Tomohiro Tanaka

Kohei Matsuura, Takanori Ashihara, Takafumi Moriya et al.

End-to-end speech summarization (E2E SSum) is a technique to directly generate summary sentences from speech. Compared with the cascade approach, which combines automatic speech recognition (ASR) and text summarization models, the E2E approach is more promising because it mitigates ASR errors, incorporates nonverbal information, and simplifies the overall system. However, since collecting a large amount of paired data (i.e., speech and summary) is difficult, the training data is usually insufficient to train a robust E2E SSum system. In this paper, we present two novel methods that leverage a large amount of external text summarization data for E2E SSum training. The first technique is to utilize a text-to-speech (TTS) system to generate synthesized speech, which is used for E2E SSum training with the text summary. The second is a TTS-free method that directly inputs phoneme sequence instead of synthesized speech to the E2E SSum model. Experiments show that our proposed TTS- and phoneme-based methods improve several metrics on the How2 dataset. In particular, our best system outperforms a previous state-of-the-art one by a large margin (i.e., METEOR score improvements of more than 6 points). To the best of our knowledge, this is the first work to use external language resources for E2E SSum. Moreover, we report a detailed analysis of the How2 dataset to confirm the validity of our proposed E2E SSum system.

Takanori Ashihara, Takafumi Moriya, Kohei Matsuura et al.

Self-supervised learning (SSL) for speech representation has been successfully applied in various downstream tasks, such as speech and speaker recognition. More recently, speech SSL models have also been shown to be beneficial in advancing spoken language understanding tasks, implying that the SSL models have the potential to learn not only acoustic but also linguistic information. In this paper, we aim to clarify if speech SSL techniques can well capture linguistic knowledge. For this purpose, we introduce SpeechGLUE, a speech version of the General Language Understanding Evaluation (GLUE) benchmark. Since GLUE comprises a variety of natural language understanding tasks, SpeechGLUE can elucidate the degree of linguistic ability of speech SSL models. Experiments demonstrate that speech SSL models, although inferior to text-based SSL models, perform better than baselines, suggesting that they can acquire a certain amount of general linguistic knowledge from just unlabeled speech data.

Kohei Matsuura, Takanori Ashihara, Takafumi Moriya et al.

End-to-end speech summarization (E2E SSum) directly summarizes input speech into easy-to-read short sentences with a single model. This approach is promising because it, in contrast to the conventional cascade approach, can utilize full acoustical information and mitigate to the propagation of transcription errors. However, due to the high cost of collecting speech-summary pairs, an E2E SSum model tends to suffer from training data scarcity and output unnatural sentences. To overcome this drawback, we propose for the first time to integrate a pre-trained language model (LM), which is highly capable of generating natural sentences, into the E2E SSum decoder via transfer learning. In addition, to reduce the gap between the independently pre-trained encoder and decoder, we also propose to transfer the baseline E2E SSum encoder instead of the commonly used automatic speech recognition encoder. Experimental results show that the proposed model outperforms baseline and data augmented models.

Takanori Ashihara, Takafumi Moriya, Kohei Matsuura et al.

Self-supervised learning (SSL) is seen as a very promising approach with high performance for several speech downstream tasks. Since the parameters of SSL models are generally so large that training and inference require a lot of memory and computational cost, it is desirable to produce compact SSL models without a significant performance degradation by applying compression methods such as knowledge distillation (KD). Although the KD approach is able to shrink the depth and/or width of SSL model structures, there has been little research on how varying the depth and width impacts the internal representation of the small-footprint model. This paper provides an empirical study that addresses the question. We investigate the performance on SUPERB while varying the structure and KD methods so as to keep the number of parameters constant; this allows us to analyze the contribution of the representation introduced by varying the model architecture. Experiments demonstrate that a certain depth is essential for solving content-oriented tasks (e.g. automatic speech recognition) accurately, whereas a certain width is necessary for achieving high performance on several speaker-oriented tasks (e.g. speaker identification). Based on these observations, we identify, for SUPERB, a more compressed model with better performance than previous studies.

Ryo Masumura, Naoki Makishima, Taiga Yamane et al.

