Emotional speech recognition is generally considered more difficult than non-emotional speech recognition. The acoustic characteristics of emotional speech differ from those of non-emotional speech. Additionally, acoustic characteristics vary significantly depending on the type and intensity of emotions. Regarding linguistic features, emotional and colloquial expressions are also observed in their utterances. To solve these problems, we aim to improve recognition performance by adapting acoustic and language models to emotional speech. We used Japanese Twitter-based Emotional Speech (JTES) as an emotional speech corpus. This corpus consisted of tweets and had an emotional label assigned to each utterance. Corpus adaptation is possible using the utterances contained in this corpus. However, regarding the language model, the amount of adaptation data is insufficient. To solve this problem, we propose an adaptation of the language model by using online tweet data downloaded from the internet. The sentences used for adaptation were extracted from the tweet data based on certain rules. We extracted the data of 25.86 M words and used them for adaptation. In the recognition experiments, the baseline word error rate was 36.11%, whereas that with the acoustic and language model adaptation was 17.77%. The results demonstrated the effectiveness of the proposed method.

While deep image compression performs better than traditional codecs like JPEG on natural images, it faces a challenge as a learning-based approach: compression performance drastically decreases for out-of-domain images. To investigate this problem, we introduce a novel task that we call universal deep image compression, which involves compressing images in arbitrary domains, such as natural images, line drawings, and comics. Furthermore, we propose a content-adaptive optimization framework to tackle this task. This framework adapts a pre-trained compression model to each target image during testing for addressing the domain gap between pre-training and testing. For each input image, we insert adapters into the decoder of the model and optimize the latent representation extracted by the encoder and the adapter parameters in terms of rate-distortion, with the adapter parameters transmitted per image. To achieve the evaluation of the proposed universal deep compression, we constructed a benchmark dataset containing uncompressed images of four domains: natural images, line drawings, comics, and vector arts. We compare our proposed method with non-adaptive and existing adaptive compression methods, and the results show that our method outperforms them. Our code and dataset are publicly available at https://github.com/kktsubota/universal-dic.

We investigated the influence of horizontal shifts of the input images for one stage object detection method. We found that the object detector class scores drop when the target object center is at the grid boundary. Many approaches have focused on reducing the aliasing effect of down-sampling to achieve shift-invariance. However, down-sampling does not completely solve this problem at the grid boundary; it is necessary to suppress the dispersion of features in pixels close to the grid boundary into adjacent grid cells. Therefore, this paper proposes two approaches focused on the grid boundary to improve this weak point of current object detection methods. One is the Sub-Grid Feature Extraction Module, in which the sub-grid features are added to the input of the classification head. The other is Grid-Aware Data Augmentation, where augmented data are generated by the grid-level shifts and are used in training. The effectiveness of the proposed approaches is demonstrated using the COCO validation set after applying the proposed method to the FCOS architecture.

For a long time, the compensation effect of hearing aid is mainly evaluated subjectively, and there are fewer studies of objective evaluation. Furthermore, a pure speech signal is generally required as a reference in the existing objective evaluation methods, which restricts the practicality in a real-world environment. Therefore, this paper presents a non-intrusive speech quality evaluation method for hearing aid, which combines the audiogram and weighted frequency information. The proposed model mainly includes an audiogram information extraction network, a frequency information extraction network, and a quality score mapping network. The audiogram is the input of the audiogram information extraction network, which helps the system capture the information related to hearing loss. In addition, the low-frequency bands of speech contain loudness information and the medium and high-frequency components contribute to semantic comprehension. The information of two frequency bands is input to the frequency information extraction network to obtain time-frequency information. When obtaining the high-level features of different frequency bands and audiograms, they are fused into two groups of tensors that distinguish the information of different frequency bands and used as the input of the attention layer to calculate the corresponding weight distribution. Finally, a dense layer is employed to predict the score of speech quality. The experimental results show that it is reasonable to combine the audiogram and the weight of the information from two frequency bands, which can effectively realize the evaluation of the speech quality of the hearing aid.

