Singing voice conversion is to convert a singer's voice to another one's voice without changing singing content. Recent work shows that unsupervised singing voice conversion can be achieved with an autoencoder-based approach [1]. However, the converted singing voice can be easily out of key, showing that the existing approach cannot model the pitch information precisely. In this paper, we propose to advance the existing unsupervised singing voice conversion method proposed in [1] to achieve more accurate pitch translation and flexible pitch manipulation. Specifically, the proposed Pitch-Net added an adversarially trained pitch regression network to enforce the encoder network to learn pitch invariant phoneme representation, and a separate module to feed pitch extracted from the source audio to the decoder network. Our evaluation shows that the proposed method can greatly improve the quality of the converted singing voice (2.92 vs 3.75 in MOS). We also demonstrate that the pitch of converted singing can be easily controlled during generation by changing the levels of the extracted pitch before passing it to the decoder network.

In this work, we propose minimum Bayes risk (MBR) training of RNN-Transducer (RNN-T) for end-to-end speech recognition.Specifically, initialized with a RNN-T trained model, MBR training is conducted via minimizing the expected edit distance between the reference label sequence and on-thefly generated N-best hypothesis.We also introduce a heuristic to incorporate an external neural network language model (NNLM) in RNN-T beam search decoding and explore MBR training with the external NNLM.Experimental results demonstrate an MBR trained model outperforms a RNN-T trained model substantially and further improvements can be achieved if trained with an external NNLM.Our best MBR trained system achieves absolute character error rate (CER) reductions of 1.2% and 0.5% on read and spontaneous Mandarin speech respectively over a strong convolution and transformer based RNN-T baseline trained on ∼21,000 hours of speech.

Peking Opera has been the most dominant form of Chinese performing art since around 200 years ago.A Peking Opera singer usually exhibits a very strong personal style via introducing improvisation and expressiveness on stage which leads the actual rhythm and pitch contour to deviate significantly from the original music score.This inconsistency poses a great challenge in Peking Opera singing voice synthesis from a music score.In this work, we propose to deal with this issue and synthesize expressive Peking Opera singing from the music score based on the Duration Informed Attention Network (DurIAN) framework.To tackle the rhythm mismatch, Lagrange multiplier is used to find the optimal output phoneme duration sequence with the constraint of the given note duration from music score.As for the pitch contour mismatch, instead of directly inferring from music score, we adopt a pseudo music score generated from the real singing and feed it as input during training.The experiments demonstrate that with the proposed system we can synthesize Peking Opera singing voice with high-quality timbre, pitch and expressiveness.

Singing voice conversion is converting the timbre in the source singing to the target speaker's voice while keeping singing content the same. However, singing data for target speaker is much more difficult to collect compared with normal speech this http URL this paper, we introduce a singing voice conversion algorithm that is capable of generating high quality target speaker's singing using only his/her normal speech data. First, we manage to integrate the training and conversion process of speech and singing into one framework by unifying the features used in standard speech synthesis system and singing synthesis system. In this way, normal speech data can also contribute to singing voice conversion training, making the singing voice conversion system more robust especially when the singing database is small.Moreover, in order to achieve one-shot singing voice conversion, a speaker embedding module is developed using both speech and singing data, which provides target speaker identify information during conversion. Experiments indicate proposed sing conversion system can convert source singing to target speaker's high-quality singing with only 20 seconds of target speaker's enrollment speech data.

In this paper, we present a Sequence-to-Sequence Attentional Siamese Neural Network (Seq2Seq-ASNN) that leverages temporal alignment information for end-to-end speaker verification. In prior works of speaker discriminative neural networks, utterance-level evaluation/enrollment speaker representations are usually calculated. Our proposed model, utilizing a sequence-to-sequence (Seq2Seq) attention mechanism, maps the frame-level evaluation representation into enrollment feature domain and further generates an utterance-level evaluation-enrollment joint vector for final similarity measure. Feature learning, attention mechanism, and metric learning are jointly optimized using an end-to-end loss function. Experimental results show that our proposed model outperforms various baseline methods, including the traditional i-Vector/PLDA method, multi-enrollment end-to-end speaker verification models, d-vector approaches, and a self attention model, for text-dependent speaker verification on a Tencent internal voice wake-up dataset.

Acoustic model and language model (LM) have been two major components in conventional speech recognition systems. They are normally trained independently, but recently there has been a trend to optimize both components simultaneously in a unified end-to-end (E2E) framework. However, the performance gap between the E2E systems and the traditional hybrid systems suggests that some knowledge has not yet been fully utilized in the new framework. An observation is that the current attention-based E2E systems could produce better recognition results when decoded with LMs which are independently trained with the same resource. In this paper, we focus on how to improve attention-based E2E systems without increasing model complexity or resorting to extra data. A novel training strategy is proposed for multi-task training with the connectionist temporal classification (CTC) loss. The sequence-based minimum Bayes risk (MBR) loss is also investigated. Our experiments on SWB 300hrs showed that both loss functions could significantly improve the baseline model performance. The additional gain from joint-LM decoding remains the same for CTC trained model but is only marginal for MBR trained model. This implies that while CTC loss function is able to capture more acoustic knowledge, MBR loss function exploits more word/character dependency.