This paper proposes a novel automatic speech recognition (ASR) system that can transcribe individual speaker's speech while identifying whether they are target or non-target speakers from multi-talker overlapped speech. Target-speaker ASR systems are a promising way to only transcribe a target speaker's speech by enrolling the target speaker's information. However, in conversational ASR applications, transcribing both the target speaker's speech and non-target speakers' ones is often required to understand interactive information. To naturally consider both target and non-target speakers in a single ASR model, our idea is to extend autoregressive modeling-based multi-talker ASR systems to utilize the enrollment speech of the target speaker. Our proposed ASR is performed by recursively generating both textual tokens and tokens that represent target or non-target speakers. Our experiments demonstrate the effectiveness of our proposed method.

Ryota Yoshihashi, Shuhei Nishimura, Dai Yonebayashi et al.

Siamese-network-based self-supervised learning (SSL) suffers from slow convergence and instability in training. To alleviate this, we propose a framework to exploit intermediate self-supervisions in each stage of deep nets, called the Ladder Siamese Network. Our self-supervised losses encourage the intermediate layers to be consistent with different data augmentations to single samples, which facilitates training progress and enhances the discriminative ability of the intermediate layers themselves. While some existing work has already utilized multi-level self supervisions in SSL, ours is different in that 1) we reveal its usefulness with non-contrastive Siamese frameworks in both theoretical and empirical viewpoints, and 2) ours improves image-level classification, instance-level detection, and pixel-level segmentation simultaneously. Experiments show that the proposed framework can improve BYOL baselines by 1.0% points in ImageNet linear classification, 1.2% points in COCO detection, and 3.1% points in PASCAL VOC segmentation. In comparison with the state-of-the-art methods, our Ladder-based model achieves competitive and balanced performances in all tested benchmarks without causing large degradation in one.

Satoshi Suzuki, Shin'ya Yamaguchi, Shoichiro Takeda et al.

Contrastive pre-trained vision-language models, such as CLIP, demonstrate strong generalization abilities in zero-shot classification by leveraging embeddings extracted from image and text encoders. This paper aims to robustly fine-tune these vision-language models on in-distribution (ID) data without compromising their generalization abilities in out-of-distribution (OOD) and zero-shot settings. Current robust fine-tuning methods tackle this challenge by reusing contrastive learning, which was used in pre-training, for fine-tuning. However, we found that these methods distort the geometric structure of the embeddings, which plays a crucial role in the generalization of vision-language models, resulting in limited OOD and zero-shot performance. To address this, we propose Difference Vector Equalization (DiVE), which preserves the geometric structure during fine-tuning. The idea behind DiVE is to constrain difference vectors, each of which is obtained by subtracting the embeddings extracted from the pre-trained and fine-tuning models for the same data sample. By constraining the difference vectors to be equal across various data samples, we effectively preserve the geometric structure. Therefore, we introduce two losses: average vector loss (AVL) and pairwise vector loss (PVL). AVL preserves the geometric structure globally by constraining difference vectors to be equal to their weighted average. PVL preserves the geometric structure locally by ensuring a consistent multimodal alignment. Our experiments demonstrate that DiVE effectively preserves the geometric structure, achieving strong results across ID, OOD, and zero-shot metrics.

Ryo Masumura, Shota Orihashi, Mana Ihori et al.

This paper proposes a joint modeling method of the Big Five, which has long been studied, and HEXACO, which has recently attracted attention in psychology, for automatically recognizing apparent personality traits from multimodal human behavior. Most previous studies have used the Big Five for multimodal apparent personality-trait recognition. However, no study has focused on apparent HEXACO which can evaluate an Honesty-Humility trait related to displaced aggression and vengefulness, social-dominance orientation, etc. In addition, the relationships between the Big Five and HEXACO when modeled by machine learning have not been clarified. We expect awareness of multimodal human behavior to improve by considering these relationships. The key advance of our proposed method is to optimize jointly recognizing the Big Five and HEXACO. Experiments using a self-introduction video dataset demonstrate that the proposed method can effectively recognize the Big Five and HEXACO.

