Search

Current end-to-end code-switching Text-to-Speech (TTS) can already generate high quality two languages speech in the same utterance with single speaker bilingual corpora. When the speakers of the bilingual corpora are different, the naturalness and consistency of the code-switching TTS will be poor. The cross-lingual embedding layers structure we proposed makes similar syllables in different languages relevant, thus improving the naturalness and consistency of generated speech. In the end-to-end code-switching TTS, there exists problem of prosody instability when synthesizing paragraph text. The text enhancement method we proposed makes the input contain prosodic information and sentence-level context information, thus improving the prosody stability of paragraph text. Experimental results demonstrate the effectiveness of the proposed methods in the naturalness, consistency, and prosody stability. In addition to Mandarin and English, we also apply these methods to Shanghaiese and Cantonese corpora, proving that the methods we proposed can be extended to other languages to build end-to-end code-switching TTS system., Comment: accepted in ISCSLP 2021

Neural end-to-end TTS such as Tacotron like network can generate very high-quality synthesized speech, and even close to human recording for similar domain text. However, it performs unsatisfactory when scaling it to some challenging test sets. One concern is that the encoder-decoder with attention-based network adopts autoregressive generative sequence model with the limitation of “exposure bias”: errors made early could be quickly amplified, harming subsequent sequence generation. To address this issue, we propose two novel methods, which aim at predicting future by improving the agreement between forward and backward decoding sequence. The first one (denoted as MRBA) is achieved by adding divergence regularization terms to model training objective to maximize the agreement between two directional models, namely L2R (which generates targets from left-to-right) and R2L (which generates targets from right-to-left). While the second one (denoted as BDR) operates on decoder-level and exploits the future information during decoding. By introducing regularization term into the training objective of forward-backward decoders, the forward-decoder's hidden states are forced to be close to the backward-decoder's. Thus, the hidden representations of a unidirectional decoder are encouraged to embed some useful information about the future. Moreover, in order to make forward and backward decoding to improve each other in an interactive process, a joint training method is designed. Experimental results on both English and Mandarin dataset show that our proposed methods especially the second one (BDR), lead to a significantly improvement on both robustness and overall naturalness, as achieving obvious preference advantages in a challenging test, and achieving state-of-the-art performance (outperforming baseline “the revised version of Tacotron2” with a gap of 0.13 and 0.12 for English and Mandarin in MOS, respectively) on a general test.

The acoustic model trained using the knowledge from the shared hidden layer (SHL) model outperforms the model trained only by using the target language, especially under low resource conditions. However, the shared features may contain some unnecessary language dependent information. It will degrade the performance of the target model. Therefore, this paper proposes language-adversarial transfer learning to alleviate this problem. Adversarial learning is used to ensure that the shared layers of the SHL-model can learn more language invariant features. Experiments are conducted on IARPA Babel datasets. The results show that the target model trained using the knowledge transferred from the adversarial SHL-model achieves up to 10.1% relative word error rate reduction when compared with the target model trained using the knowledge transferred from the SHL-model.

The paper describes the CASIA speech synthesis system entry for challenge M2VoC 2021. The low similarity and naturalness of synthesized speech remains a challenging problem for speaker adaptation with few resources. Since the end-to-end acoustic model is too complex to interpret, overfitting will occur when training with few data. To prevent the model from overfitting, this paper proposes a novel speaker adaptation framework that decomposes the prosody and voice characteristics in the end-to-end model. A prosody control attention is proposed to control the phonemes’ duration of different speakers. To make the attention controlled by the prosody information, a set of phoneme-level transition tokens is auto-learned from the prosody encoder in our framework and these transition tokens can determine the duration of phonemes in the attention mechanism. Secondly, when we need to use small data set for speaker adaptation, we just need to adapt the speaker related prosody model and decoder, which can prevent the model from overfitting. Further, we use a data puring model to automatically optimize the quality of datasets. Experiments demonstrate the effectiveness of speaker adaptation based on our method, and we (team identifier is T03) get the top three results in competition M2VoC by using this framework.