Conventional acoustic models for automatic speech recognition (ASR) are usually constructed from sub-word unit (e.g., context-dependent phoneme, grapheme, wordpiece etc.). Recent studies demonstrate that connectionist temporal classification (CTC) based acoustic-to-word (A2W) models are also promising for ASR. Such structures have drawn increasing attention as they can directly target words as output units, which simplify ASR pipeline by avoiding additional pronunciation lexicon, or even language model. In this study, we systematically explore to use word as acoustic modeling unit for conversational speech recognition. By replacing senone alignment with word alignment in a convolutional bidirectional LSTM architecture and employing a lexicon-free weighted finite-state transducer (WFST) based decoding, we greatly simplify conventional hybrid speech recognition system. On Hub5-2000 Switchboard/CallHome test sets with 300-hour training data, we achieve a WER that is close to the senone based hybrid systems with a WFST based decoding.

This paper advances the design of CTC-based all-neural (or end-to-end) speech recognizers. We propose a novel symbol inventory, and a novel iterated-CTC method in which a second system is used to transform a noisy initial output into a cleaner version. We present a number of stabilization and initialization methods we have found useful in training these networks. We evaluate our system on the commonly used NIST 2000 conversational telephony test set, and significantly exceed the previously published performance of similar systems, both with and without the use of an external language model and decoding technology.

This document briefly describes the systems submitted by the Center for Robust Speech Systems (CRSS) from The University of Texas at Dallas (UTD) to the 2016 National Institute of Standards and Technology (NIST) Speaker Recognition Evaluation (SRE). We developed several UBM and DNN i-Vector based speaker recognition systems with different data sets and feature representations. Given that the emphasis of the NIST SRE 2016 is on language mismatch between training and enrollment/test data, so-called domain mismatch, in our system development we focused on: (1) using unlabeled in-domain data for centralizing data to alleviate the domain mismatch problem, (2) finding the best data set for training LDA/PLDA, (3) using newly proposed dimension reduction technique incorporating unlabeled in-domain data before PLDA training, (4) unsupervised speaker clustering of unlabeled data and using them alone or with previous SREs for PLDA training, (5) score calibration using only unlabeled data and combination of unlabeled and development (Dev) data as separate experiments.

Pitch estimation in diverse naturalistic audio streams remains a challenge for speech processing and spoken language technology. In this study, we investigate the use of robust harmonic features for classification-based pitch estimation. The proposed pitch estimation algorithm is composed of two stages: pitch candidate generation and target pitch selection. Based on energy intensity and spectral envelope shape, five types of robust harmonic features are proposed to reflect pitch associated harmonic structure. A neural network is adopted for modeling the relationship between input harmonic features and output pitch salience for each specific pitch candidate. In the test stage, each pitch candidate is assessed with an output salience that indicates the potential as a true pitch value, based on its input feature vector processed through the neural network. Finally, according to the temporal continuity of pitch values, pitch contour tracking is performed using a hidden Markov model (HMM), and the Viterbi algorithm is used for HMM decoding. Experimental results show that the proposed algorithm outperforms several state-of-the-art pitch estimation methods in terms of accuracy in both high and low levels of additive noise.

Most speaker diarization research has focused on unsupervised scenarios, where no human supervision is available. However, in many real-world applications, a certain amount of human input could be expected, especially when minimal human supervision brings significant performance improvement. In this study, we propose an active learning based bottom-up speaker clustering algorithm to effectively improve speaker diarization performance with limited human input. Specifically, the proposed active learning based speaker clustering has two different stages: explore and constrained clustering. The explore stage is to quickly discover at least one sample for each speaker for boosting speaker clustering process with reliable initial speaker clusters. After discovering all, or a majority, of the involved speakers during explore stage, the constrained clustering is performed. Constrained clustering is similar to traditional bottom-up clustering process with an important difference that the clusters created during explore stage are restricted from merging with each other. Constrained clustering continues until only the clusters generated from the explore stage are left. Since the objective of active learning based speaker clustering algorithm is to provide good initial speaker models, performance saturates as soon as sufficient examples are ensured for each cluster. To further improve diarization performance with increasing human input, we propose a second method which actively select speech segments that account for the largest expected speaker error from existing cluster assignments for human evaluation and reassignment. The algorithms are evaluated on our recently created Apollo Mission Control Center dataset as well as augmented multiparty interaction meeting corpus. The results indicate that the proposed active learning algorithms are able to reduce diarization error rate significantly with a relatively small amount of human supervision.

