This study presents a two-stage spoken term detection (STD) method that uses the same STD engine twice and a support vector machine (SVM)-based classifier to verify detected terms from the STD engine's output. In a front-end process, the STD engine is used to pre-index target spoken documents from a keyword list built from an automatic speech recognition result. The STD result includes a set of keywords and their detection intervals (positions) in the spoken documents. For keywords having competitive intervals, we rank them based on the STD matching cost and select the one having the longest duration among competitive detections. The selected keywords are registered in the pre-index. They are then used to train an SVM-based classifier. In a query term search process, a query term is searched by the same STD engine, and the output candidates are verified by the SVM-based classifier. Our proposed two-stage STD method with pre-indexing was evaluated using the NTCIR-10 SpokenDoc-2 STD task and it drastically outperformed the traditional STD method based on dynamic time warping and a confusion network-based index.

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.

Text detection is a crucial pre-processing step in optical character recognition (OCR) for the accurate recognition of text, including both fonts and handwritten characters, in documents. While current deep learning-based text detection tools can detect text regions with high accuracy, they often treat multiple lines of text as a single region. To perform line-based character recognition, it is necessary to divide the text into individual lines, which requires a line detection technique. This paper focuses on the development of a new approach to single-line detection in OCR that is based on the existing Character Region Awareness For Text detection (CRAFT) model and incorporates a deep neural network specialized in line segmentation. However, this new method may still detect multiple lines as a single text region when multi-line text with narrow spacing is present. To address this, we also introduce a post-processing algorithm to detect single text regions using the output of the single-line segmentation. Our proposed method successfully detects single lines, even in multi-line text with narrow line spacing, and hence improves the accuracy of OCR.

There have been lots of previous studies on fluency evaluation of spontaneous speech. However, most of them focus on lexical cues, and little emphasis is placed on how diverse acoustic features and deep end-to-end models contribute to improving the performance. In this paper, we describe multi-layer neural network to investigate not only lexical features extracted from transcription, but also consider utterance-level acoustic features from audio data. We also conduct the experiments to investigate the performance of end-to-end approaches with mel-spectrogram in this task. As the speech fluency evaluation task, we evaluate our proposed method in two binary classification tasks of fluent speech detection and disfluent speech detection. Speech data of around 10 seconds duration each with the annotation of the three classes of “fluent,” “neutral,” and “disfluent” is used for evaluation. According to the two way splits of those three classes, the task of fluent speech detection is defined as binary classification of fluent vs. neutral and disfluent, while that of disfluent speech detection is defined as binary classification of fluent and neutral vs. disfluent. We then conduct experiments with the purpose of comparative evaluation of multi-layer neural network with diverse features as well as end-to-end models. For the fluent speech detection, in the comparison of utterance-level disfluency-based, prosodic, and acoustic features with multi-layer neural network, disfluency-based and prosodic features only are better. More specifically, the performance improved a lot when removing all of the acoustic features from the full set of features, while the performance is damaged a lot if fillers related features are removed. Overall, however, the end-to-end Transformer+VGGNet model with mel-spectrogram achieves the best results. For the disfluent speech detection, the multi-layer neural network using disfluency-based, prosodic, and acoustic features without fillers achieves the best results. The end-to-end Transformer+VGGNet architecture also obtains high scores, whereas it is exceeded by the best results with the multi-layer neural network with significant difference. Thus, unlike in the fluent speech detection, disfluency-based and prosodic features other than fillers are still necessary in the disfluent speech detection.

This paper presents speech-driven Web retrieval models which accept spoken search topics (queries) in the NTCIR-3 Web retrieval task. The major focus of this paper is on improving speech recognition accuracy of spoken queries and then improving retrieval accuracy in speech-driven Web retrieval. We experimentally evaluated the techniques of combining outputs of multiple LVCSR models in recognition of spoken queries. As model combination techniques, we compared the SVM learning technique with conventional voting schemes such as ROVER. In addition, for investigating the effects on the retrieval performance in vocabulary size of the language model, we prepared two kinds of language models: the one's vocabulary size was 20,000, the other's one was 60,000. Then, we evaluated the differences in the recognition rates of the spoken queries and the retrieval performance. We showed that the techniques of multiple LVCSR model combination could achieve improvement both in speech recognition and retrieval accuracies in speech-driven text retrieval. Comparing with the retrieval accuracies when an LM with a 20,000/60,000 vocabulary size is used in an LVCSR system, we found that the larger the vocabulary size is, the better the retrieval accuracy is.

This paper describes an accurate unsupervised speaker adaptation method for lecture style spontaneous speech recognition using multiple LVCSR systems. In an unsupervised speaker adaptation framework, the improvement of recognition performance by adapting acoustic models remarkably depends on the accuracy of labels such as phonemes and syllables. Therefore, extraction of the adaptation data guided by confidence measure is effective for unsupervised adaptation. In this paper, we looked for the high confidence portions based on the agreement between two LVCSR systems, adapted acoustic models using the portions attached with high accurate labels, and then improved the recognition accuracy. We applied our method to the Corpus of Spontaneous Japanese (CSJ) and the method improved the recognition rate by about 2.1% in comparison with a traditional method.

This study proposes a two-pass spoken term detection (STD) method. The first pass uses a phoneme-based dynamic time warping (DTW)-based STD, and the second pass recomputes detection scores produced by the first pass using conditional random fields (CRF)-based triphone detectors. In the second-pass, we treat STD as a sequence labeling problem. We use CRF-based triphone detection models based on features generated from multiple types of phoneme-based transcriptions. The models train recognition error patterns such as phoneme-to-phoneme confusions in the CRF framework. Consequently, the models can detect a triphone comprising a query term with a detection probability. In the experimental evaluation of two types of test collections, the CRF-based approach worked well in the re-ranking process for the DTW-based detections. CRF-based re-ranking showed 2.1% and 2.0% absolute improvements in F-measure for each of the two test collections.