Tomohiro Tanaka, Narumasa Tsutsumida

Floods are increasingly frequent natural disasters causing extensive human and economic damage, highlighting the critical need for rapid and accurate flood inundation mapping. While remote sensing technologies have advanced flood monitoring capabilities, operational challenges persist: single-sensor approaches face weather-dependent data availability and limited revisit periods, while multi-sensor fusion methods require substantial computational resources and large-scale labeled datasets. To address these limitations, this study introduces a novel sensor-flexible flood detection methodology by fine-tuning Presto, a lightweight ($\sim$0.4M parameters) multi-modal pre-trained transformer that processes both Synthetic Aperture Radar (SAR) and multispectral (MS) data at the pixel level. Our approach uniquely enables flood mapping using SAR-only, MS-only, or combined SAR+MS inputs through a single model architecture, addressing the critical operational need for rapid response with whatever sensor data becomes available first during disasters. We evaluated our method on the Sen1Floods11 dataset against the large-scale Prithvi-100M baseline ($\sim$100M parameters) across three realistic data availability scenarios. The proposed model achieved superior performance with an F1 score of 0.896 and mIoU of 0.886 in the optimal sensor-fusion scenario, outperforming the established baseline. Crucially, the model demonstrated robustness by maintaining effective performance in MS-only scenarios (F1: 0.893) and functional capabilities in challenging SAR-only conditions (F1: 0.718), confirming the advantage of multi-modal pre-training for operational flood mapping. Our parameter-efficient, sensor-flexible approach offers an accessible and robust solution for real-world disaster scenarios requiring immediate flood extent assessment regardless of sensor availability constraints.

Taiga Yamane, Satoshi Suzuki, Ryo Masumura et al.

Multi-View Pedestrian Detection (MVPD) aims to detect pedestrians in the form of a bird's eye view (BEV) from multi-view images. In MVPD, end-to-end trainable deep learning methods have progressed greatly. However, they often struggle to detect pedestrians with consistently small or large scales in views or with vastly different scales between views. This is because they do not exploit multi-scale image features to generate the BEV feature and detect pedestrians. To overcome this problem, we propose a novel MVPD method, called Multi-Scale Multi-View Detection (MSMVD). MSMVD generates multi-scale BEV features by projecting multi-scale image features extracted from individual views into the BEV space, scale-by-scale. Each of these BEV features inherits the properties of its corresponding scale image features from multiple views. Therefore, these BEV features help the precise detection of pedestrians with consistently small or large scales in views. Then, MSMVD combines information at different scales of multiple views by processing the multi-scale BEV features using a feature pyramid network. This improves the detection of pedestrians with vastly different scales between views. Extensive experiments demonstrate that exploiting multi-scale image features via multi-scale BEV features greatly improves the detection performance, and MSMVD outperforms the previous highest MODA by $4.5$ points on the GMVD dataset.

Shuntaro Okada, Kenji Doi, Ryota Yoshihashi et al.

We propose a general approach to optimize noise schedules for training and sampling in diffusion models. Our approach optimizes the noise schedules to ensure a constant rate of change in the probability distribution of diffused data throughout the diffusion process. Any distance metric for measuring the probability-distributional change is applicable to our approach, and we introduce three distance metrics. We evaluated the effectiveness of our approach on unconditional and class-conditional image-generation tasks using the LSUN (Horse, Bedroom, Church), ImageNet, FFHQ, and CIFAR10 datasets. Through extensive experiments, we confirmed that our approach broadly improves the performance of pixel-space and latent-space diffusion models regardless of the dataset, sampler, and number of function evaluations ranging from 5 to 250. Notably, by using our approach for optimizing both training and sampling schedules, we achieved a state-of-the-art FID score of 2.03 without sacrificing mode coverage on LSUN Horse 256 $\times$ 256.

Ryota Yoshihashi, Yuya Otsuka, Kenji Doi et al.