Aiming at efficiently predicting acoustic features with high naturalness and robustness, this paper proposes PATNet, a neural acoustic model for speech synthesis using phoneme-level autoregression. PATNet accepts phoneme sequences as input and is built based on Transformer structure. PATNet adopts a duration model instead of attention mechanism for sequence alignment. The decoder of PATNet predicts multi-frame spectra within one phoneme in parallel given the predicted spectra of previous phonemes. Such phoneme-level autoregression enables PATNet to achieve higher inference efficiency than the models with frame-level autoregression, such as Transformer-TTS, and improves the robustness of acoustic feature prediction by utilizing phoneme boundaries explicitly. Experimental results show that the speech synthesized by PATNet obtained lower character error rate (CER) than Tacotron, Transfomer-TTS and FastSpeech when evaluated by a speech recognition engine. Besides, PATNet achieved 10 times faster inference speed than Transformer-TTS and significantly better naturalness than FastSpeech.

End-to-end framework can generate high-quality and high-similarity speech in the personalized speech synthesis task. However, the generalization of out-of-domain texts is still a challenging task. Limited target data leads to unacceptable errors and poor prosody and similarity performance of the synthetic speech. In this paper, we present a bi-level function decoupling framework to realise separate modeling and controlling for solving above problems. Firstly, on the style representation modeling level, compared with the conventional methods that use single embedding to model all the text dependent discrepancies, it is proposed that the speaker embedding and prosody embedding are modeled separately based on the reference audio and phonetic posteriorgram (PPG) by a multi-head attention mechanism. Secondly, on the model structure level, the decoder model structure is factored into average-net and adaptation-net, where the duration prosody controlling and speaker timbre imitation are mainly designed in relatively separate areas. Experimental results on Mandarin dataset show that the proposed methods lead to an improvement on both robustness, naturalness and similarity.

Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because the decoder predicts text tokens (such as characters or words) in an autoregressive manner, it is difficult for an AED model to predict all tokens in parallel. This makes the inference speed relatively slow. We believe that because the encoder already captures the whole speech utterance, which has the token-level relationship implicitly, we can predict a token without explicitly autoregressive language modeling. When the prediction of a token does not rely on other tokens, the parallel prediction of all tokens in the sequence is realizable. Based on this idea, we propose a non-autoregressive speech recognition model called LASO (Listen Attentively, and Spell Once). The model consists of an encoder, a decoder, and a position dependent summarizer (PDS). The three modules are based on basic attention blocks. The encoder extracts high-level representations from the speech. The PDS uses positional encodings corresponding to tokens to convert the acoustic representations into token-level representations. The decoder further captures token-level relationships with the self-attention mechanism. At last, the probability distribution on the vocabulary is computed for each token position. Therefore, speech recognition is re-formulated as a position-wise classification problem. Further, we propose a cross-modal transfer learning method to refine semantics from a large-scale pre-trained language model BERT for improving the performance., 14 pages, 7 figures

Attention-based models have recently shown powerful representation learning ability in speaker recognition. However, most of the attention mechanism based models primarily focus on pooling layers. In this work, we present an end-to-end speaker verification system which leverage time-frequency and channel features hierarchically. To further improve system performance, we employ Large Margin Cosine Loss to optimize the model to determine the optimal loss function. We carry out experiments on the VoxCeleb1 datasets to evaluate the effectiveness of our methods. The results suggest that our best system outperforms the i-vector + PLDA and x-vector system by 53.3% and 7.6%, respectively.

Transducer-based models, such as RNN-Transducer and transformer-transducer, have achieved great success in speech recognition. A typical transducer model decodes the output sequence conditioned on the current acoustic state and previously predicted tokens step by step. Statistically, The number of blank tokens in the prediction results accounts for nearly 90\% of all tokens. It takes a lot of computation and time to predict the blank tokens, but only the non-blank tokens will appear in the final output sequence. Therefore, we propose a method named fast-skip regularization, which tries to align the blank position predicted by a transducer with that predicted by a CTC model. During the inference, the transducer model can predict the blank tokens in advance by a simple CTC project layer without many complicated forward calculations of the transducer decoder and then skip them, which will reduce the computation and improve the inference speed greatly. All experiments are conducted on a public Chinese mandarin dataset AISHELL-1. The results show that the fast-skip regularization can indeed help the transducer model learn the blank position alignments. Besides, the inference with fast-skip can be speeded up nearly 4 times with only a little performance degradation., Comment: Submitted to INTERSPEECH2021

