Berlin Chen

Bi-Cheng Yan, Hsin-Wei Wang, Yi-Cheng Wang et al.

Automatic pronunciation assessment (APA) manages to quantify the pronunciation proficiency of a second language (L2) learner in a language. Prevailing approaches to APA normally leverage neural models trained with a regression loss function, such as the mean-squared error (MSE) loss, for proficiency level prediction. Despite most regression models can effectively capture the ordinality of proficiency levels in the feature space, they are confronted with a primary obstacle that different phoneme categories with the same proficiency level are inevitably forced to be close to each other, retaining less phoneme-discriminative information. On account of this, we devise a phonemic contrast ordinal (PCO) loss for training regression-based APA models, which aims to preserve better phonemic distinctions between phoneme categories meanwhile considering ordinal relationships of the regression target output. Specifically, we introduce a phoneme-distinct regularizer into the MSE loss, which encourages feature representations of different phoneme categories to be far apart while simultaneously pulling closer the representations belonging to the same phoneme category by means of weighted distances. An extensive set of experiments carried out on the speechocean762 benchmark dataset suggest the feasibility and effectiveness of our model in relation to some existing state-of-the-art models.

Jiun-Ting Li, Bi-Cheng Yan, Tien-Hong Lo et al.

Automated speaking assessment in conversation tests (ASAC) aims to evaluate the overall speaking proficiency of an L2 (second-language) speaker in a setting where an interlocutor interacts with one or more candidates. Although prior ASAC approaches have shown promising performance on their respective datasets, there is still a dearth of research specifically focused on incorporating the coherence of the logical flow within a conversation into the grading model. To address this critical challenge, we propose a hierarchical graph model that aptly incorporates both broad inter-response interactions (e.g., discourse relations) and nuanced semantic information (e.g., semantic words and speaker intents), which is subsequently fused with contextual information for the final prediction. Extensive experimental results on the NICT-JLE benchmark dataset suggest that our proposed modeling approach can yield considerable improvements in prediction accuracy with respect to various assessment metrics, as compared to some strong baselines. This also sheds light on the importance of investigating coherence-related facets of spoken responses in ASAC.

Fu-An Chao, Berlin Chen

Prior efforts in building computer-assisted pronunciation training (CAPT) systems often treat automatic pronunciation assessment (APA) and mispronunciation detection and diagnosis (MDD) as separate fronts: the former aims to provide multiple pronunciation aspect scores across diverse linguistic levels, while the latter focuses instead on pinpointing the precise phonetic pronunciation errors made by non-native language learners. However, it is generally expected that a full-fledged CAPT system should perform both functionalities simultaneously and efficiently. In response to this surging demand, we in this work first propose HMamba, a novel CAPT approach that seamlessly integrates APA and MDD tasks in parallel. In addition, we introduce a novel loss function, decoupled cross-entropy loss (deXent), specifically tailored for MDD to facilitate better-supervised learning for detecting mispronounced phones, thereby enhancing overall performance. A comprehensive set of empirical results on the speechocean762 benchmark dataset demonstrates the effectiveness of our approach on APA. Notably, our proposed approach also yields a considerable improvement in MDD performance over a strong baseline, achieving an F1-score of 63.85%. Our codes are made available at https://github.com/Fuann/hmamba

Hung-Yang Sung, Chien-Chun Wang, Kuan-Tang Huang et al.

Automatic speech recognition (ASR) for low-resource languages such as Taiwanese Hokkien is difficult due to the scarcity of annotated data. However, direct fine-tuning on Han-character transcriptions often fails to capture detailed phonetic and tonal cues, while training only on romanization lacks lexical and syntactic coverage. In addition, prior studies have rarely explored staged strategies that integrate both annotation types. To address this gap, we present CLiFT-ASR, a cross-lingual fine-tuning framework that builds on Mandarin HuBERT models and progressively adapts them to Taiwanese Hokkien. The framework employs a two-stage process in which it first learns acoustic and tonal representations from phonetic Tai-lo annotations and then captures vocabulary and syntax from Han-character transcriptions. This progressive adaptation enables effective alignment between speech sounds and orthographic structures. Experiments on the TAT-MOE corpus demonstrate that CLiFT-ASR achieves a 24.88\% relative reduction in character error rate (CER) compared with strong baselines. The results indicate that CLiFT-ASR provides an effective and parameter-efficient solution for Taiwanese Hokkien ASR and that it has potential to benefit other low-resource language scenarios.

