In this paper, we propose a new training framework named the INmfCA algorithm for nonparallel voice conversion (VC) systems. To train conversion models, traditional VC frameworks require parallel corpora, in which source and target speakers utter the same linguistic contents. Although the frameworks have achieved high-quality VC, they are not applicable in situations where parallel corpora are unavailable. To acquire conversion models without parallel corpora, nonparallel methods are widely studied. Although the frameworks achieve VC under nonparallel conditions, they tend to require huge background knowledge or many training utterances. This is because of difficulty in disentangling linguistic and speaker information without a large amount of data. In this work, we tackle this problem by exploiting NMF, which can factorize acoustic features into time-variant and time-invariant components in an unsupervised manner. The method acquires alignment between the acoustic features of a source speaker's utterances and a target dictionary and uses the obtained alignment as activation of NMF to train the source speaker's dictionary without parallel corpora. The acquisition method is based on the INCA algorithm, which obtains the alignment of nonparallel corpora. In contrast to the INCA algorithm, the alignment is not restricted to observed samples, and thus the proposed method can efficiently utilize small nonparallel corpora. The results of subjective experiments show that the combination of the proposed algorithm and the INCA algorithm outperformed not only an INCA-based nonparallel framework but also CycleGAN-VC, which performs nonparallel VC without any additional training data. The results also indicate that a one-shot VC framework, which does not need to train source speakers, can be constructed on the basis of the proposed method.

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 paper proposes a voice conversion (VC) method based on a model that links linguistic and acoustic representations via latent phonological distinctive features. Our method, called speech chain VC, is inspired by the concept of the speech chain, where speech communication consists of a chain of events linking the speaker's brain with the listener's brain. We assume that speaker identity information, which appears in the acoustic level, is embedded in two steps — where phonological information is encoded into articulatory movements (linguistic to physiological) and where articulatory movements generate sound waves (physiological to acoustic). Speech chain VC represents these event links by using an adaptive restricted Boltzmann machine (ARBM) introducing phoneme labels and acoustic features as two classes of visible units and latent phonological distinctive features associated with articulatory movements as hidden units. Subjective evaluation experiments showed that intelligibility of the converted speech significantly improved compared with the conventional ARBM-based method. The speaker-identity conversion quality of the proposed method was comparable to that of a Gaussian mixture model (GMM)-based method. Analyses on the representations of the hidden layer of the speech chain VC model supported that some of the hidden units actually correspond to phonological distinctive features. Final part of this paper proposes approaches to achieve one-shot VC by using the speech chain VC model. Subjective evaluation experiments showed that when a target speaker is the same gender as a source speaker, the proposed methods can achieve one-shot VC based on each single source and target speaker's utterance.

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.

This paper describes a novel approach to flexible control of speaker characteristics using tensor representation of multiple Gaussian mixture models (GMM). In voice conversion studies, realization of conversion from/to an arbitrary speaker's voice is one of the important objectives. For this purpose, eigenvoice conversion (EVC) based on an eigenvoice GMM (EV-GMM) was proposed. In the EVC, a speaker space is constructed based on GMM supervectors which are high-dimensional vectors derived by concatenating the mean vectors of each of the speaker GMMs. In the speaker space, each speaker is represented by a small number of weight parameters of eigen-supervectors. In this paper, we revisit construction of the speaker space by introducing the tensor factor analysis of training data set. In our approach, each speaker is represented as a matrix of which the row and the column respectively correspond to the dimension of the mean vector and the Gaussian component. The speaker space is derived by the tensor factor analysis of the set of the matrices. Our approach can solve an inherent problem of supervector representation, and it improves the performance of voice conversion. In addition, in this paper, effects of speaker adaptive training before factorization are also investigated. Experimental results of one-to-many voice conversion demonstrate the effectiveness of the proposed approach.

To improve the likability of speech, we propose a voice conversion algorithm by controlling the fundamental frequency (F0) and the spectral envelope and carry out a subjective evaluation. The subjects can manipulate these two speech parameters. From the result, the subjects preferred speech with a parameter related to higher brightness.

One of the problems associated with voice conversion from a nonparallel corpus is how to find the best match or alignment between the source and the target vector sequences without linguistic information. In a previous study, alignment was achieved by minimizing the distance between the source vector and the transformed vector. This method, however, yielded a sequence of feature vectors that were not well matched with the underlying speaker model. In this letter, the vectors were selected from the candidates by maximizing the overall likelihood of the selected vectors with respect to the target model in the HMM context. Both objective and subjective evaluations were carried out using the CMU ARCTIC database to verify the effectiveness of the proposed method.

