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.

The recurrent neural network (RNN) has been used in audio and speech processing, such as language translation and speech recognition. Although RNN-based architecture can be applied to speech synthesis, the long computing time is still the primary concern. This research proposes a fast gated recurrent neural network, a fast RNN-based architecture, for speech synthesis based on the minimal gated unit (MGU). Our architecture removes the unit state history from some equations in MGU. Our MGU-based architecture is about twice faster, with equally good sound quality than the other MGU-based architectures.

This paper describes a method to control prosodic features using phonetic and prosodic symbols as input of attention-based sequence-to-sequence (seq2seq) acoustic modeling (AM) for neural text-to-speech (TTS). The method involves inserting a sequence of prosodic symbols between phonetic symbols that are then used to reproduce prosodic acoustic features, i.e. accents, pauses, accent breaks, and sentence endings, in several seq2seq AM methods. The proposed phonetic and prosodic labels have simple descriptions and a low production cost. By contrast, the labels of conventional statistical parametric speech synthesis methods are complicated, and the cost of time alignments such as aligning the boundaries of phonemes is high. The proposed method does not need the boundary positions of phonemes. We propose an automatic conversion method for conventional labels and show how to automatically reproduce pitch accents and phonemes. The results of objective and subjective evaluations show the effectiveness of our method.

As deep learning-based speech recognition systems are spotlighted, the need for large-scale speech databases for acoustic model training is increasing. Broadcast data can be easily used for database construction, since it contains transcripts for the hearing impaired. However, the subtitle timestamps have not been used to extract speech data because they are often inaccurate due to the inherent characteristics of closed captioning. Thus, we propose to build a large-scale speech database from multi-genre broadcast data with inaccurate subtitle timestamps. The proposed method first extracts the most likely speech intervals by removing subtitle texts with low subtitle quality index, concatenating adjacent subtitle texts into a merged subtitle text, and adding a margin to the timestamp of the merged subtitle text. Next, a speech recognizer is used to extract a hypothesis text of a speech segment corresponding to the merged subtitle text, and then the hypothesis text obtained from the decoder is recursively aligned with the merged subtitle text. Finally, the speech database is constructed by selecting the sub-parts of the merged subtitle text that match the hypothesis text. Our method successfully refines a large amount of broadcast data with inaccurate subtitle timestamps, taking about half of the time compared with the previous methods. Consequently, our method is useful for broadcast data processing, where bulk speech data can be collected every hour.

Deep neural networks (DNNs) have achieved significant success in the field of automatic speech recognition. One main advantage of DNNs is automatic feature extraction without human intervention. However, adaptation under limited available data remains a major challenge for DNN-based systems because of their enormous free parameters. In this paper, we propose a filterbank-incorporated DNN that incorporates a filterbank layer that presents the filter shape/center frequency and a DNN-based acoustic model. The filterbank layer and the following networks of the proposed model are trained jointly by exploiting the advantages of the hierarchical feature extraction, while most systems use pre-defined mel-scale filterbank features as input acoustic features to DNNs. Filters in the filterbank layer are parameterized to represent speaker characteristics while minimizing a number of parameters. The optimization of one type of parameters corresponds to the Vocal Tract Length Normalization (VTLN), and another type corresponds to feature-space Maximum Linear Likelihood Regression (fMLLR) and feature-space Discriminative Linear Regression (fDLR). Since the filterbank layer consists of just a few parameters, it is advantageous in adaptation under limited available data. In the experiment, filterbank-incorporated DNNs showed effectiveness in speaker/gender adaptations under limited adaptation data. Experimental results on CSJ task demonstrate that the adaptation of proposed model showed 5.8% word error reduction ratio with 10 utterances against the un-adapted model.