Yi-Cheng Wang, Hsin-Wei Wang, Bi-Cheng Yan et al.

End-to-end automatic speech recognition (E2E ASR) systems often suffer from mistranscription of domain-specific phrases, such as named entities, sometimes leading to catastrophic failures in downstream tasks. A family of fast and lightweight named entity correction (NEC) models for ASR have recently been proposed, which normally build on phonetic-level edit distance algorithms and have shown impressive NEC performance. However, as the named entity (NE) list grows, the problems of phonetic confusion in the NE list are exacerbated; for example, homophone ambiguities increase substantially. In view of this, we proposed a novel Description Augmented Named entity CorrEctoR (dubbed DANCER), which leverages entity descriptions to provide additional information to facilitate mitigation of phonetic confusion for NEC on ASR transcription. To this end, an efficient entity description augmented masked language model (EDA-MLM) comprised of a dense retrieval model is introduced, enabling MLM to adapt swiftly to domain-specific entities for the NEC task. A series of experiments conducted on the AISHELL-1 and Homophone datasets confirm the effectiveness of our modeling approach. DANCER outperforms a strong baseline, the phonetic edit-distance-based NEC model (PED-NEC), by a character error rate (CER) reduction of about 7% relatively on AISHELL-1 for named entities. More notably, when tested on Homophone that contain named entities of high phonetic confusion, DANCER offers a more pronounced CER reduction of 46% relatively over PED-NEC for named entities.

Jhen-Ke Lin, Hao-Chien Lu, Chung-Chun Wang et al.

Verbatim transcription for automatic speaking assessment demands accurate capture of disfluencies, crucial for downstream tasks like error analysis and feedback. However, many ASR systems discard or generalize hesitations, losing important acoustic details. We fine-tune Whisper models on the Speak & Improve 2025 corpus using low-rank adaptation (LoRA), without recourse to external audio training data. We compare three annotation schemes: removing hesitations (Pure), generic tags (Rich), and acoustically precise fillers inferred by Gemini 2.0 Flash from existing audio-transcript pairs (Extra). Our challenge system achieved 6.47% WER (Pure) and 5.81% WER (Extra). Post-challenge experiments reveal that fine-tuning Whisper Large V3 Turbo with the "Extra" scheme yielded a 5.5% WER, an 11.3% relative improvement over the "Pure" scheme (6.2% WER). This demonstrates that explicit, realistic filled-pause labeling significantly enhances ASR accuracy for verbatim L2 speech transcription.

Jiun-Ting Li, Bi-Cheng Yan, Yi-Cheng Wang et al.

Automatic pronunciation assessment (APA) analyzes second-language (L2) learners' speech by providing fine-grained pronunciation feedback at various linguistic levels. Most existing efforts on APA typically adopt segmental-level features as inputs and predict pronunciation scores at different granularities via hierarchical (or parallel) pronunciation modeling. This, however, inevitably causes assessments across linguistic levels (e.g., phone, word, and utterance) to rely solely on phoneme-level pronunciation features, nearly sidelining supra-segmental pronunciation cues. To address this limitation, we introduce multi-task pretraining (MTP) for APA, a simple yet effective strategy that attempts to capture long-term temporal pronunciation cues while strengthening the intrinsic structures within an utterance via the objective of reconstructing input features. Specifically, for a phoneme-level encoder of an APA model, the proposed MTP strategy randomly masks segmental-level pronunciation features and reconstructs the masked ones based on their surrounding pronunciation context. Furthermore, current APA systems lack integration with automated speaking assessment (ASA), limiting holistic proficiency evaluation. Drawing on empirical studies and prior knowledge in ASA, our framework bridges this gap by incorporating handcrafted features (HCFs), such as fluency (speech rate, silence duration) and stress (pitch accent strength), derived from human-designed formulas via regressors to generate interpretable proficiency scores. Experiments on speechocean762 show improved pronunciation scoring and ASA proficiency correlation, enabling targeted training and comprehensive proficiency assessment.

