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

In this paper, we propose a low-complexity and accurate noise suppression based on an a priori SNR (Speech to Noise Ratio) model for greater robustness w.r.t. short-term noise-fluctuation. The a priori SNR, the ratio of speech spectra and noise spectra in the spectral domain, represents the difference between speech features and noise features in the feature domain, including the mel-cepstral domain and the logarithmic power spectral domain. This is because logarithmic operations are used for domain conversions. Therefore, an a priori SNR model can easily be expressed in terms of the difference between the speech model and the noise model, which are modeled by the Gaussian mixture models, and it can be generated with low computational cost. By using a priori SNRs accurately estimated on the basis of an a priori SNR model, it is possible to calculate accurate coefficients of noise suppression filters taking into account the variance of noise, without serious increase in computational cost over that of a conventional model-based Wiener filter (MBW). We have conducted in-car speech recognition evaluation using the CENSREC-2 database, and a comparison of the proposed method with a conventional MBW showed that the recognition error rate for all noise environments was reduced by 9%, and that, notably, that for audio-noise environments was reduced by 11%. We show that the proposed method can be processed with low levels of computational and memory resources through implementation on a digital signal processor.

In industry, automatic speech recognition has come to be a competitive feature for embedded products with poor hardware resources. In this work, we propose a tiny end-to-end speech recognition model that is lightweight and easily deployable on edge platforms. First, instead of sophisticated network structures, such as recurrent neural networks, transformers, etc., the model we propose mainly uses convolutional neural networks as its backbone. This ensures that our model is supported by most software development kits for embedded devices. Second, we adopt the basic unit of MobileNet-v3, which performs well in computer vision tasks, and integrate the features of the hidden layer at different scales, thus compressing the number of parameters of the model to less than 1 M and achieving an accuracy greater than that of some traditional models. Third, in order to further reduce the CPU computation, we directly extract acoustic representations from 1-dimensional speech waveforms and use a self-supervised learning approach to encourage the convergence of the model. Finally, to solve some problems where hardware resources are relatively weak, we use a prefix beam search decoder to dynamically extend the search path with an optimized pruning strategy and an additional initialism language model to capture the probability of between-words in advance and thus avoid premature pruning of correct words. In our experiments, according to a number of evaluation categories, our end-to-end model outperformed several tiny speech recognition models used for embedded devices in related work.

In this work, we propose a new automatic speech recognition (ASR) system based on feature learning and an end-to-end training procedure for air traffic control (ATC) systems. The proposed model integrates the feature learning block, recurrent neural network (RNN), and connectionist temporal classification loss to build an end-to-end ASR model. Facing the complex environments of ATC speech, instead of the handcrafted features, a learning block is designed to extract informative features from raw waveforms for acoustic modeling. Both the SincNet and 1D convolution blocks are applied to process the raw waveforms, whose outputs are concatenated to the RNN layers for the temporal modeling. Thanks to the ability to learn representations from raw waveforms, the proposed model can be optimized in a complete end-to-end manner, i.e., from waveform to text. Finally, the multilingual issue in the ATC domain is also considered to achieve the ASR task by constructing a combined vocabulary of Chinese characters and English letters. The proposed approach is validated on a multilingual real-world corpus (ATCSpeech), and the experimental results demonstrate that the proposed approach outperforms other baselines, achieving a 6.9% character error rate.

Recently, robust speech recognition for real-world applications has attracted much attention. This paper proposes a robust speech recognition method based on the teacher-student learning framework for domain adaptation. In particular, the student network will be trained based on a novel optimization criterion defined by the encoder outputs of both teacher and student networks rather than the final output posterior probabilities, which aims to make the noisy audio map to the same embedding space as clean audio, so that the student network is adaptive in the noise domain. Comparative experiments demonstrate that the proposed method obtained good robustness against noise.