This paper presents a channel operating margin (COM) based high-speed serial link optimization using machine learning (ML). COM that is proposed for evaluating serial link is calculated at first and during the calculation several important equalization parameters corresponding to the best configuration are extracted which can be used for the ML modeling of serial link. Then a deep neural network containing hidden layers are investigated to model a whole serial link equalization including transmitter feed forward equalizer (FFE), receiver continuous time linear equalizer (CTLE) and decision feedback equalizer (DFE). By training, validating and testing a lot of samples that meet the COM specification of 400GAUI-8 C2C, an effective ML model is generated and the maximum relative error is only 0.1 compared with computation results. At last 3 link configurations are discussed from the view of tradeoff between the link performance and cost, illustrating that our COM based ML modeling method can be applied to advanced serial link design for NRZ, PAM4 or even other higher level pulse amplitude modulation signal.

In recent years, deep neural networks (DNNs) have made a significant impact on a variety of research fields and applications. One drawback of DNNs is that it requires a huge amount of dataset for training. Since it is very expensive to ask experts to label the data, many non-expert data collection methods such as web crawling have been proposed. However, dataset created by non-experts often contain corrupted labels, and DNNs trained on such dataset are unreliable. Since DNNs have an enormous number of parameters, it tends to overfit to noisy labels, resulting in poor generalization performance. This problem is called Learning with Noisy labels (LNL). Recent studies showed that DNNs are robust to the noisy labels in the early stage of learning before over-fitting to noisy labels because DNNs learn the simple patterns first. Therefore DNNs tend to output true labels for samples with noisy labels in the early stage of learning, and the number of false predictions for samples with noisy labels is higher than for samples with clean labels. Based on these observations, we propose a new sample selection approach for LNL using the number of false predictions. Our method periodically collects the records of false predictions during training, and select samples with a low number of false predictions from the recent records. Then our method iteratively performs sample selection and training a DNNs model using the updated dataset. Since the model is trained with more clean samples and records more accurate false predictions for sample selection, the generalization performance of the model gradually increases. We evaluated our method on two benchmark datasets, CIFAR-10 and CIFAR-100 with synthetically generated noisy labels, and the obtained results which are better than or comparative to the-state-of-the-art approaches.

A new software based in-situ training (SBIST) method to achieve high accuracies is proposed for binarized neural networks inference accelerator chips in which measured offsets in sense amplifiers (activation binarizers) are transformed into biases in the training software. To expedite this individual training, the initial values for the weights are taken from results of a common forming training process which is conducted in advance by using the offset fluctuation distribution averaged over the fabrication line. SPICE simulation inference results for the accelerator predict that the accuracy recovers to higher than 90% even when the amplifier offset is as large as 40mV only after a few epochs of the individual training.

A backdoor attack is a type of attack method inducing deep neural network (DNN) misclassification. An adversary mixes poison data, which consist of images tampered with adversarial marks at specific locations and of adversarial target classes, into a training dataset. The backdoor model classifies only images with adversarial marks into an adversarial target class and other images into the correct classes. However, the attack performance degrades sharply when the location of the adversarial marks is slightly shifted. An adversarial mark that induces the misclassification of a DNN is usually applied when a picture is taken, so the backdoor attack will have difficulty succeeding in the physical world because the adversarial mark position fluctuates. This paper proposes a new approach in which an adversarial mark is applied using fault injection on the mobile industry processor interface (MIPI) between an image sensor and the image recognition processor. Two independent attack drivers are electrically connected to the MIPI data lane in our attack system. While almost all image signals are transferred from the sensor to the processor without tampering by canceling the attack signal between the two drivers, the adversarial mark is injected into a given location of the image signal by activating the attack signal generated by the two attack drivers. In an experiment, the DNN was implemented on a Raspberry pi 4 to classify MNIST handwritten images transferred from the image sensor over the MIPI. The adversarial mark successfully appeared in a specific small part of the MNIST images using our attack system. The success rate of the backdoor attack using this adversarial mark was 91%, which is much higher than the 18% rate achieved using conventional input image tampering.