Yu-Hsuan Fang, Tien-Hong Lo, Yao-Ting Sung et al.

Traditional Automated Speaking Assessment (ASA) systems exhibit inherent modality limitations: text-based approaches lack acoustic information while audio-based methods miss semantic context. Multimodal Large Language Models (MLLM) offer unprecedented opportunities for comprehensive ASA by simultaneously processing audio and text within unified frameworks. This paper presents a very first systematic study of MLLM for comprehensive ASA, demonstrating the superior performance of MLLM across the aspects of content and language use . However, assessment on the delivery aspect reveals unique challenges, which is deemed to require specialized training strategies. We thus propose Speech-First Multimodal Training (SFMT), leveraging a curriculum learning principle to establish more robust modeling foundations of speech before cross-modal synergetic fusion. A series of experiments on a benchmark dataset show MLLM-based systems can elevate the holistic assessment performance from a PCC value of 0.783 to 0.846. In particular, SFMT excels in the evaluation of the delivery aspect, achieving an absolute accuracy improvement of 4% over conventional training approaches, which also paves a new avenue for ASA.

Hsuan-Yu Wang, Pei-Ying Lee, Berlin Chen

In this paper, we investigate the impact of incorporating timestamp-based alignment between Automatic Speech Recognition (ASR) transcripts and Speaker Diarization (SD) outputs on Speech Emotion Recognition (SER) accuracy. Misalignment between these two modalities often reduces the reliability of multimodal emotion recognition systems, particularly in conversational contexts. To address this issue, we introduce an alignment pipeline utilizing pre-trained ASR and speaker diarization models, systematically synchronizing timestamps to generate accurately labeled speaker segments. Our multimodal approach combines textual embeddings extracted via RoBERTa with audio embeddings from Wav2Vec, leveraging cross-attention fusion enhanced by a gating mechanism. Experimental evaluations on the IEMOCAP benchmark dataset demonstrate that precise timestamp alignment improves SER accuracy, outperforming baseline methods that lack synchronization. The results highlight the critical importance of temporal alignment, demonstrating its effectiveness in enhancing overall emotion recognition accuracy and providing a foundation for robust multimodal emotion analysis.

Tzu-Hsuan Yang, Yue-Yang He, Berlin Chen

Effective pronunciation feedback is critical in second language (L2) learning, for which computer-assisted pronunciation training (CAPT) systems often encompass two key tasks: automatic pronunciation assessment (APA) and mispronunciation detection and diagnosis (MDD). Recent work has shown that joint modeling of these two tasks can yield mutual benefits. Our unified framework leverages Mamba, a selective state space model (SSM), while integrating phonological features and think token strategies to jointly enhance interpretability and fine-grained temporal reasoning in APA and MDD. To our knowledge, this is the first study to combine phonological attribution, SSM-based modeling, and prompting in CAPT. A series of experiments conducted on the speechocean762 benchmark demonstrate that our model consistently outperforms prior methods, particularly on the MDD task.

Tzu-Ting Yang, Hsin-Wei Wang, Yi-Cheng Wang et al.