This paper proposes Deep Neural Network (DNN)-based Voice Conversion (VC) using input-to-output highway networks. VC is a speech synthesis technique that converts input features into output speech parameters, and DNN-based acoustic models for VC are used to estimate the output speech parameters from the input speech parameters. Given that the input and output are often in the same domain (e.g., cepstrum) in VC, this paper proposes a VC using highway networks connected from the input to output. The acoustic models predict the weighted spectral differentials between the input and output spectral parameters. The architecture not only alleviates over-smoothing effects that degrade speech quality, but also effectively represents the characteristics of spectral parameters. The experimental results demonstrate that the proposed architecture outperforms Feed-Forward neural networks in terms of the speech quality and speaker individuality of the converted speech.

As one of the techniques enabling individual singers to produce the varieties of voice timbre beyond their own physical constraints, a statistical voice timbre control technique based on the perceived age has been developed. In this technique, the perceived age of a singing voice, which is the age of the singer as perceived by the listener, is used as one of the intuitively understandable measures to describe voice characteristics of the singing voice. The use of statistical voice conversion (SVC) with a singer-dependent multiple-regression Gaussian mixture model (MR-GMM), which effectively models the voice timbre variations caused by a change of the perceived age, makes it possible for individual singers to manipulate the perceived ages of their own singing voices while retaining their own singer identities. However, there still remain several issues; e.g., 1) a controllable range of the perceived age is limited; 2) quality of the converted singing voice is significantly degraded compared to that of a natural singing voice; and 3) each singer needs to sing the same phrase set as sung by a reference singer to develop the singer-dependent MR-GMM. To address these issues, we propose the following three methods; 1) a method using gender-dependent modeling to expand the controllable range of the perceived age; 2) a method using direct waveform modification based on spectrum differential to improve quality of the converted singing voice; and 3) a rapid unsupervised adaptation method based on maximum a posteriori (MAP) estimation to easily develop the singer-dependent MR-GMM. The experimental results show that the proposed methods achieve a wider controllable range of the perceived age, a significant quality improvement of the converted singing voice, and the development of the singer-dependnet MR-GMM using only a few arbitrary phrases as adaptation data.

This paper describes a novel statistical bandwidth extension (BWE) technique based on a Gaussian mixture model (GMM) and a sub-band basis spectrum model (SBM), in which each dimensional component represents a specific acoustic space in the frequency domain. The proposed method can achieve the BWE from speech data with an arbitrary frequency bandwidth whereas the conventional methods perform the conversion from fixed narrow-band data. In the proposed method, we train a GMM with SBM parameters extracted from full-band spectra in advance. According to the bandwidth of input signal, the trained GMM is reconstructed to the GMM of the joint probability density between low-band SBM and high-band SBM components. Then high-band SBM components are estimated from low-band SBM components of the input signal based on the reconstructed GMM. Finally, BWE is achieved by adding the spectra decoded from estimated high-band SBM components to the ones of the input signal. To construct the full-band signal from the narrow-band one, we apply this method to log-amplitude spectra and aperiodic components. Objective and subjective evaluation results show that the proposed method extends the bandwidth of speech data robustly for the log-amplitude spectra. Experimental results also indicate that the aperiodic component extracted from the upsampled narrow-band signal realizes the same performance as the restored and the full-band aperiodic components in the proposed method.

This paper presents a novel statistical sample-based approach for Gaussian Mixture Model (GMM)-based Voice Conversion (VC). Although GMM-based VC has the promising flexibility of model adaptation, quality in converted speech is significantly worse than that of natural speech. This paper addresses the problem of inaccurate modeling, which is one of the main reasons causing the quality degradation. Recently, we have proposed statistical sample-based speech synthesis using rich context models for high-quality and flexible Hidden Markov Model (HMM)-based Text-To-Speech (TTS) synthesis. This method makes it possible not only to produce high-quality speech by introducing ideas from unit selection synthesis, but also to preserve flexibility of the original HMM-based TTS. In this paper, we apply this idea to GMM-based VC. The rich context models are first trained for individual joint speech feature vectors, and then we gather them mixture by mixture to form a Rich context-GMM (R-GMM). In conversion, an iterative generation algorithm using R-GMMs is used to convert speech parameters, after initialization using over-trained probability distributions. Because the proposed method utilizes individual speech features, and its formulation is the same as that of conventional GMM-based VC, it makes it possible to produce high-quality speech while keeping flexibility of the original GMM-based VC. The experimental results demonstrate that the proposed method yields significant improvements in term of speech quality and speaker individuality in converted speech.