This paper proposes two high-fidelity and computationally efficient neural voice conversion (VC) methods based on a direct waveform modification using spectral differentials. The conventional spectral-differential VC method with a minimum-phase filter achieves high-quality conversion for narrow-band (16 kHz-sampled) VC but requires heavy computational cost in filtering. This is because the minimum phase obtained using a fixed lifter of the Hilbert transform often results in a long-tap filter. Furthermore, when we extend the method to full-band (48 kHz-sampled) VC, the computational cost is heavy due to increased sampling points, and the converted-speech quality degrades due to large fluctuations in the high-frequency band. To construct a short-tap filter, we propose a lifter-training method for data-driven phase reconstruction that trains a lifter of the Hilbert transform by taking into account filter truncation. We also propose a frequency-band-wise modeling method based on sub-band multi-rate signal processing (sub-band modeling method) for full-band VC. It enhances the computational efficiency by reducing sampling points of signals converted with filtering and improves converted-speech quality by modeling only the low-frequency band. We conducted several objective and subjective evaluations to investigate the effectiveness of the proposed methods through implementation of the real-time, online, full-band VC system we developed, which is based on the proposed methods. The results indicate that 1) the proposed lifter-training method for narrow-band VC can shorten the tap length to 1/16 without degrading the converted-speech quality, and 2) the proposed sub-band modeling method for full-band VC can improve the converted-speech quality while reducing the computational cost, and 3) our real-time, online, full-band VC system can convert 48 kHz-sampled speech in real time attaining the converted speech with a 3.6 out of 5.0 mean opinion score of naturalness.

This work presents a DNN accelerator architecture specifically designed for performing efficient inference on compressed and sparse DNN models. Leveraging the data sparsity, a runtime processing scheme is proposed to deal with the encoded weights and activations directly in the compressed domain without decompressing. Furthermore, a new data flow is proposed to facilitate the reusage of input activations across the fully-connected (FC) layers. The proposed design is implemented and verified using the Xilinx Virtex-7 FPGA. Experimental results show it achieves 1.99×, 1.95× faster and 20.38×, 3.04× more energy efficient than CPU and mGPU platforms, respectively, running AlexNet.

Knowledge distillation approaches can transfer information from a large network (teacher network) to a small network (student network) to compress and accelerate deep neural networks. This paper proposes a novel knowledge distillation approach called multi-knowledge distillation (MKD). MKD consists of two stages. In the first stage, it employs autoencoders to learn compact and precise representations of the feature maps (FM) from the teacher network and the student network, these representations can be treated as the essential of the FM, i.e., EFM. In the second stage, MKD utilizes multiple kinds of knowledge, i.e., the magnitude of individual sample's EFM and the similarity relationships among several samples' EFM to enhance the generalization ability of the student network. Compared with previous approaches that employ FM or the handcrafted features from FM, the EFM learned from autoencoders can be transferred more efficiently and reliably. Furthermore, the rich information provided by the multiple kinds of knowledge guarantees the student network to mimic the teacher network as closely as possible. Experimental results also show that MKD is superior to the-state-of-arts.

A model extraction attack is a security issue in deep neural networks (DNNs). Information on a trained DNN model is an attractive target for an adversary not only in terms of intellectual property but also of security. Thus, an adversary tries to reveal the sensitive information contained in the trained DNN model from machine-learning services. Previous studies on model extraction attacks assumed that the victim provides a machine-learning cloud service and the adversary accesses the service through formal queries. However, when a DNN model is implemented on an edge device, adversaries can physically access the device and try to reveal the sensitive information contained in the implemented DNN model. We call these physical model extraction attacks model reverse-engineering (MRE) attacks to distinguish them from attacks on cloud services. Power side-channel analyses are often used in MRE attacks to reveal the internal operation from power consumption or electromagnetic leakage. Previous studies, including ours, evaluated MRE attacks against several types of DNN processors with power side-channel analyses. In this paper, information leakage from a systolic array which is used for the matrix multiplication unit in the DNN processors is evaluated. We utilized correlation power analysis (CPA) for the MRE attack and reveal weight parameters of a DNN model from the systolic array. Two types of the systolic array were implemented on field-programmable gate array (FPGA) to demonstrate that CPA reveals weight parameters from those systolic arrays. In addition, we applied an extended analysis approach called “chain CPA” for robust CPA analysis against the systolic arrays. Our experimental results indicate that an adversary can reveal trained model parameters from a DNN accelerator even if the DNN model parameters in the off-chip bus are protected with data encryption. Countermeasures against side-channel leaks will be important for implementing a DNN accelerator on a FPGA or application-specific integrated circuit (ASIC).