Deep neural network (DNN)-based speech synthesis can produce more natural synthesized speech than the conventional HMM-based speech synthesis. However, it is not revealed whether the synthesized speech quality can be improved by utilizing a multi-speaker speech corpus. To address this problem, this paper proposes DNN-based speech synthesis using speaker codes as a method to improve the performance of the conventional speaker dependent DNN-based method. In order to model speaker variation in the DNN, the augmented feature (speaker codes) is fed to the hidden layer(s) of the conventional DNN. This paper investigates the effectiveness of introducing speaker codes to DNN acoustic models for speech synthesis for two tasks: multi-speaker modeling and speaker adaptation. For the multi-speaker modeling task, the method we propose trains connection weights of the whole DNN using a multi-speaker speech corpus. When performing multi-speaker synthesis, the speaker code corresponding to the selected target speaker is fed to the DNN to generate the speaker's voice. When performing speaker adaptation, a set of connection weights of the multi-speaker model is re-estimated to generate a new target speaker's voice. We investigated the relationship between the prediction performance and architecture of the DNNs through objective measurements. Objective evaluation experiments revealed that the proposed model outperformed conventional methods (HMMs, speaker dependent DNNs and multi-speaker DNNs based on a shared hidden layer structure). Subjective evaluation experimental results showed that the proposed model again outperformed the conventional methods (HMMs, speaker dependent DNNs), especially when using a small number of target speaker utterances.

We propose a new method to combine multiple acoustic models in Gaussian mixture model (GMM) spaces for robust speech recognition. Even though large vocabulary continuous speech recognition (LVCSR) systems are recently widespread, they often make egregious recognition errors resulting from unavoidable mismatch of speaking styles or environments between the training and real conditions. To handle this problem, a multi-style training approach has been used conventionally to train a large acoustic model by using a large speech database with various kinds of speaking styles and environment noise. But, in this work, we combine multiple sub-models trained for different speaking styles or environment noise into a large acoustic model by maximizing the log-likelihood of the sub-model states sharing the same phonetic context and position. Then the combined acoustic model is used in a new target system, which is robust to variation in speaking style and diverse environment noise. Experimental results show that the proposed method significantly outperforms the conventional methods in two tasks: Non-native English speech recognition for second-language learning systems and noise-robust point-of-interest (POI) recognition for car navigation systems.

The paper addresses a scheme of lightly supervised training of an acoustic model, which exploits a large amount of data with closed caption texts but not faithful transcripts. In the proposed scheme, a sequence of the closed caption text and that of the ASR hypothesis by the baseline system are aligned. Then, a set of dedicated classifiers is designed and trained to select the correct one among them or reject both. It is demonstrated that the classifiers can effectively filter the usable data for acoustic model training. The scheme realizes automatic training of the acoustic model with an increased amount of data. A significant improvement in the ASR accuracy is achieved from the baseline system and also in comparison with the conventional method of lightly supervised training based on simple matching.

This paper proposes a novel approach for integrating spectral feature extraction and acoustic modeling in hidden Markov model (HMM) based speech synthesis. The statistical modeling process of speech waveforms is typically divided into two component modules: the frame-by-frame feature extraction module and the acoustic modeling module. In the feature extraction module, the statistical mel-cepstral analysis technique has been used and the objective function is the likelihood of mel-cepstral coefficients for given speech waveforms. In the acoustic modeling module, the objective function is the likelihood of model parameters for given mel-cepstral coefficients. It is important to improve the performance of each component module for achieving higher quality synthesized speech. However, the final objective of speech synthesis systems is to generate natural speech waveforms from given texts, and the improvement of each component module does not always lead to the improvement of the quality of synthesized speech. Therefore, ideally all objective functions should be optimized based on an integrated criterion which well represents subjective speech quality of human perception. In this paper, we propose an approach to model speech waveforms directly and optimize the final objective function. Experimental results show that the proposed method outperformed the conventional methods in objective and subjective measures.