The advance of generative models for images has inspired various training techniques for image recognition utilizing synthetic images. In semantic segmentation, one promising approach is extracting pseudo-masks from attention maps in text-to-image diffusion models, which enables real-image-and-annotation-free training. However, the pioneering training method using the diffusion-synthetic images and pseudo-masks, i.e., DiffuMask has limitations in terms of mask quality, scalability, and ranges of applicable domains. To overcome these limitations, this work introduces three techniques for diffusion-synthetic semantic segmentation training. First, reliability-aware robust training, originally used in weakly supervised learning, helps segmentation with insufficient synthetic mask quality. %Second, large-scale pretraining of whole segmentation models, not only backbones, on synthetic ImageNet-1k-class images with pixel-labels benefits downstream segmentation tasks. Second, we introduce prompt augmentation, data augmentation to the prompt text set to scale up and diversify training images with a limited text resources. Finally, LoRA-based adaptation of Stable Diffusion enables the transfer to a distant domain, e.g., auto-driving images. Experiments in PASCAL VOC, ImageNet-S, and Cityscapes show that our method effectively closes gap between real and synthetic training in semantic segmentation.

Takanori Ashihara, Takafumi Moriya, Kohei Matsuura et al.

Self-supervised learning (SSL) has been dramatically successful not only in monolingual but also in cross-lingual settings. However, since the two settings have been studied individually in general, there has been little research focusing on how effective a cross-lingual model is in comparison with a monolingual model. In this paper, we investigate this fundamental question empirically with Japanese automatic speech recognition (ASR) tasks. First, we begin by comparing the ASR performance of cross-lingual and monolingual models for two different language tasks while keeping the acoustic domain as identical as possible. Then, we examine how much unlabeled data collected in Japanese is needed to achieve performance comparable to a cross-lingual model pre-trained with tens of thousands of hours of English and/or multilingual data. Finally, we extensively investigate the effectiveness of SSL in Japanese and demonstrate state-of-the-art performance on multiple ASR tasks. Since there is no comprehensive SSL study for Japanese, we hope this study will guide Japanese SSL research.

Shota Orihashi, Yoshihiro Yamazaki, Naoki Makishima et al.

This paper presents a novel training method for end-to-end scene text recognition. End-to-end scene text recognition offers high recognition accuracy, especially when using the encoder-decoder model based on Transformer. To train a highly accurate end-to-end model, we need to prepare a large image-to-text paired dataset for the target language. However, it is difficult to collect this data, especially for resource-poor languages. To overcome this difficulty, our proposed method utilizes well-prepared large datasets in resource-rich languages such as English, to train the resource-poor encoder-decoder model. Our key idea is to build a model in which the encoder reflects knowledge of multiple languages while the decoder specializes in knowledge of just the resource-poor language. To this end, the proposed method pre-trains the encoder by using a multilingual dataset that combines the resource-poor language's dataset and the resource-rich language's dataset to learn language-invariant knowledge for scene text recognition. The proposed method also pre-trains the decoder by using the resource-poor language's dataset to make the decoder better suited to the resource-poor language. Experiments on Japanese scene text recognition using a small, publicly available dataset demonstrate the effectiveness of the proposed method.

Shota Orihashi, Yoshihiro Yamazaki, Naoki Makishima et al.

This paper presents a novel knowledge distillation method for dialogue sequence labeling. Dialogue sequence labeling is a supervised learning task that estimates labels for each utterance in the target dialogue document, and is useful for many applications such as dialogue act estimation. Accurate labeling is often realized by a hierarchically-structured large model consisting of utterance-level and dialogue-level networks that capture the contexts within an utterance and between utterances, respectively. However, due to its large model size, such a model cannot be deployed on resource-constrained devices. To overcome this difficulty, we focus on knowledge distillation which trains a small model by distilling the knowledge of a large and high performance teacher model. Our key idea is to distill the knowledge while keeping the complex contexts captured by the teacher model. To this end, the proposed method, hierarchical knowledge distillation, trains the small model by distilling not only the probability distribution of the label classification, but also the knowledge of utterance-level and dialogue-level contexts trained in the teacher model by training the model to mimic the teacher model's output in each level. Experiments on dialogue act estimation and call scene segmentation demonstrate the effectiveness of the proposed method.

Tomohiro Tanaka, Ryo Masumura, Mana Ihori et al.