Fake audio attack becomes a major threat to the speaker verification system. Although current detection approaches have achieved promising results on dataset-specific scenarios, they encounter difficulties on unseen spoofing data. Fine-tuning and retraining from scratch have been applied to incorporate new data. However, fine-tuning leads to performance degradation on previous data. Retraining takes a lot of time and computation resources. Besides, previous data are unavailable due to privacy in some situations. To solve the above problems, this paper proposes detecting fake without forgetting, a continual-learning-based method, to make the model learn new spoofing attacks incrementally. A knowledge distillation loss is introduced to loss function to preserve the memory of original model. Supposing the distribution of genuine voice is consistent among different scenarios, an extra embedding similarity loss is used as another constraint to further do a positive sample alignment. Experiments are conducted on the ASVspoof2019 dataset. The results show that our proposed method outperforms fine-tuning by the relative reduction of average equal error rate up to 81.62%., Comment: 5 pages, conference

Although attention based end-to-end models have achieved promising performance in speech recognition, the multi-pass forward computation in beam-search increases inference time cost, which limits their practical applications. To address this issue, we propose a non-autoregressive end-to-end speech recognition system called LASO (listen attentively, and spell once). Because of the non-autoregressive property, LASO predicts a textual token in the sequence without the dependence on other tokens. Without beam-search, the one-pass propagation much reduces inference time cost of LASO. And because the model is based on the attention based feedforward structure, the computation can be implemented in parallel efficiently. We conduct experiments on publicly available Chinese dataset AISHELL-1. LASO achieves a character error rate of 6.4%, which outperforms the state-of-the-art autoregressive transformer model (6.7%). The average inference latency is 21 ms, which is 1/50 of the autoregressive transformer model., Comment: accepted by INTERSPEECH2020

End-to-end speech synthesis can generate high-quality synthetic speech and achieve high similarity scores with low-resource adaptation data. However, the generalization of out-domain texts is still a challenging task. The limited adaptation data leads to unacceptable errors and the poor prosody performance of the synthetic speech. In this paper, we present two novel methods to handle the above problems by focusing on the attention. Firstly, compared with the conventional methods that extract prosody embeddings for conditioning input, a duration controller with feedback mechanism is proposed, which can control the states transition in the sequence-to-sequence model more directly and precisely. Secondly, to alleviate the unmatching text-audio pairs’ impact on model, an adaptative optimization strategy which would consider the matching degree of the training sample is also proposed. Experimental results on Mandarin dataset show that proposed methods lead to an improvement on both robustness and overall naturalness.

Non-autoregressive transformer models have achieved extremely fast inference speed and comparable performance with autoregressive sequence-to-sequence models in neural machine translation. Most of the non-autoregressive transformers decode the target sequence from a predefined-length mask sequence. If the predefined length is too long, it will cause a lot of redundant calculations. If the predefined length is shorter than the length of the target sequence, it will hurt the performance of the model. To address this problem and improve the inference speed, we propose a spike-triggered non-autoregressive transformer model for end-to-end speech recognition, which introduces a CTC module to predict the length of the target sequence and accelerate the convergence. All the experiments are conducted on a public Chinese mandarin dataset AISHELL-1. The results show that the proposed model can accurately predict the length of the target sequence and achieve a competitive performance with the advanced transformers. What's more, the model even achieves a real-time factor of 0.0056, which exceeds all mainstream speech recognition models., Comment: 5 pages

Despite the recent significant advances witnessed in end-to-end (E2E) ASR system for code-switching, hunger for audio-text paired data limits the further improvement of the models' performance. In this paper, we propose a decoupled transformer model to use monolingual paired data and unpaired text data to alleviate the problem of code-switching data shortage. The model is decoupled into two parts: audio-to-phoneme (A2P) network and phoneme-to-text (P2T) network. The A2P network can learn acoustic pattern scenarios using large-scale monolingual paired data. Meanwhile, it generates multiple phoneme sequence candidates for single audio data in real-time during the training process. Then the generated phoneme-text paired data is used to train the P2T network. This network can be pre-trained with large amounts of external unpaired text data. By using monolingual data and unpaired text data, the decoupled transformer model reduces the high dependency on code-switching paired training data of E2E model to a certain extent. Finally, the two networks are optimized jointly through attention fusion. We evaluate the proposed method on the public Mandarin-English code-switching dataset. Compared with our transformer baseline, the proposed method achieves 18.14% relative mix error rate reduction., Comment: 5 pages, 1 figures