This paper presents a method for many-to-one voice conversion using phonetic posteriorgrams (PPGs) based on an adversarial training of deep neural networks (DNNs). A conventional method for many-to-one VC can learn a mapping function from input acoustic features to target acoustic features through separately trained DNN-based speech recognition and synthesis models. However, 1) the differences among speakers observed in PPGs and 2) an over-smoothing effect of generated acoustic features degrade the converted speech quality. Our method performs a domain-adversarial training of the recognition model for reducing the PPG differences. In addition, it incorporates a generative adversarial network into the training of the synthesis model for alleviating the over-smoothing effect. Unlike the conventional method, ours jointly trains the recognition and synthesis models so that they are optimized for many-to-one VC. Experimental evaluation demonstrates that the proposed method significantly improves the converted speech quality compared with conventional VC methods.

Some non-acoustic modalities have the ability to reveal certain speech attributes that can be used for synthesizing speech signals without acoustic signals. This study validated the use of ultrasonic Doppler frequency shifts caused by facial movements to implement a silent speech interface system. A 40kHz ultrasonic beam is incident to a speaker's mouth region. The features derived from the demodulated received signals were used to estimate the speech parameters. A nonlinear regression approach was employed in this estimation where the relationship between ultrasonic features and corresponding speech is represented by deep neural networks (DNN). In this study, we investigated the discrepancies between the ultrasonic signals of audible and silent speech to validate the possibility for totally silent communication. Since reference speech signals are not available in silently mouthed ultrasonic signals, a nearest-neighbor search and alignment method was proposed, wherein alignment was achieved by determining the optimal pair of ultrasonic and audible features in the sense of a minimum mean square error criterion. The experimental results showed that the performance of the ultrasonic Doppler-based method was superior to that of EMG-based speech estimation, and was comparable to an image-based method.

Conventional approaches to statistical parametric speech synthesis use context-dependent hidden Markov models (HMMs) clustered using decision trees to generate speech parameters from linguistic features. However, decision trees are not always appropriate to model complex context dependencies of linguistic features efficiently. An alternative scheme that replaces decision trees with deep neural networks (DNNs) was presented as a possible way to overcome the difficulty. By training the network to represent high-dimensional feedforward dependencies from linguistic features to acoustic features, DNN-based speech synthesis systems convert a text into a speech. To improved the naturalness of the synthesized speech, this paper presents a novel pre-training method for DNN-based statistical parametric speech synthesis systems. In our method, a deep relational model (DRM), which represents a joint probability of two visible variables, is applied to describe the joint distribution of acoustic and linguistic features. As with DNNs, a DRM consists several hidden layers and two visible layers. Although DNNs represent feedforward dependencies from one visible variables (inputs) to other visible variables (outputs), a DRM has an ability to represent the bidirectional dependencies between two visible variables. During the maximum-likelihood (ML) -based training, the model optimizes its parameters (connection weights between two adjacent layers, and biases) of a deep architecture considering the bidirectional conversion between 1) acoustic features given linguistic features, and 2) linguistic features given acoustic features generated from itself. Owing to considering whether the generated acoustic features are recognizable, our method can obtain reasonable parameters for speech synthesis. Experimental results in a speech synthesis task show that pre-trained DNN-based systems using our proposed method outperformed randomly-initialized DNN-based systems, especially when the amount of training data is limited. Additionally, speaker-dependent speech recognition experimental results also show that our method outperformed DNN-based systems, by setting the initial parameters of our method are the same as that in the synthesis experiments.