Although end-to-end based speech recognition research for Mandarin-English code-switching has attracted increasing interests, it remains challenging due to data scarcity. Meta-learning approach is popular with low-resource modeling using high-resource data, but it does not make full use of low-resource code-switching data. Therefore we propose a two-fold cross-validation training framework combined with meta-learning approach. Experiments on the SEAME corpus demonstrate the effects of our method.

Real-time machine speech translation systems mimic human interpreters and translate incoming speech from a source language to the target language in real-time. Such systems can be achieved by performing low-latency processing in ASR (automatic speech recognition) module before passing the output to MT (machine translation) and TTS (text-to-speech synthesis) modules. Although several studies recently proposed sequence mechanisms for neural incremental ASR (ISR), these frameworks have a more complicated training mechanism than the standard attention-based ASR because they have to decide the incremental step and learn the alignment between speech and text. In this paper, we propose attention-transfer ISR (AT-ISR) that learns the knowledge from attention-based non-incremental ASR for a low delay end-to-end speech recognition. ISR comes with a trade-off between delay and performance, so we investigate how to reduce AT-ISR delay without a significant performance drop. Our experiment shows that AT-ISR achieves a comparable performance to the non-incremental ASR when the incremental recognition begins after the speech utterance reaches 25% of the complete utterance length. Additional experiments to investigate the effect of ISR on translation tasks are also performed. The focus is to find the optimum granularity of the output unit. The results reveal that our end-to-end subword-level ISR resulted in the best translation quality with the lowest WER and the lowest uncovered-word rate.

Speech recognition is a technique that recognizes words and sentences in audio form and converts them into text sentences. Currently, with the advancement of deep learning technologies, speech recognition has achieved very satisfactory results close to human abilities. However, there are still limitations in identification results such as lack of punctuation, capitalization, and standardized numerical data. Vietnamese also contains local words, homonyms, etc, which make it difficult to read and understand the identification results for users as well as to perform the next tasks in Natural Language Processing (NLP). In this paper, we propose to combine the transformer decoder with conditional random field (CRF) to restore punctuation and capitalization for the Vietnamese automatic speech recognition (ASR) output. By chunking input sentences and merging output sequences, it is possible to handle longer strings with greater accuracy. Experiments show that the method proposed in the Vietnamese post-speech recognition dataset delivers the best results.

End-to-end (E2E) multilingual automatic speech recognition (ASR) systems aim to recognize multilingual speeches in a unified framework. In the current E2E multilingual ASR framework, the output prediction for a specific language lacks constraints on the output scope of modeling units. In this paper, a language supervision training strategy is proposed with language masks to constrain the neural network output distribution. To simulate the multilingual ASR scenario with unknown language identity information, a language identification (LID) classifier is applied to estimate the language masks. On four Babel corpora, the proposed E2E multilingual ASR system achieved an average absolute word error rate (WER) reduction of 2.6% compared with the multilingual baseline system.

Performance in Automatic Speech Recognition (ASR) degrades dramatically in noisy environments. To alleviate this problem, a variety of deep networks based on convolutional neural networks and recurrent neural networks were proposed by applying L1 or L2 loss. In this Letter, we propose a new orthogonal gradient penalty (OGP) method for Wasserstein Generative Adversarial Networks (WGAN) applied to denoising and despeeching models. WGAN integrates a multi-task autoencoder which estimates not only speech features but also noise features from noisy speech. While achieving 14.1% improvement in Wasserstein distance convergence rate, the proposed OGP enhanced features are tested in ASR and achieve 9.7%, 8.6%, 6.2%, and 4.8% WER improvements over DDAE, MTAE, R-CED(CNN) and RNN models.