Recently, language identity information has been utilized to improve the performance of end-to-end code-switching (CS) speech recognition task. However, previous work use an additional language identification (LID) model as an auxiliary module, which increases computation cost. In this work, we propose an improved recurrent neural network transducer (RNN-T) model with language bias to alleviate the problem. We use the language identities to bias the model to predict the CS points. This promotes the model to learn the language identity information directly from transcriptions, and no additional LID model is needed. We evaluate the approach on a Mandarin-English CS corpus SEAME. Compared to our RNN-T baseline, the RNN-T with language bias can achieve 16.2% and 12.9% relative mixed error reduction on two test sets, respectively.

Recurrent neural networks (RNNs) have shown significant improvements in recent years for speech enhancement. However, the model complexity and inference time cost of RNNs are much higher than deep feed-forward neural networks (DNNs). Therefore, these limit the applications of speech enhancement. This paper proposes a deep time delay neural network (TDNN) for speech enhancement with full data learning. The TDNN has excellent potential for capturing long range temporal contexts, which utilizes a modular and incremental design. Besides, the TDNN preserves the feed-forward structure so that its inference cost is comparable to standard DNN. To make full use of the training data, we propose a full data learning method for speech enhancement. More specifically, we not only use the noisy-to-clean (input-to-target) to train the enhanced model, but also the clean-to-clean and noise-to-silence data. Therefore, all of the training data can be used to train the enhanced model. Our experiments are conducted on TIMIT dataset. Experimental results show that our proposed method could achieve a better performance than DNN and comparable even better performance than BLSTM. Meanwhile, compared with the BLSTM, the proposed method drastically reduce the inference time., Comment: Accepted by ISCSLP 2021

The joint training framework for speech enhancement and recognition methods have obtained quite good performances for robust end-to-end automatic speech recognition (ASR). However, these methods only utilize the enhanced feature as the input of the speech recognition component, which are affected by the speech distortion problem. In order to address this problem, this paper proposes a gated recurrent fusion (GRF) method with joint training framework for robust end-to-end ASR. The GRF algorithm is used to dynamically combine the noisy and enhanced features. Therefore, the GRF can not only remove the noise signals from the enhanced features, but also learn the raw fine structures from the noisy features so that it can alleviate the speech distortion. The proposed method consists of speech enhancement, GRF and speech recognition. Firstly, the mask based speech enhancement network is applied to enhance the input speech. Secondly, the GRF is applied to address the speech distortion problem. Thirdly, to improve the performance of ASR, the state-of-the-art speech transformer algorithm is used as the speech recognition component. Finally, the joint training framework is utilized to optimize these three components, simultaneously. Our experiments are conducted on an open-source Mandarin speech corpus called AISHELL-1. Experimental results show that the proposed method achieves the relative character error rate (CER) reduction of 10.04\% over the conventional joint enhancement and transformer method only using the enhanced features. Especially for the low signal-to-noise ratio (0 dB), our proposed method can achieves better performances with 12.67\% CER reduction, which suggests the potential of our proposed method., Comment: Accepted by IEEE/ACM Transactions on Audio, Speech, and Language Processing

Query-by-example (QbE) keyword spotting is convenient for users to define their own keywords, so it is useful in device control. However, conventional regular softmax, which is commonly used for training QbE models, has two limitations. First, the learned features are not discriminative enough. Second, norm variations of the unnormalized features affect computing cosine similarities. To address these issues, this paper introduces normalization and additive margin into residual networks for QbE keyword spotting. Features and weights are normalized on a hypersphere of fixed radius. Additive margin further helps to reduce the intra-class variations and increase inter-class differences. Based on public datasets AISHELL-1 and HelloNPU, we design three different test sets, namely in-vocabulary, out-of-vocabulary, and cross-corpus, to evaluate our proposed method. Experiments show that our proposed method can learn more discriminative embedding features. For totally unseen situation, our proposed method achieves a relative false rejection rate reduction of 46.60% when the false alarm rate is 2% in cross-corpus evaluation, compared with regular softmax.