This paper presents recognition of anomalously deformed Kana sequences in Japanese historical documents, for which a contest was held by IEICE PRMU 2017. The contest was divided into three levels in accordance with the number of characters to be recognized: level 1: single characters, level 2: sequences of three vertically written Kana characters, and level 3: unrestricted sets of characters composed of three or more characters possibly in multiple lines. This paper focuses on the methods for levels 2 and 3 that won the contest. We basically follow the segmentation-free approach and employ the hierarchy of a Convolutional Neural Network (CNN) for feature extraction, Bidirectional Long Short-Term Memory (BLSTM) for frame prediction, and Connectionist Temporal Classification (CTC) for text recognition, which is named a Deep Convolutional Recurrent Network (DCRN). We compare the pretrained CNN approach and the end-to-end approach with more detailed variations for level 2. Then, we propose a method of vertical text line segmentation and multiple line concatenation before applying DCRN for level 3. We also examine a two-dimensional BLSTM (2DBLSTM) based method for level 3. We present the evaluation of the best methods by cross validation. We achieved an accuracy of 89.10% for the three-Kana-character sequence recognition and an accuracy of 87.70% for the unrestricted Kana recognition without employing linguistic context. These results prove the performances of the proposed models on the level 2 and 3 tasks.

A monaural speech enhancement method combining deep neural network (DNN) with low rank analysis and speech present probability is proposed in this letter. Low rank and sparse analysis is first applied on the noisy speech spectrogram to get the approximate low rank representation of noise. Then a joint feature training strategy for DNN based speech enhancement is presented, which helps the DNN better predict the target speech. To reduce the residual noise in highly overlapping regions and high frequency domain, speech present probability (SPP) weighted post-processing is employed to further improve the quality of the speech enhanced by trained DNN model. Compared with the supervised non-negative matrix factorization (NMF) and the conventional DNN method, the proposed method obtains improved speech enhancement performance under stationary and non-stationary conditions.

Microblog data contains rich information of real-world events with great commercial values, so microblog-oriented natural language processing (NLP) tasks have grabbed considerable attention of researchers. However, the performance of microblog-oriented Chinese Word Segmentation (CWS) based on deep neural networks (DNNs) is still not satisfying. One critical reason is that the existing microblog-oriented training corpus is inadequate to train effective weight matrices for DNNs. In this paper, we propose a novel active learning method to extend the scale of the training corpus for DNNs. However, due to a large amount of partially overlapped sentences in the microblogs, it is difficult to select samples with high annotation values from raw microblogs during the active learning procedure. To select samples with higher annotation values, parameter λ is introduced to control the number of repeatedly selected samples. Meanwhile, various strategies are adopted to measure the overall annotation values of a sample during the active learning procedure. Experiments on the benchmark datasets of NLPCC 2015 show that our λ-active learning method outperforms the baseline system and the state-of-the-art method. Besides, the results also demonstrate that the performances of the DNNs trained on the extended corpus are significantly improved.

In this study, toward the development of smartphone-based monitoring system for life logging, we collect over 1,400 hours of data by recording including both the outdoor and indoor daily activities of 19 subjects, under practical conditions with a smartphone and a small camera. We then construct a huge human activity database which consists of an environmental sound signal, triaxial acceleration signals and manually annotated activity tags. Using our constructed database, we evaluate the activity recognition performance of deep neural networks (DNNs), which have achieved great performance in various fields, and apply DNN-based adaptation techniques to improve the performance with only a small amount of subject-specific training data. We experimentally demonstrate that; 1) the use of multi-modal signal, including environmental sound and triaxial acceleration signals with a DNN is effective for the improvement of activity recognition performance, 2) the DNN can discriminate specified activities from a mixture of ambiguous activities, and 3) DNN-based adaptation methods are effective even if only a small amount of subject-specific training data is available.

In this letter, we explore joint optimization of perceptual gain function and deep neural networks (DNNs) for a single-channel speech enhancement task. A DNN architecture is proposed which incorporates the masking properties of the human auditory system to make the residual noise inaudible. This new DNN architecture directly trains a perceptual gain function which is used to estimate the magnitude spectrum of clean speech from noisy speech features. Experimental results demonstrate that the proposed speech enhancement approach can achieve significant improvements over the baselines when tested with TIMIT sentences corrupted by various types of noise, no matter whether the noise conditions are included in the training set or not.