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

This paper demonstrates latent word recurrent neural network language models (LW-RNN-LMs) for enhancing automatic speech recognition (ASR). LW-RNN-LMs are constructed so as to pick up advantages in both recurrent neural network language models (RNN-LMs) and latent word language models (LW-LMs). The RNN-LMs can capture long-range context information and offer strong performance, and the LW-LMs are robust for out-of-domain tasks based on the latent word space modeling. However, the RNN-LMs cannot explicitly capture hidden relationships behind observed words since a concept of a latent variable space is not present. In addition, the LW-LMs cannot take into account long-range relationships between latent words. Our idea is to combine RNN-LM and LW-LM so as to compensate individual disadvantages. The LW-RNN-LMs can support both a latent variable space modeling as well as LW-LMs and a long-range relationship modeling as well as RNN-LMs at the same time. From the viewpoint of RNN-LMs, LW-RNN-LM can be considered as a soft class RNN-LM with a vast latent variable space. In contrast, from the viewpoint of LW-LMs, LW-RNN-LM can be considered as an LW-LM that uses the RNN structure for latent variable modeling instead of an n-gram structure. This paper also details a parameter inference method and two kinds of implementation methods, an n-gram approximation and a Viterbi approximation, for introducing the LW-LM to ASR. Our experiments show effectiveness of LW-RNN-LMs on a perplexity evaluation for the Penn Treebank corpus and an ASR evaluation for Japanese spontaneous speech tasks.

The aim of this research is to support real-time drawingin talking by using multimodal user interface technologies. In this situation, if talking and drawing are considered as commands by mistake during presentation, it will disturb users' natural talking and drawing. To prevent this problem, we introduce two modes of a command mode and a free mode, and explore smooth mode switching techniques that does not interfere with users' natural talking and drawing. We evaluate four techniques. Among them, a technique that specifies the command mode after actions using a pen gesture was the most effective. In this technique, users could quickly draw diagrams, and specifying mode switching didn't interfere with users' natural talk.

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

The long short-term memory recurrent neural network (LSTM) has achieved tremendous success for automatic speech recognition (ASR). However, the complicated gating mechanism of LSTM introduces a massive computational cost and limits the application of LSTM in some scenarios. In this paper, we describe our work on accelerating the decoding speed and improving the decoding accuracy. First, we propose an architecture, which is called Projected Gated Recurrent Unit (PGRU), for ASR tasks, and show that the PGRU can consistently outperform the standard GRU. Second, to improve the PGRU generalization, particularly on large-scale ASR tasks, we propose the Output-gate PGRU (OPGRU). In addition, the time delay neural network (TDNN) and normalization methods are found beneficial for OPGRU. In this paper, we apply the OPGRU for both the acoustic model and recurrent neural network language model (RNN-LM). Finally, we evaluate the PGRU on the total Eval2000 / RT03 test sets, and the proposed OPGRU single ASR system achieves 0.9% / 0.9% absolute (8.2% / 8.6% relative) reduction in word error rate (WER) compared to our previous best LSTM single ASR system. Furthermore, the OPGRU ASR system achieves significant speed-up on both acoustic model and language model rescoring.

For voice-enabled car navigation systems that use a multi-purpose cloud speech recognition service (cloud ASR), utterance classification that is robust against speech recognition errors is needed to realize a user-friendly voice interface. The purpose of this study is to improve the accuracy of utterance classification for voice-enabled car navigation systems when inputs to a classifier are error-prone speech recognition results obtained from a cloud ASR. The role of utterance classification is to predict which car navigation function a user wants to execute from a spontaneous utterance. A cloud ASR causes speech recognition errors due to the noises that occur when traveling in a car, and the errors degrade the accuracy of utterance classification. There are many methods for reducing the number of speech recognition errors by modifying the inside of a speech recognizer. However, application developers cannot apply these methods to cloud ASRs because they cannot customize the ASRs. In this paper, we propose a system for improving the accuracy of utterance classification by modifying both speech-signal inputs to a cloud ASR and recognized-sentence outputs from an ASR. First, our system performs speech enhancement on a user's utterance and then sends both enhanced and non-enhanced speech signals to a cloud ASR. Speech recognition results from both speech signals are merged to reduce the number of recognition errors. Second, to reduce that of utterance classification errors, we propose a data augmentation method, which we call “optimal doping,” where not only accurate transcriptions but also error-prone recognized sentences are added to training data. An evaluation with real user utterances spoken to car navigation products showed that our system reduces the number of utterance classification errors by 54% from a baseline condition. Finally, we propose a semi-automatic upgrading approach for classifiers to benefit from the improved performance of cloud ASRs.