Code-switching-where multilingual speakers alternately switch between languages during conversations-still poses significant challenges to end-to-end (E2E) automatic speech recognition (ASR) systems due to phenomena of both acoustic and semantic confusion. This issue arises because ASR systems struggle to handle the rapid alternation of languages effectively, which often leads to significant performance degradation. Our main contributions are at least threefold: First, we incorporate language identification (LID) information into several intermediate layers of the encoder, aiming to enrich output embeddings with more detailed language information. Secondly, through the novel application of language boundary alignment loss, the subsequent ASR modules are enabled to more effectively utilize the knowledge of internal language posteriors. Third, we explore the feasibility of using language posteriors to facilitate deep interaction between shared encoder and language-specific encoders. Through comprehensive experiments on the SEAME corpus, we have verified that our proposed method outperforms the prior-art method, disentangle based mixture-of-experts (D-MoE), further enhancing the acuity of the encoder to languages.

Chung-Wen Wu, Berlin Chen

Automatic Speech Assessment (ASA) has seen notable advancements with the utilization of self-supervised features (SSL) in recent research. However, a key challenge in ASA lies in the imbalanced distribution of data, particularly evident in English test datasets. To address this challenge, we approach ASA as an ordinal classification task, introducing Weighted Vectors Ranking Similarity (W-RankSim) as a novel regularization technique. W-RankSim encourages closer proximity of weighted vectors in the output layer for similar classes, implying that feature vectors with similar labels would be gradually nudged closer to each other as they converge towards corresponding weighted vectors. Extensive experimental evaluations confirm the effectiveness of our approach in improving ordinal classification performance for ASA. Furthermore, we propose a hybrid model that combines SSL and handcrafted features, showcasing how the inclusion of handcrafted features enhances performance in an ASA system.

Tzu-Ting Yang, Hsin-Wei Wang, Berlin Chen

In recent years, end-to-end speech recognition has emerged as a technology that integrates the acoustic, pronunciation dictionary, and language model components of the traditional Automatic Speech Recognition model. It is possible to achieve human-like recognition without the need to build a pronunciation dictionary in advance. However, due to the relative scarcity of training data on code-switching, the performance of ASR models tends to degrade drastically when encountering this phenomenon. Most past studies have simplified the learning complexity of the model by splitting the code-switching task into multiple tasks dealing with a single language and then learning the domain-specific knowledge of each language separately. Therefore, in this paper, we attempt to introduce language identification information into the middle layer of the ASR model's encoder. We aim to generate acoustic features that imply language distinctions in a more implicit way, reducing the model's confusion when dealing with language switching.

Hsin-Wei Wang, Bi-Cheng Yan, Hsuan-Sheng Chiu et al.

End-to-end (E2E) neural modeling has emerged as one predominant school of thought to develop computer-assisted language training (CAPT) systems, showing competitive performance to conventional pronunciation-scoring based methods. However, current E2E neural methods for CAPT are faced with at least two pivotal challenges. On one hand, most of the E2E methods operate in an autoregressive manner with left-to-right beam search to dictate the pronunciations of an L2 learners. This however leads to very slow inference speed, which inevitably hinders their practical use. On the other hand, E2E neural methods are normally data greedy and meanwhile an insufficient amount of nonnative training data would often reduce their efficacy on mispronunciation detection and diagnosis (MD&D). In response, we put forward a novel MD&D method that leverages non-autoregressive (NAR) E2E neural modeling to dramatically speed up the inference time while maintaining performance in line with the conventional E2E neural methods. In addition, we design and develop a pronunciation modeling network stacked on top of the NAR E2E models of our method to further boost the effectiveness of MD&D. Empirical experiments conducted on the L2-ARCTIC English dataset seems to validate the feasibility of our method, in comparison to some top-of-the-line E2E models and an iconic pronunciation-scoring based method built on a DNN-HMM acoustic model.