This paper proposes a simple yet effective unsupervised speaker adaptation approach for batch normalization based deep neural network acoustic models. The basic idea of this approach is to recompute means and variances in all batch normalization layers over the test data for every speaker. Thus the distribution of the test data can be close to the training data. This approach doesn't need to adjust any trainable parameters of the acoustic model. Experiments are conducted on CHiME-3 datasets. The results show that the proposed adaptation obtains improvement on the real test set by 2.17 % relative average word error rate (WER) reduction when compared with the scaling and shifting factors (SSF) adaptation. Combining our proposed MV adaptation with the SSF adaptation obtains further improvement.

Short utterances dialect identification is a challenging task because of the substantial similarity between dialects. The previous cross-entropy loss function does not consider the category and probability of prediction error, which result in insensitivity to easily misclassified and unbalanced samples. To solve this problem, we propose to use an improved cross-entropy loss function, namely focal loss, introducing category weights and tunable focusing parameter to improve the classification accuracy. Experiments are carried out on AI Dialect Contest database. The results demonstrate that our proposed end-to-end model trained with focal loss achieves better performance than the model trained with cross-entropy loss function.

Speech enhancement generative adversarial network (SEGAN) is an end-to-end deep learning architecture, which only uses the clean speech as the training targets. However, when the signal-to-noise ratio (SNR) is very low, predicting clean speech signals could be very difficult as the speech is dominated by the noise. In order to address this problem, in this paper, we propose a gated convolutional neural network (CNN) SEGAN (GSEGAN) with noise prior knowledge learning to address this problem. The proposed model not only estimates the clean speech, but also learns the noise prior knowledge to assist the speech enhancement. In addition, gated CNN has an excellent potential for capturing long-term temporal dependencies than regular CNN. Motivated by this, we use a gated CNN architecture to acquire more detailed information at waveform level instead of regular CNN. We evaluate the proposed method GSEGAN on Voice Bank corpus. Experimental results show that the proposed method GSEGAN outperforms the SEGAN baseline, with a relative improvement of 0.7%, 28.2% and 43.9% for perceptual evaluation of speech quality (PESQ), overall Signal-to-Noise Ratio (SNRovl) and Segmental Signal-to-Noise Ratio (SNRseg), respectively.

In order to improve the performance of distant speech recognition tasks, this paper proposes to distill knowledge from the close-talking model to the distant model using parallel data. The close-talking model is called the teacher model. The distant model is called the student model. The student model is trained to imitate the output distributions of the teacher model. This constraint can be realized by minimizing the Kullback-Leibler (KL) divergence between the output distribution of the student model and the teacher model. Experimental results on AMI datasets show that the best student model achieves up to 8.5% relative word error rate (WER) reduction when compared with the conventionally-trained baseline models.

Recurrent neural network transducers (RNN-T) have been successfully applied in end-to-end speech recognition. However, the recurrent structure makes it difficult for parallelization . In this paper, we propose a self-attention transducer (SA-T) for speech recognition. RNNs are replaced with self-attention blocks, which are powerful to model long-term dependencies inside sequences and able to be efficiently parallelized. Furthermore, a path-aware regularization is proposed to assist SA-T to learn alignments and improve the performance. Additionally, a chunk-flow mechanism is utilized to achieve online decoding. All experiments are conducted on a Mandarin Chinese dataset AISHELL-1. The results demonstrate that our proposed approach achieves a 21.3% relative reduction in character error rate compared with the baseline RNN-T. In addition, the SA-T with chunk-flow mechanism can perform online decoding with only a little degradation of the performance.