This paper proposes gain relaxation in signal enhancement designed for speech recognition. Gain relaxation selectively applies softer enhancement of a target signal to eliminate potential degradation in speech recognition caused by small undesirable distortion in the target signal components. The softer enhancement is a solution to overlooked performance degradation in signal enhancement combined with speech recognition which is encountered in commercial products with an unaware small local noise source. Evaluation of directional interference suppression with signals recorded by a commercial PC (personal computer) demonstrates that signal enhancement over the input is achieved without sacrificing the performance for clean speech.

In this letter, we explored the usage of spatio-temporal information in one unified framework to improve the performance of multichannel speech recognition. Generalized cross correlation (GCC) is served as spatial feature compensation, and an attention mechanism across time is embedded within long short-term memory (LSTM) neural networks. Experiments on the AMI meeting corpus show that the proposed method provides a 8.2% relative improvement in word error rate (WER) over the model trained directly on the concatenation of multiple microphone outputs.

This paper proposes a novel domain adaptation method that can utilize out-of-domain text resources and partially domain matched text resources in language modeling. A major problem in domain adaptation is that it is hard to obtain adequate adaptation effects from out-of-domain text resources. To tackle the problem, our idea is to carry out model merger in a latent variable space created from latent words language models (LWLMs). The latent variables in the LWLMs are represented as specific words selected from the observed word space, so LWLMs can share a common latent variable space. It enables us to perform flexible mixture modeling with consideration of the latent variable space. This paper presents two types of mixture modeling, i.e., LWLM mixture models and LWLM cross-mixture models. The LWLM mixture models can perform a latent word space mixture modeling to mitigate domain mismatch problem. Furthermore, in the LWLM cross-mixture models, LMs which individually constructed from partially matched text resources are split into two element models, each of which can be subjected to mixture modeling. For the approaches, this paper also describes methods to optimize mixture weights using a validation data set. Experiments show that the mixture in latent word space can achieve performance improvements for both target domain and out-of-domain compared with that in observed word space.

Previous research has shown that transcripts generated by automatic speech recognition (ASR) technologies can improve the listening comprehension of non-native speakers (NNSs). However, we still lack a detailed understanding of how ASR transcripts affect listening comprehension of NNSs. To explore this issue, we conducted two studies. The first study examined how the current presentation of ASR transcripts impacted NNSs' listening comprehension. 20 NNSs engaged in two listening tasks, each in different conditions: C1) audio only and C2) audio+ASR transcripts. The participants pressed a button whenever they encountered a comprehension problem, and explained each problem in the subsequent interviews. From our data analysis, we found that NNSs adopted different strategies when using the ASR transcripts; some followed the transcripts throughout the listening; some only checked them when necessary. NNSs also appeared to face difficulties following imperfect and slightly delayed transcripts while listening to speech - many reported difficulties concentrating on listening/reading or shifting between the two. The second study explored how different display methods of ASR transcripts affected NNSs' listening experiences. We focused on two display methods: 1) accuracy-oriented display which shows transcripts only after the completion of speech input analysis, and 2) speed-oriented display which shows the interim analysis results of speech input. We conducted a laboratory experiment with 22 NNSs who engaged in two listening tasks with ASR transcripts presented via the two display methods. We found that the more the NNSs paid attention to listening to the audio, the more they tended to prefer the speed-oriented transcripts, and vice versa. Mismatched transcripts were found to have negative effects on NNSs' listening comprehension. Our findings have implications for improving the presentation methods of ASR transcripts to more effectively support NNSs.