Most of the current voice conversion methods are conducted based on parallel speech, which is not easily obtained in practice. In this letter, a novel iterative speaker model alignment (ISMA) method is proposed to address this problem. First, the source and target speaker models are each trained from the background model by adopting maximum a posteriori (MAP) algorithm. Then, a novel ISMA method is presented for alignment and transformation of spectral features. Finally, the proposed ISMA approach is further combined with a Gaussian mixture model (GMM) to improve the conversion performance. A series of objective and subjective experiments are carried out on CMU ARCTIC dataset, and the results demonstrate that the proposed method significantly outperforms the state-of-the-art approach.

In this paper, we use neural networks (NNs) for cross-dialectal (Mandarin-Shanghainese) voice conversion using a bi-dialectal speakers' recordings. This system employs a nonlinear mapping function, which is trained by parallel mandarin features of source and target speakers, to convert source speaker's Shanghainese features to those of target speaker. This study investigates three training aspects: a) Frequency warping, which is supposed to be language independent; b) Pre-training, which drives weights to a better starting point than random initialization or be regarded as unsupervised feature learning; and c) Sequence training, which minimizes sequence-level errors and matches objectives used in training and converting. Experimental results show that the performance of cross-dialectal voice conversion is close to that of intra-dialectal. This benefit is likely from the strong learning capabilities of NNs, e.g., exploiting feature correlations between fundamental frequency (F0) and spectrum. The objective measures: log spectral distortion (LSD) and root mean squared error (RMSE) of F0, both show that pre-training and sequence training outperform the frame-level mean square error (MSE) training. The naturalness of the converted Shanghainese speech and the similarity between converted Shanghainese speech and target Mandarin speech are significantly improved.

The perceived age of a singing voice is the age of the singer as perceived by the listener, and is one of the notable characteristics that determines perceptions of a song. In this paper, we describe an investigation of acoustic features that have an effect on the perceived age, and a novel voice timbre control technique based on the perceived age for singing voice conversion (SVC). Singers can sing expressively by controlling prosody and voice timbre, but the varieties of voices that singers can produce are limited by physical constraints. Previous work has attempted to overcome this limitation through the use of statistical voice conversion. This technique makes it possible to convert singing voice timbre of an arbitrary source singer into those of an arbitrary target singer. However, it is still difficult to intuitively control singing voice characteristics by manipulating parameters corresponding to specific physical traits, such as gender and age. In this paper, we first perform an investigation of the factors that play a part in the listener's perception of the singer's age at first. Then, we applied a multiple-regression Gaussian mixture models (MR-GMM) to SVC for the purpose of controlling voice timbre based on the perceived age and we propose SVC based on the modified MR-GMM for manipulating the perceived age while maintaining singer's individuality. The experimental results show that 1) the perceived age of singing voices corresponds relatively well to the actual age of the singer, 2) prosodic features have a larger effect on the perceived age than spectral features, 3) the individuality of a singer is influenced more heavily by segmental features than prosodic features 4) the proposed voice timbre control method makes it possible to change the singer's perceived age while not having an adverse effect on the perceived individuality.

This paper presents a voice conversion technique using speaker-dependent Restricted Boltzmann Machines (RBM) to build high-order eigen spaces of source/target speakers, where it is easier to convert the source speech to the target speech than in the traditional cepstrum space. We build a deep conversion architecture that concatenates the two speaker-dependent RBMs with neural networks, expecting that they automatically discover abstractions to express the original input features. Under this concept, if we train the RBMs using only the speech of an individual speaker that includes various phonemes while keeping the speaker individuality unchanged, it can be considered that there are fewer phonemes and relatively more speaker individuality in the output features of the hidden layer than original acoustic features. Training the RBMs for a source speaker and a target speaker, we can then connect and convert the speaker individuality abstractions using Neural Networks (NN). The converted abstraction of the source speaker is then back-propagated into the acoustic space (e.g., MFCC) using the RBM of the target speaker. We conducted speaker-voice conversion experiments and confirmed the efficacy of our method with respect to subjective and objective criteria, comparing it with the conventional Gaussian Mixture Model-based method and an ordinary NN.