Deep clustering (DC) and utterance-level permutation invariant training (uPIT) have been demonstrated promising for speaker-independent speech separation. DC is usually formulated as two-step processes: embedding learning and embedding clustering, which results in complex separation pipelines and a huge obstacle in directly optimizing the actual separation objectives. As for uPIT, it only minimizes the chosen permutation with the lowest mean square error, doesn't discriminate it with other permutations. In this paper, we propose a discriminative learning method for speaker-independent speech separation using deep embedding features. Firstly, a DC network is trained to extract deep embedding features, which contain each source's information and have an advantage in discriminating each target speakers. Then these features are used as the input for uPIT to directly separate the different sources. Finally, uPIT and DC are jointly trained, which directly optimizes the actual separation objectives. Moreover, in order to maximize the distance of each permutation, the discriminative learning is applied to fine tuning the whole model. Our experiments are conducted on WSJ0-2mix dataset. Experimental results show that the proposed models achieve better performances than DC and uPIT for speaker-independent speech separation., 5 pages, 1 figure, accepted by INTERSPEECH 2019

Integrating an external language model into a sequence-to-sequence speech recognition system is non-trivial. Previous works utilize linear interpolation or a fusion network to integrate external language models. However, these approaches introduce external components, and increase decoding computation. In this paper, we instead propose a knowledge distillation based training approach to integrating external language models into a sequence-to-sequence model. A recurrent neural network language model, which is trained on large scale external text, generates soft labels to guide the sequence-to-sequence model training. Thus, the language model plays the role of the teacher. This approach does not add any external component to the sequence-to-sequence model during testing. And this approach is flexible to be combined with shallow fusion technique together for decoding. The experiments are conducted on public Chinese datasets AISHELL-1 and CLMAD. Our approach achieves a character error rate of 9.3%, which is relatively reduced by 18.42% compared with the vanilla sequence-to-sequence model., 5 pages, 3 figures, accepted by INTERSPEECH 2019

This paper proposes to learn language-invariant bottleneck features from an adversarial end-to-end acoustic model for low resource languages. The multilingual end-to-end model is trained with a connectionist temporal classification loss function. The model has shared and private layers. The shared layers are the hidden layers utilized to learn universal features for all the languages. The private layers are the language-dependent layers used to capture language-specific features. Attention based adversarial end-to-end language identification is used to capture enough language information. Furthermore, orthogonality constraints are used to make private and shared features dissimilar. Experiments are conducted on IARPA Babel datasets. The results show that the target model trained with the proposed language-invariant bottleneck features outperforms the target model trained with the conventional multilingual bottleneck features by up to 9.7% relative word error rate reduction.

This paper proposes to use self-attention based model to predict punctuation marks for word sequences. The model is trained using word and speech embedding features which are obtained from the pre-trained Word2Vec and Speech2Vec, respectively. Thus, the model can use any kind of textual data and speech data. Experiments are conducted on English IWSLT2011 datasets. The results show that the self-attention based model trained using word and speech embedding features outperforms the previous state-of-the-art single model by up to 7.8% absolute overall F 1 -score. The results also show that it obtains performance improvement by up to 4.7% absolute overall F 1 -score against the previous best ensemble model.

Attention-based encoder-decoder (AED) models have achieved promising performance in speech recognition. However, because of the end-to-end training, an AED model is usually trained with speech-text paired data. It is challenging to incorporate external text-only data into AED models. Another issue of the AED model is that it does not use the right context of a text token while predicting the token. To alleviate the above two issues, we propose a unified method called LST (Learn Spelling from Teachers) to integrate knowledge into an AED model from the external text-only data and leverage the whole context in a sentence. The method is divided into two stages. First, in the representation stage, a language model is trained on the text. It can be seen as that the knowledge in the text is compressed into the LM. Then, at the transferring stage, the knowledge is transferred to the AED model via teacher-student learning. To further use the whole context of the text sentence, we propose an LM called causal cloze completer (COR), which estimates the probability of a token, given both the left context and the right context of it. Therefore, with LST training, the AED model can leverage the whole context in the sentence. Different from fusion based methods, which use LM during decoding, the proposed method does not increase any extra complexity at the inference stage. We conduct experiments on two scales of public Chinese datasets AISHELL-1 and AISHELL-2. The experimental results demonstrate the effectiveness of leveraging external text-only data and the whole context in a sentence with our proposed method, compared with baseline hybrid systems and AED model based systems., Comment: Submitted TASLP