This paper proposes an acoustic modeling technique based on Bayesian framework using multiple model structures for speech recognition. The aim of the Bayesian approach is to obtain good prediction of observation by marginalizing all variables related to generative processes. Although the effectiveness of marginalizing model parameters was recently reported in speech recognition, most of these systems use only “one” model structure, e.g., topologies of HMMs, the number of states and mixtures, types of state output distributions, and parameter tying structures. However, it is insufficient to represent a true model distribution, because a family of such models usually does not include a true distribution in most practical cases. One of solutions of this problem is to use multiple model structures. Although several approaches using multiple model structures have already been proposed, the consistent integration of multiple model structures based on the Bayesian approach has not seen in speech recognition. This paper focuses on integrating multiple phonetic decision trees based on the Bayesian framework in HMM based acoustic modeling. The proposed method is derived from a new marginal likelihood function which includes the model structures as a latent variable in addition to HMM state sequences and model parameters, and the posterior distributions of these latent variables are obtained using the variational Bayesian method. Furthermore, to improve the optimization algorithm, the deterministic annealing EM (DAEM) algorithm is applied to the training process. The proposed method effectively utilizes multiple model structures, especially in the early stage of training and this leads to better predictive distributions and improvement of recognition performance.

We propose an active learning framework for speech recognition that reduces the amount of data required for acoustic modeling. This framework consists of two steps. We first obtain a phone-error distribution using an acoustic model estimated from transcribed speech data. Then, from a text corpus we select a sentence whose phone-occurrence distribution is close to the phone-error distribution and collect its speech data. We repeat this process to increase the amount of transcribed speech data. We applied this framework to speaker adaptation and acoustic model training. Our evaluation results showed that it significantly reduced the amount of transcribed data while maintaining the same level of accuracy.

A novel speech feature generation-based acoustic model training method for robust speaker-independent speech recognition is proposed. For decades, speaker adaptation methods have been widely used. All of these adaptation methods need adaptation data. However, our proposed method aims to create speaker-independent acoustic models that cover not only known but also unknown speakers. We achieve this by adopting inverse maximum likelihood linear regression (MLLR) transformation-based feature generation, and then we train our models using these features. First we obtain MLLR transformation matrices from a limited number of existing speakers. Then we extract the bases of the MLLR transformation matrices using PCA. The distribution of the weight parameters to express the transformation matrices for the existing speakers are estimated. Next, we construct pseudo-speaker transformations by sampling the weight parameters from the distribution, and apply the transformation to the normalized features of the existing speaker to generate the features of the pseudo-speakers. Finally, using these features, we train the acoustic models. Evaluation results show that the acoustic models trained using our proposed method are robust for unknown speakers.

In this paper, a logarithmic adaptive quantization projection (LAQP) algorithm for digital watermarking is proposed. Conventional quantization index modulation uses a fixed quantization step in the watermarking embedding procedure, which leads to poor fidelity. Moreover, the conventional methods are sensitive to value-metric scaling attack. The LAQP method combines the quantization projection scheme with a perceptual model. In comparison to some conventional quantization methods with a perceptual model, the LAQP only needs to calculate the perceptual model in the embedding procedure, avoiding the decoding errors introduced by the difference of the perceptual model used in the embedding and decoding procedure. Experimental results show that the proposed watermarking scheme keeps a better fidelity and is robust against the common signal processing attack. More importantly, the proposed scheme is invariant to value-metric scaling attack.

This paper proposes likelihood smoothing techniques to improve decision tree-based acoustic models, where decision trees are used as replacements for Gaussian mixture models to compute the observation likelihoods for a given HMM state in a speech recognition system. Decision trees have a number of advantageous properties, such as not imposing restrictions on the number or types of features, and automatically performing feature selection. This paper describes basic configurations of decision tree-based acoustic models and proposes two methods to improve the robustness of the basic model: DT mixture models and soft decisions for continuous features. Experimental results for the Aurora 2 speech database show that a system using decision trees offers state-of-the-art performance, even without taking advantage of its full potential and soft decisions improve the performance of DT-based acoustic models with 16.8% relative error rate reduction over hard decisions.