We propose a semi-supervised learning method for building end-to-end rich transcription-style automatic speech recognition (RT-ASR) systems from small-scale rich transcription-style and large-scale common transcription-style datasets. In spontaneous speech tasks, various speech phenomena such as fillers, word fragments, laughter and coughs, etc. are often included. While common transcriptions do not give special awareness to these phenomena, rich transcriptions explicitly convert them into special phenomenon tokens as well as textual tokens. In previous studies, the textual and phenomenon tokens were simultaneously estimated in an end-to-end manner. However, it is difficult to build accurate RT-ASR systems because large-scale rich transcription-style datasets are often unavailable. To solve this problem, our training method uses a limited rich transcription-style dataset and common transcription-style dataset simultaneously. The Key process in our semi-supervised learning is to convert the common transcription-style dataset into a pseudo-rich transcription-style dataset. To this end, we introduce style tokens which control phenomenon tokens are generated or not into transformer-based autoregressive modeling. We use this modeling for generating the pseudo-rich transcription-style datasets and for building RT-ASR system from the pseudo and original datasets. Our experiments on spontaneous ASR tasks showed the effectiveness of the proposed method.

Tomohiro Tanaka, Ryo Masumura, Mana Ihori et al.

We propose a cross-modal transformer-based neural correction models that refines the output of an automatic speech recognition (ASR) system so as to exclude ASR errors. Generally, neural correction models are composed of encoder-decoder networks, which can directly model sequence-to-sequence mapping problems. The most successful method is to use both input speech and its ASR output text as the input contexts for the encoder-decoder networks. However, the conventional method cannot take into account the relationships between these two different modal inputs because the input contexts are separately encoded for each modal. To effectively leverage the correlated information between the two different modal inputs, our proposed models encode two different contexts jointly on the basis of cross-modal self-attention using a transformer. We expect that cross-modal self-attention can effectively capture the relationships between two different modals for refining ASR hypotheses. We also introduce a shallow fusion technique to efficiently integrate the first-pass ASR model and our proposed neural correction model. Experiments on Japanese natural language ASR tasks demonstrated that our proposed models achieve better ASR performance than conventional neural correction models.

Ryo Masumura, Daiki Okamura, Naoki Makishima et al.

In this paper, we present a novel modeling method for single-channel multi-talker overlapped automatic speech recognition (ASR) systems. Fully neural network based end-to-end models have dramatically improved the performance of multi-taker overlapped ASR tasks. One promising approach for end-to-end modeling is autoregressive modeling with serialized output training in which transcriptions of multiple speakers are recursively generated one after another. This enables us to naturally capture relationships between speakers. However, the conventional modeling method cannot explicitly take into account the speaker attributes of individual utterances such as gender and age information. In fact, the performance deteriorates when each speaker is the same gender or is close in age. To address this problem, we propose unified autoregressive modeling for joint end-to-end multi-talker overlapped ASR and speaker attribute estimation. Our key idea is to handle gender and age estimation tasks within the unified autoregressive modeling. In the proposed method, transformer-based autoregressive model recursively generates not only textual tokens but also attribute tokens of each speaker. This enables us to effectively utilize speaker attributes for improving multi-talker overlapped ASR. Experiments on Japanese multi-talker overlapped ASR tasks demonstrate the effectiveness of the proposed method.

Naoki Makishima, Mana Ihori, Tomohiro Tanaka et al.

We present a novel personalized voice activity detection (PVAD) learning method that does not require enrollment data during training. PVAD is a task to detect the speech segments of a specific target speaker at the frame level using enrollment speech of the target speaker. Since PVAD must learn speakers' speech variations to clarify the boundary between speakers, studies on PVAD used large-scale datasets that contain many utterances for each speaker. However, the datasets to train a PVAD model are often limited because substantial cost is needed to prepare such a dataset. In addition, we cannot utilize the datasets used to train the standard VAD because they often lack speaker labels. To solve these problems, our key idea is to use one utterance as both a kind of enrollment speech and an input to the PVAD during training, which enables PVAD training without enrollment speech. In our proposed method, called enrollment-less training, we augment one utterance so as to create variability between the input and the enrollment speech while keeping the speaker identity, which avoids the mismatch between training and inference. Our experimental results demonstrate the efficacy of the method.