Deep neural network (DNN) based acoustic models have greatly improved the performance of automatic speech recognition (ASR) for various tasks. Further performance improvements have been reported when making DNNs aware of the acoustic context (e.g. speaker or environment) for example by adding auxiliary features to the input, such as noise estimates or speaker i-vectors. We have recently proposed a context adaptive DNN (CA-DNN), which is another approach to exploit the acoustic context information within a DNN. A CA-DNN is a DNN that has one or several factorized layers, i.e. layers that use a different set of parameters to process each acoustic context class. The output of a factorized layer is obtained by the weighted sum over the contribution of the different context classes, given weights over the context classes. In our previous work, the class weights were computed independently of the recognizer. In this paper, we extend our previous work by introducing the joint training of the CA-DNN parameters and the class weights computation. Consequently, the class weights and the associated class definitions can be optimized for ASR. We report experimental results on the AURORA4 noisy speech recognition task showing the potential of our approach for fast unsupervised adaptation.

This study proposes a novel deep neural network (DNN) based approach to language identification (LID) for the NIST 2015 Language Recognition (LRE) i-Vector Machine Learning Challenge. State-of-the-art DNN based LID systems utilize large amounts of labeled training data. The 2015 LRE i-Vector Machine Learning Challenge limits the access to only ready-to-use i-Vectors for LID system training and testing. This poses unique challenges in designing DNN based LID systems, since optimized front-ends and network architectures can no longer be used. We propose to use the training i-Vectors to train an initial DNN for LID. Next, we present a novel strategy to use this initial DNN to estimate out-of-set language labels from the development data. The final DNN for LID is trained using the original training data, and the estimated out-of-set language data. We show that augmenting the training set with out-of-set labels leads to significant improvement in the LID performance. Our approach obtains very competitive costs (defined by NIST) of 26.56, and 25.98 respectively, on the progress and evaluation subsets of the challenge. Since the amount of training data is very limited (300 i-Vectors per language), this study outlines a successful recipe for DNN based LID using very limited resources.

In this paper, we present the system developed by the Center for Robust Speech Systems (CRSS), University of Texas at Dallas, for the NIST 2015 language recognition i-vector machine learning challenge. Our system includes several subsystems, based on Linear Discriminant Analysis - Support Vector Machine (LDA-SVM) and deep neural network (DNN) approaches. An important feature of this challenge is the emphasis on out-of-set language detection. As a result, our system development focuses mainly on the evaluation and comparison of two different out-of-set language detection strategies: direct out-of-set detection and indirect out-of-set detection. These out-of-set detection strategies differ mainly on whether the unlabeled development data are used or not. The experimental results indicate that indirect out-of-set detection strategies used in our system could efficiently exploit the unlabeled development data, and therefore consistently outperform the direct out-of-set detection approach. Finally, by fusing four variants of indirect out-of-set detection based subsystems, our system achieves a relative performance gain of up to 45%, compared to the baseline cosine distance scoring (CDS) system provided by organizer.

CHiME-3 is a research community challenge organised in 2015 to evaluate speech recognition systems for mobile multi-microphone devices used in noisy daily environments. This paper describes NTT's CHiME-3 system, which integrates advanced speech enhancement and recognition techniques. Newly developed techniques include the use of spectral masks for acoustic beam-steering vector estimation and acoustic modelling with deep convolutional neural networks based on the "network in network" concept. In addition to these improvements, our system has several key differences from the official baseline system. The differences include multi-microphone training, dereverberation, and cross adaptation of neural networks with different architectures. The impacts that these techniques have on recognition performance are investigated. By combining these advanced techniques, our system achieves a 3.45% development error rate and a 5.83% evaluation error rate. Three simpler systems are also developed to perform evaluations with constrained set-ups.

It is common for subjects to produce speech while performing a physical task where speech technology may be used. Variabilities are introduced to speech since physical task can influence human speech production. These variabilities degrade the performance of most speech systems. It is vital to detect speech under physical stress variabilities for subsequent algorithm processsing. This study presents a method for detecting physical task stress from speech. Inspired by the fact that i-vectors can generally model total factors from speech, a state-of-the-art ivector framework is investigated with MFCCs and our previously formulated TEO-CB-Auto-Env features for neutral/physical task stress detection. Since MFCCs are derived from a linear speech production model and TEO-CB-Auto-Env features employ a nonlinear operator, these two features are believed to have complementary effects on physical task stress detection. Two alternative fusion strategies (feature-level and score-level fusion) are investigated to validate this hypothesis. Experiments over the UT-Scope Physical Corpus demonstrate that a relative accuracy gain of 2.68% is obtained when fusing different feature based i-vectors. An additional relative performance boost with of 6.52% in accuracy is achieved using score level fusion.