In this paper, we propose a hybrid model adaptation approach in which pronunciation and acoustic models are adapted by incorporating the pronunciation and acoustic variabilities of non-native speech in order to improve the performance of non-native automatic speech recognition (ASR). Specifically, the proposed hybrid model adaptation can be performed at either the state-tying or triphone-modeling level, depending at which acoustic model adaptation is performed. In both methods, we first analyze the pronunciation variant rules of non-native speakers and then classify each rule as either a pronunciation variant or an acoustic variant. The state-tying level hybrid method then adapts pronunciation models and acoustic models by accommodating the pronunciation variants in the pronunciation dictionary and by clustering the states of triphone acoustic models using the acoustic variants, respectively. On the other hand, the triphone-modeling level hybrid method initially adapts pronunciation models in the same way as in the state-tying level hybrid method; however, for the acoustic model adaptation, the triphone acoustic models are then re-estimated based on the adapted pronunciation models and the states of the re-estimated triphone acoustic models are clustered using the acoustic variants. From the Korean-spoken English speech recognition experiments, it is shown that ASR systems employing the state-tying and triphone-modeling level adaptation methods can relatively reduce the average word error rates (WERs) by 17.1% and 22.1% for non-native speech, respectively, when compared to a baseline ASR system.

Statistical speech recognition using continuous-density hidden Markov models (CDHMMs) has yielded many practical applications. However, in general, mismatches between the training data and input data significantly degrade recognition accuracy. Various acoustic model adaptation techniques using a few input utterances have been employed to overcome this problem. In this article, we survey these adaptation techniques, including maximum a posteriori (MAP) estimation, maximum likelihood linear regression (MLLR), and eigenvoice. We also present a schematic view called the adaptation pyramid to illustrate how these methods relate to each other.

We present a new discriminative method of acoustic model adaptation that deals with a task-dependent speech variability. We have focused on differences of expressions or speaking styles between tasks and set the objective of this method as improving the recognition accuracy of indistinctly pronounced phrases dependent on a speaking style. The adaptation appends subword models for frequently observable variants of subwords in the task. To find the task-dependent variants, low-confidence words are statistically selected from words with higher frequency in the task's adaptation data by using their word lattices. HMM parameters of subword models dependent on the words are discriminatively trained by using linear transforms with a minimum phoneme error (MPE) criterion. For the MPE training, subword accuracy discriminating between the variants and the originals is also investigated. In speech recognition experiments, the proposed adaptation with the subword variants reduced the word error rate by 12.0% relative in a Japanese conversational broadcast task.

In this paper, we propose an acoustic model that is robust to multiple noise environments, as well as a method for adapting the acoustic model to an environment to improve the model. The model is called "the multi-mixture model," which is based on a mixture of different HMMs each of which is trained using speech under different noise conditions. Speech recognition experiments showed that the proposed model performs better than the conventional multi-condition model. The method for adaptation is based on the aspect model, which is a "mixture-of-mixture" model. To realize adaptation using extremely small amount of adaptation data (i.e., a few seconds), we train a small number of mixture models, which can be interpreted as models for "clusters" of noise environments. Then, the models are mixed using weights, which are determined according to the adaptation data. The experimental results showed that the adaptation based on the aspect model improved the word accuracy in a heavy noise environment and showed no performance deterioration for all noise conditions, while the conventional methods either did not improve the performance or showed both improvement and degradation of recognition performance according to noise conditions.

This paper proposes an efficient FFT algorithm for the Psycho-Acoustic Model (PAM) of MPEG-4 AAC. The proposed algorithm synthesizes FFT coefficients using MDCT and MDST coefficients through circular convolution. The complexity of the MDCT and MDST coefficients is approximately half of the original FFT. We also design a new PAM based on the proposed FFT algorithm, which has 15% lower computational complexity than the original PAM without degradation of sound quality. Subjective as well as objective test results are presented to confirm the efficiency of the proposed FFT computation algorithm and the PAM.

For the digital watermarking patchwork algorithm originally given by Bender et al., this paper proposes two improvements applicable to audio watermarking. First, the watermark embedding strength is psychoacoustically adapted, using the Bark frequency scale. Second, whereas previous approaches leave the samples that do not correspond to the data untouched, in this paper, these are modified to reduce the probability of misdetection, a method called full index embedding. In simulations, the proposed combination of these two proposed methods has higher resistance to a variety of attacks than prior algorithms.