For most of the attention-based sequence-to-sequence models, the decoder predicts the output sequence conditioned on the entire input sequence processed by the encoder. The asynchronous problem between the encoding and decoding makes these models difficult to be applied for online speech recognition. In this paper, we propose a model named synchronous transformer to address this problem, which can predict the output sequence chunk by chunk. Once a fixed-length chunk of the input sequence is processed by the encoder, the decoder begins to predict symbols immediately. During training, a forward-backward algorithm is introduced to optimize all the possible alignment paths. Our model is evaluated on a Mandarin dataset AISHELL-1. The experiments show that the synchronous transformer is able to perform encoding and decoding synchronously, and achieves a character error rate of 8.91% on the test set., Comment: Accepted by ICASSP 2020

In order to solve the problem of teachers’ excessive energy dissipation due to interaction with teaching equipment in traditional classrooms, an interactive and intelligent teaching interface is proposed to enable teachers to use the gestures to give students a geometry lesson. The traditional algorithm of gesture recognition mainly consists of feature extraction and classifier, which requires human-designed features. The recognition is mainly based on static gesture or dynamic gesture singular state recognition algorithm. The recognition accuracy is not robust enough and different people Identification results do not have the universality and ease of operation. In order to solve this problem, we propose a multi-state gesture recognition algorithm based on the deep learning network, which combines the large database of hand gestures and the deep learning algorithms. The innovation of this algorithm is as follows: Aiming at the static gesture images, a sequence reduction algorithm is proposed. According to the sequence of dynamic gestures, the first and last frame fixed and intermediate frame traversal combination algorithm are proposed to get the dynamic and static fusion gesture training datasets, and then the dynamic and static fusion datasets are input to the deep learning network GoogLeNet for training. After repeated training, we found the optimal rule of deep learning network training. According to the optimization law, we got GoogLeNet_model which can recognize 23 kinds of dynamic and static fusion gestures, the recognition rate is 97.09%. We use this model in interactive teaching interface, and achieved good application effect.

A language model (LM) is an important part of a speech recognition system. Language model adaptation techniques use a large amount of source domain data and limited target domain data to improve the performance of language models in target domain. Even though text datasets are easy to obtain, there is no public Chinese text dataset for language model adaptation tasks. This paper presents a language model adaptation dataset which consists of four different domains of news data, i.e., sport, stock, fashion, finance. The discrepancy between the domains of data is evaluated. Model combination based adaptation of n-gram is evaluated on the dataset. Three different fine-tuning adaptation methods of recurrent neural network language models (RNNLMs) are evaluated. WER results on AIShell speech data with the language models trained on this dataset are also provided. The absolute WER reduction of lattice rescoring with adapted RNNLM is 4.74%.

The challenge in deep learning for speaker independent speech separation comes from the label ambiguity or permutation problem. Utterance-level permutation invariant training (uPIT) technique solves this problem by minimizing the mean square error (MSE) over all permutations between outputs and targets. It is a state-of-the-art deep learning architecture. However, uPIT only minimizes the chosen permutation with the lowest MSE, not discriminates it with other permutations. This may lead to increase the possibility of remixing the separated sources. In this paper, we propose a uPIT with discriminative learning (uPITDL) method to solve this problem by adding one regularization at the cost function. In other words, we minimize the difference between the outputs of model and their corresponding reference signals. Moreover, the dissimilarity between the prediction and the targets of other sources is maximized. We evaluate the proposed model on WSJ0-2mix dataset. Experimental results show 22.0% and 24.8% relative improvements under both closed and open conditions compared with the uPIT baseline.

As a major branch of speech processing, speech emotion recognition has drawn much attention of researchers. Prior works have proposed a variety of models and feature sets for training a system. In this paper, we propose to use semi-supervised learning with ladder networks to generate robust feature representation for speech emotion recognition. In our method, the input of ladder network is the normalized static acoustic features and is mapped to high level hidden representations. The model is trained to simultaneously minimize the sum of supervised and unsupervised cost functions by back-propagation. The extracted hidden representations are used as emotional features in SVM model for speech emotion recognition. The experimental results, performed on IEMOCAP database, show 2.6% higher performance than denoising auto-encoder, and 5.3% than the static acoustic features.