This paper presents a voice conversion (VC) technique for noisy environments based on a sparse representation of speech. Sparse representation-based VC using Non-negative matrix factorization (NMF) is employed for noise-added spectral conversion between different speakers. In our previous exemplar-based VC method, source exemplars and target exemplars are extracted from parallel training data, having the same texts uttered by the source and target speakers. The input source signal is represented using the source exemplars and their weights. Then, the converted speech is constructed from the target exemplars and the weights related to the source exemplars. However, this exemplar-based approach needs to hold all training exemplars (frames), and it requires high computation times to obtain the weights of the source exemplars. In this paper, we propose a framework to train the basis matrices of the source and target exemplars so that they have a common weight matrix. By using the basis matrices instead of the exemplars, the VC is performed with lower computation times than with the exemplar-based method. The effectiveness of this method was confirmed by comparing its effectiveness (in speaker conversion experiments using noise-added speech data) with that of an exemplar-based method and a conventional Gaussian mixture model (GMM)-based method.

This paper presents an electrolaryngeal (EL) speech enhancement method capable of significantly improving naturalness of EL speech while causing no degradation in its intelligibility. An electrolarynx is an external device that artificially generates excitation sounds to enable laryngectomees to produce EL speech. Although proficient laryngectomees can produce quite intelligible EL speech, it sounds very unnatural due to the mechanical excitation produced by the device. Moreover, the excitation sounds produced by the device often leak outside, adding to EL speech as noise. To address these issues, there are mainly two conventional approached to EL speech enhancement through either noise reduction or statistical voice conversion (VC). The former approach usually causes no degradation in intelligibility but yields only small improvements in naturalness as the mechanical excitation sounds remain essentially unchanged. On the other hand, the latter approach significantly improves naturalness of EL speech using spectral and excitation parameters of natural voices converted from acoustic parameters of EL speech, but it usually causes degradation in intelligibility owing to errors in conversion. We propose a hybrid approach using a noise reduction method for enhancing spectral parameters and statistical voice conversion method for predicting excitation parameters. Moreover, we further modify the prediction process of the excitation parameters to improve its prediction accuracy and reduce adverse effects caused by unvoiced/voiced prediction errors. The experimental results demonstrate the proposed method yields significant improvements in naturalness compared with EL speech while keeping intelligibility high enough.

In this paper, we propose voice conversion (VC) based on articulatory features (AF) to vocal-tract parameters (VTP) mapping. An artificial neural network (ANN) is applied to map AF to VTP and to convert a speaker's voice to a target-speaker's voice. The proposed system is not only text-independent VC, in which it does not need parallel utterances between source and target-speakers, but can also be used for an arbitrary source-speaker. This means that our approach does not require source-speaker data to build the VC model. We are also focusing on a small number of target-speaker training data. For comparison, a baseline system based on Gaussian mixture model (GMM) approach is conducted. The experimental results for a small number of training data show that the converted voice of our approach is intelligible and has speaker individuality of the target-speaker.

This paper presents a voice conversion (VC) technique for noisy environments, where parallel exemplars are introduced to encode the source speech signal and synthesize the target speech signal. The parallel exemplars (dictionary) consist of the source exemplars and target exemplars, having the same texts uttered by the source and target speakers. The input source signal is decomposed into the source exemplars, noise exemplars and their weights (activities). Then, by using the weights of the source exemplars, the converted signal is constructed from the target exemplars. We carried out speaker conversion tasks using clean speech data and noise-added speech data. The effectiveness of this method was confirmed by comparing its effectiveness with that of a conventional Gaussian Mixture Model (GMM)-based method.

We propose a segment-based voice conversion technique using hidden Markov model (HMM)-based speech synthesis with nonparallel training data. In the proposed technique, the phoneme information with durations and a quantized F0 contour are extracted from the input speech of a source speaker, and are transmitted to a synthesis part. In the synthesis part, the quantized F0 symbols are used as prosodic context. A phonetically and prosodically context-dependent label sequence is generated from the transmitted phoneme and the F0 symbols. Then, converted speech is generated from the label sequence with durations using the target speaker's pre-trained context-dependent HMMs. In the model training, the models of the source and target speakers can be trained separately, hence there is no need to prepare parallel speech data of the source and target speakers. Objective and subjective experimental results show that the segment-based voice conversion with phonetic and prosodic contexts works effectively even if the parallel speech data is not available.