Mana Ihori, Naoki Makishima, Tomohiro Tanaka et al.

In this paper, we propose a novel spoken-text-style conversion method that can simultaneously execute multiple style conversion modules such as punctuation restoration and disfluency deletion without preparing matched datasets. In practice, transcriptions generated by automatic speech recognition systems are not highly readable because they often include many disfluencies and do not include punctuation marks. To improve their readability, multiple spoken-text-style conversion modules that individually model a single conversion task are cascaded because matched datasets that simultaneously handle multiple conversion tasks are often unavailable. However, the cascading is unstable against the order of tasks because of the chain of conversion errors. Besides, the computation cost of the cascading must be higher than the single conversion. To execute multiple conversion tasks simultaneously without preparing matched datasets, our key idea is to distinguish individual conversion tasks using the on-off switch. In our proposed zero-shot joint modeling, we switch the individual tasks using multiple switching tokens, enabling us to utilize a zero-shot learning approach to executing simultaneous conversions. Our experiments on joint modeling of disfluency deletion and punctuation restoration demonstrate the effectiveness of our method.

Ryota Yoshihashi, Tomohiro Tanaka, Kenji Doi et al.

In the deployment of scene-text spotting systems on mobile platforms, lightweight models with low computation are preferable. In concept, end-to-end (E2E) text spotting is suitable for such purposes because it performs text detection and recognition in a single model. However, current state-of-the-art E2E methods rely on heavy feature extractors, recurrent sequence modellings, and complex shape aligners to pursue accuracy, which means their computations are still heavy. We explore the opposite direction: How far can we go without bells and whistles in E2E text spotting? To this end, we propose a text-spotting method that consists of simple convolutions and a few post-processes, named Context-Free TextSpotter. Experiments using standard benchmarks show that Context-Free TextSpotter achieves real-time text spotting on a GPU with only three million parameters, which is the smallest and fastest among existing deep text spotters, with an acceptable transcription quality degradation compared to heavier ones. Further, we demonstrate that our text spotter can run on a smartphone with affordable latency, which is valuable for building stand-alone OCR applications.

Naoki Makishima, Mana Ihori, Akihiko Takashima et al.

We present an audio-visual speech separation learning method that considers the correspondence between the separated signals and the visual signals to reflect the speech characteristics during training. Audio-visual speech separation is a technique to estimate the individual speech signals from a mixture using the visual signals of the speakers. Conventional studies on audio-visual speech separation mainly train the separation model on the audio-only loss, which reflects the distance between the source signals and the separated signals. However, conventional losses do not reflect the characteristics of the speech signals, including the speaker's characteristics and phonetic information, which leads to distortion or remaining noise. To address this problem, we propose the cross-modal correspondence (CMC) loss, which is based on the cooccurrence of the speech signal and the visual signal. Since the visual signal is not affected by background noise and contains speaker and phonetic information, using the CMC loss enables the audio-visual speech separation model to remove noise while preserving the speech characteristics. Experimental results demonstrate that the proposed method learns the cooccurrence on the basis of CMC loss, which improves separation performance.

Ryo Masumura, Mana Ihori, Akihiko Takashima et al.

This paper is the first study to apply deep mutual learning (DML) to end-to-end ASR models. In DML, multiple models are trained simultaneously and collaboratively by mimicking each other throughout the training process, which helps to attain the global optimum and prevent models from making over-confident predictions. While previous studies applied DML to simple multi-class classification problems, there are no studies that have used it on more complex sequence-to-sequence mapping problems. For this reason, this paper presents a method to apply DML to state-of-the-art Transformer-based end-to-end ASR models. In particular, we propose to combine DML with recent representative training techniques. i.e., label smoothing, scheduled sampling, and SpecAugment, each of which are essential for powerful end-to-end ASR models. We expect that these training techniques work well with DML because DML has complementary characteristics. We experimented with two setups for Japanese ASR tasks: large-scale modeling and compact modeling. We demonstrate that DML improves the ASR performance of both modeling setups compared with conventional learning methods including knowledge distillation. We also show that combining DML with the existing training techniques effectively improves ASR performance.