In this study, we explore the propagation of uncertainty in the state-of-the-art speaker recognition system. Specifically, we incorporate the uncertainty associated with observation features into the i-Vector extraction framework. To prove the concept, both the oracle and practically estimated uncertainty are used for evaluation. The oracle uncertainty is calculated assuming the knowledge of clean speech features, while the estimated uncertainties are obtained using SPLICE and joint-GMM based methods. We evaluate the proposed framework on both YOHO and NIST 2010 Speaker Recognition Evaluation (SRE) corpora by artificially introducing noise at different SNRs. In the speaker verification experiments, we confirmed that the proposed uncertainty based i-Vector extraction framework shows significant robustness against noise.

Speech and language processing technology has the potential of playing an important role in future deep space missions. To be able to replicate the success of speech technologies from ground to space, it is important to understand how astronaut’s speech production mechanism changes when they are in space. In this study, we investigate the variations of astronaut’s voice charac-teristic during NASA Apollo 11 mission. While the focus is constrained to analysis of the three astronauts voices who par-ticipated in the Apollo 11 mission, it is the first step towards our long term objective of automating large components of space missions with speech and language technology. The result of this study is also significant from an historical point of view as it provides a new perspective of understanding the key mo-ment of human history- landing a man on the moon, as well as employed for future advancement in speech and language tech-nology in “non-neutral”conditions.

In state-of-the-art speaker recognition system, universal background model (UBM) plays a role of acoustic space division. Each Gaussian mixture of trained UBM represents one distinct acoustic region. The posterior probabilities of features belonging to each region are further used as core components of Baum-Welch statistics. Therefore, the quality of estimated Baum-Welch statistics depends highly on how acoustic regions are separable with each other. In this paper, we propose to transform the front end acoustical features into a space where the separability of mixtures of trained UBM can be optimized. To achieve this, an UBM was first trained from the acoustical features and a transformation matrix is estimated using linear discriminant analysis (LDA) by treating each mixture of trained UBM as independent class. Therefore, the proposed method named as UBM-based LDA (uLDA) does not require any speaker labels or other supervised information. The obtained transformation matrix is then applied to acoustic features for i-Vector extraction. Experimental results on the male part of core conditions of NIST SRE 2010 dataset confirmed the improved performance using proposed method. Index Terms: Speaker recognition, i-Vector, universal background model (UBM), Baum-Welch statistic, LDA.

State-of-the-art speaker verification systems model speaker identity by mapping i-Vectors onto a probabilistic linear discriminant analysis (PLDA) space. Compared to other modeling approaches (such as cosine distance scoring), PLDA provides a more efficient mechanism to separate speaker information from other sources of undesired variabilities and offers superior speaker verification performance. Unfortunately, this efficiency is obtained at the cost of a required large corpus of labeled development data, which is too expensive/unrealistic in many cases. This study investigates a potential solution to resolve this challenge by effectively utilizing unlabeled development data with universal imposter clustering. The proposed method offers +21.9% and +34.6% relative gains versus the baseline system on two public available corpora, respectively. This significant improvement proves the effectiveness of the proposed method.

Mask-based objective speech-intelligibility measures have been successfully proposed for evaluating the performance of binary masking algorithms. These objective measures were computed directly by comparing the estimated binary mask against the ground truth ideal binary mask (IdBM). Most of these objective measures, however, assign equal weight to all time-frequency (T-F) units. In this study, we propose to improve the existing mask-based objective measures by weighting each T-F unit according to its target or masker loudness. The proposed objective measure shows significantly better performance than two other existing mask-based objective measures.

NASA’s Apollo program stands as one of mankind’s greatest achievements in the 20th century. During a span of 4 years (from 1968 to 1972), a total of 9 lunar missions were launched and 12 astronauts walked on the surface of the moon. It was one the most complex operations executed from scientific, techno-logical and operational perspectives. In this paper, we describe our recent efforts in gathering and organizing the Apollo pro-gram data. It is important to note that the audio content captured during the 7-10 day missions represent the coordinated efforts of hundreds of individuals within NASA Mission Control, re-sulting in well over 100k hours of data for the entire program. It is our intention to make the material stemming from this ef-fort available to the research community to further research ad-vancements in speech and language processing. Particularly, we describe the speech and text aspects of the Apollo data while pointing out its applicability to several classical speech pro-cessing and natural language processing problems such as au-dio processing, speech and speaker recognition, information re-trieval, document linking and a range of other processing tasks which enable knowledge search, retrieval, and understanding.. We also highlight some of the outstanding opportunities and challenges associated with this dataset. Finally, we also present initial results for speech recognition, document linking, and au-dio processing systems.