Conventional continuous emotion recognition consists of feature extraction step followed by regression step. However, the objective of the two steps is not consistent as they are parted. Besides, there is still no consensus about appropriate emotional features. In this study, we propose an end-to-end continuous emotion recognition framework which merges feature extraction and regressor into a unified system. We employ 3D convolutional networks with Long Short-Term Memory Neutral Network (ConvLSTM) to handle spatiotemporal information for continuous emotion recognition. This model is applied on AVEC 2017 database. The experiment results reveal that ConvLSTM model makes a positive effect on the performance improvement, which outperforms the baseline results for arousal of 0.583 vs 0.525 (baseline) and for valence of 0.h54 vs 0.507.

This paper proposes an adversarial multilingual training to train bottleneck (BN) networks for the target language. A parallel shared-exclusive model is also proposed to train the BN network. Adversarial training is used to ensure that the shared layers can learn language-invariant features. Experiments are conducted on IARPA Babel datasets. The results show that the proposed adversarial multilingual BN model outperforms the baseline BN model by up to 8.9% relative word error rate (WER) reduction. The results also show that the proposed parallel shared-exclusive model achieves up to 1.7% relative WER reduction when compared with the stacked share-exclusive model.

Mobile devices have limited computing power and limited memory. Thus, large deep neural network (DNN) based acoustic models are not well suited for application on mobile devices. In order to alleviate this problem, this paper proposes to compress acoustic models by using knowledge transfer. This approach forces a large teacher model to transfer generalized knowledge to a small student model. The student model is trained with a linear interpolation of hard probabilities and soft probabilities to learn generalized knowledge from the teacher model. The hard probabilities are generated from a Gaussian mixture model hidden Markov model (GMM-HMM) system. The soft probabilities are computed from a teacher model (DNN or RNN). Experiments on AMI corpus show that a small student model obtains 2.4% relative WER improvement over a large teacher model with almost 7.6 times compression ratio.

In this paper we investigate the performance of Multitask learning (MTL) for the combined model of Convolutional, Long Short-Term Memory and Deep neural Networks (CLDNN) for low resource speech recognition tasks. We trained the multilingual CNN model followed by the MTL using the DNN layers. In the MTL framework the grapheme models are used along with the phone models in the shared hidden layers of deep neural network in order to calculate the state probability. We experimented with universal phone set (UPS) and universal grapheme set (UGS) in the DNN framework and a combination of both UPS and UGS for further accuracy of the overall system. The combined model is implemented on Prediction and Correction (PAC) model making it a multilingual PAC-MTL-CLDNN architecture. We evaluated the improvements on AP16-OLR task and using our proposed model we get 1.8% improvement on Vietnam and 2.5% improvement on Uyghur over the baseline PAC model and MDNN system. We also evaluated that extra grapheme modeling task is still efficient with one hour of training data to get 2.1% improvement on Uyghur over the baseline MDNN system making it highly beneficial for zero resource languages.

The continuous dimensional emotion can depict subtlety and complexity of emotional change, which is an inherently challenging problem with growing attention. This paper presents our automatic prediction of dimensional emotional state for Audio-Visual Emotion Challenge (AVEC 2017), which uses multi-features and fusion across all available modalities. Besides the baseline features provided by the organizers, we also extract other acoustic audio feature sets, appearance features and deep visual features as complementary features. Each type of feature is trained using Long Short-Term Memory Recurrent Neutral Network (LSTM-RNN) for every dimensional emotion prediction separately considering annotation delay and temporal pooling. To overcome overfitting problem, robust models are chosen carefully for individual model. Finally, multimodal emotion fusion is achieved by utilizing Support Vector Regression (SVR) with the estimates from different feature sets in decision level fusion. The experimental results indicate that our extracted features are beneficial to performance improvement and our system design achieves very promising results with Concordant Correlation Coefficient (CCC), which outperform the baseline system on the testing set for arousal of 0.599 vs 0.375 (baseline) and for valence of 0.721 vs 0.466 and for liking 0.295 vs 0.246.