Ryo Masumura, Naoki Makishima, Mana Ihori et al.

This paper presents a novel self-supervised learning method for handling conversational documents consisting of transcribed text of human-to-human conversations. One of the key technologies for understanding conversational documents is utterance-level sequential labeling, where labels are estimated from the documents in an utterance-by-utterance manner. The main issue with utterance-level sequential labeling is the difficulty of collecting labeled conversational documents, as manual annotations are very costly. To deal with this issue, we propose large-context conversational representation learning (LC-CRL), a self-supervised learning method specialized for conversational documents. A self-supervised learning task in LC-CRL involves the estimation of an utterance using all the surrounding utterances based on large-context language modeling. In this way, LC-CRL enables us to effectively utilize unlabeled conversational documents and thereby enhances the utterance-level sequential labeling. The results of experiments on scene segmentation tasks using contact center conversational datasets demonstrate the effectiveness of the proposed method.

Ryo Masumura, Naoki Makishima, Mana Ihori et al.

We present a novel large-context end-to-end automatic speech recognition (E2E-ASR) model and its effective training method based on knowledge distillation. Common E2E-ASR models have mainly focused on utterance-level processing in which each utterance is independently transcribed. On the other hand, large-context E2E-ASR models, which take into account long-range sequential contexts beyond utterance boundaries, well handle a sequence of utterances such as discourses and conversations. However, the transformer architecture, which has recently achieved state-of-the-art ASR performance among utterance-level ASR systems, has not yet been introduced into the large-context ASR systems. We can expect that the transformer architecture can be leveraged for effectively capturing not only input speech contexts but also long-range sequential contexts beyond utterance boundaries. Therefore, this paper proposes a hierarchical transformer-based large-context E2E-ASR model that combines the transformer architecture with hierarchical encoder-decoder based large-context modeling. In addition, in order to enable the proposed model to use long-range sequential contexts, we also propose a large-context knowledge distillation that distills the knowledge from a pre-trained large-context language model in the training phase. We evaluate the effectiveness of the proposed model and proposed training method on Japanese discourse ASR tasks.

Mana Ihori, Naoki Makishima, Tomohiro Tanaka et al.

This paper presents a self-supervised learning method for pointer-generator networks to improve spoken-text normalization. Spoken-text normalization that converts spoken-style text into style normalized text is becoming an important technology for improving subsequent processing such as machine translation and summarization. The most successful spoken-text normalization method to date is sequence-to-sequence (seq2seq) mapping using pointer-generator networks that possess a copy mechanism from an input sequence. However, these models require a large amount of paired data of spoken-style text and style normalized text, and it is difficult to prepare such a volume of data. In order to construct spoken-text normalization model from the limited paired data, we focus on self-supervised learning which can utilize unpaired text data to improve seq2seq models. Unfortunately, conventional self-supervised learning methods do not assume that pointer-generator networks are utilized. Therefore, we propose a novel self-supervised learning method, MAsked Pointer-Generator Network (MAPGN). The proposed method can effectively pre-train the pointer-generator network by learning to fill masked tokens using the copy mechanism. Our experiments demonstrate that MAPGN is more effective for pointer-generator networks than the conventional self-supervised learning methods in two spoken-text normalization tasks.

Mana Ihori, Ryo Masumura, Naoki Makishima et al.

This paper presents a novel fusion method for integrating an external language model (LM) into the Transformer based sequence-to-sequence (seq2seq) model. While paired data are basically required to train the seq2seq model, the external LM can be trained with only unpaired data. Thus, it is important to leverage memorized knowledge in the external LM for building the seq2seq model, since it is hard to prepare a large amount of paired data. However, the existing fusion methods assume that the LM is integrated with recurrent neural network-based seq2seq models instead of the Transformer. Therefore, this paper proposes a fusion method that can explicitly utilize network structures in the Transformer. The proposed method, called {\bf memory attentive fusion}, leverages the Transformer-style attention mechanism that repeats source-target attention in a multi-hop manner for reading the memorized knowledge in the LM. Our experiments on two text-style conversion tasks demonstrate that the proposed method performs better than conventional fusion methods.