% and 50.92–80.68%, respectively, compared with systems that use a single vowel. Conclusions: The results are consistent with other machine learning algorithms, including gated recurrent units, random forest, deep neural networks, and LSTM.The sensitivities for each disorder were also analyzed, and the model capabilities were visualized via principal component analysis. An alternative experiment based on a balanced dataset again confirms the advantages of using continuous speech for learning voice disorders.

Distributed speech recognition (DSR) splits the processing of data between a mobile device and a network server. In the front-end, features are extracted and compressed to transmit over a wireless channel to a back-end server, where the incoming stream is received and reconstructed for recognition tasks. In this paper, we propose a feature compression algorithm termed suppression by selecting wavelets (SSW) for DSR: minimizing memory and device requirements while also maintaining or even improving the recognition performance. The SSW approach first applies the discrete wavelet transform (DWT) to filter the incoming speech feature sequence into two temporal sub-sequences at the client terminal. Feature compression is achieved by keeping the low (modulation) frequency sub-sequence while discarding the high frequency counterpart. The low-frequency sub-sequence is then transmitted across the remote network for specific feature statistics normalization. Wavelets are favorable for resolving the temporal properties of the feature sequence, and the down-sampling process in DWT reduces the amount of data at the terminal prior to transmission across the network, which can be interpreted as data compression. Once the compressed features have arrived at the server, the feature sequence can be enhanced by statistics normalization, reconstructed with inverse DWT, and compensated with a simple post filter to alleviate any over-smoothing effects from the compression stage. Results on a standard robustness task (Aurora-4) and on a Mandarin Chinese news corpus (MATBN) showed SSW outperforms conventional noise-robustness techniques while also providing nearly a 50% compression rate during the transmission stage of DSR systems.

Conventionally, the maximum likelihood (ML) criterion is applied to train a deep belief network (DBN). We present a maximum entropy (ME) learning algorithm for DBNs, designed specifically to handle limited training data. Maximizing only the entropy of parameters in the DBN allows more effective generalization capability, less bias towards data distributions, and robustness to over-fitting compared to ML learning. Results of text classification and object recognition tasks demonstrate ME-trained DBN outperforms ML-trained DBN when training data is limited.

The detection of audio tampering plays a crucial role in ensuring the authenticity and integrity of multimedia files. This paper presents a novel approach to identifying tampered audio files by leveraging the unique Electric Network Frequency (ENF) signal, which is inherent to the power grid and serves as a reliable indicator of authenticity. The study begins by establishing a comprehensive Chinese ENF database containing diverse ENF signals extracted from audio files. The proposed methodology involves extracting the ENF signal, applying wavelet decomposition, and utilizing the autoregressive model to train effective classification models. Subsequently, the framework is employed to detect audio tampering and assess the influence of various environmental conditions and recording devices on the ENF signal. Experimental evaluations conducted on our Chinese ENF database demonstrate the efficacy of the proposed method, achieving impressive accuracy rates ranging from 91% to 93%. The results emphasize the significance of ENF-based approaches in enhancing audio file forensics and reaffirm the necessity of adopting reliable tamper detection techniques in multimedia authentication.