Tien-Hong Lo, Yao-Ting Sung, Berlin Chen

Recently, end-to-end (E2E) models, which allow to take spectral vector sequences of L2 (second-language) learners' utterances as input and produce the corresponding phone-level sequences as output, have attracted much research attention in developing mispronunciation detection (MD) systems. However, due to the lack of sufficient labeled speech data of L2 speakers for model estimation, E2E MD models are prone to overfitting in relation to conventional ones that are built on DNN-HMM acoustic models. To alleviate this critical issue, we in this paper propose two modeling strategies to enhance the discrimination capability of E2E MD models, each of which can implicitly leverage the phonetic and phonological traits encoded in a pretrained acoustic model and contained within reference transcripts of the training data, respectively. The first one is input augmentation, which aims to distill knowledge about phonetic discrimination from a DNN-HMM acoustic model. The second one is label augmentation, which manages to capture more phonological patterns from the transcripts of training data. A series of empirical experiments conducted on the L2-ARCTIC English dataset seem to confirm the efficacy of our E2E MD model when compared to some top-of-the-line E2E MD models and a classic pronunciation-scoring based method built on a DNN-HMM acoustic model.

Shao-Wei Fan Jiang, Bi-Cheng Yan, Tien-Hong Lo et al.

With the acceleration of globalization, more and more people are willing or required to learn second languages (L2). One of the major remaining challenges facing current mispronunciation and diagnosis (MDD) models for use in computer-assisted pronunciation training (CAPT) is to handle speech from L2 learners with a diverse set of accents. In this paper, we set out to mitigate the adverse effects of accent variety in building an L2 English MDD system with end-to-end (E2E) neural models. To this end, we first propose an effective modeling framework that infuses accent features into an E2E MDD model, thereby making the model more accent-aware. Going a step further, we design and present disparate accent-aware modules to perform accent-aware modulation of acoustic features in a finer-grained manner, so as to enhance the discriminating capability of the resulting MDD model. Extensive sets of experiments conducted on the L2-ARCTIC benchmark dataset show the merits of our MDD model, in comparison to some existing E2E-based strong baselines and the celebrated pronunciation scoring based method.

Shih-Hsuan Chiu, Tien-Hong Lo, Fu-An Chao et al.

How to effectively incorporate cross-utterance information cues into a neural language model (LM) has emerged as one of the intriguing issues for automatic speech recognition (ASR). Existing research efforts on improving contextualization of an LM typically regard previous utterances as a sequence of additional input and may fail to capture complex global structural dependencies among these utterances. In view of this, we in this paper seek to represent the historical context information of an utterance as graph-structured data so as to distill cross-utterances, global word interaction relationships. To this end, we apply a graph convolutional network (GCN) on the resulting graph to obtain the corresponding GCN embeddings of historical words. GCN has recently found its versatile applications on social-network analysis, text summarization, and among others due mainly to its ability of effectively capturing rich relational information among elements. However, GCN remains largely underexplored in the context of ASR, especially for dealing with conversational speech. In addition, we frame ASR N-best reranking as a prediction problem, leveraging bidirectional encoder representations from transformers (BERT) as the vehicle to not only seize the local intrinsic word regularity patterns inherent in a candidate hypothesis but also incorporate the cross-utterance, historical word interaction cues distilled by GCN for promoting performance. Extensive experiments conducted on the AMI benchmark dataset seem to confirm the pragmatic utility of our methods, in relation to some current top-of-the-line methods.

Shih-Hsuan Chiu, Berlin Chen

More recently, Bidirectional Encoder Representations from Transformers (BERT) was proposed and has achieved impressive success on many natural language processing (NLP) tasks such as question answering and language understanding, due mainly to its effective pre-training then fine-tuning paradigm as well as strong local contextual modeling ability. In view of the above, this paper presents a novel instantiation of the BERT-based contextualized language models (LMs) for use in reranking of N-best hypotheses produced by automatic speech recognition (ASR). To this end, we frame N-best hypothesis reranking with BERT as a prediction problem, which aims to predict the oracle hypothesis that has the lowest word error rate (WER) given the N-best hypotheses (denoted by PBERT). In particular, we also explore to capitalize on task-specific global topic information in an unsupervised manner to assist PBERT in N-best hypothesis reranking (denoted by TPBERT). Extensive experiments conducted on the AMI benchmark corpus demonstrate the effectiveness and feasibility of our methods in comparison to the conventional autoregressive models like the recurrent neural network (RNN) and a recently proposed method that employed BERT to compute pseudo-log-likelihood (PLL) scores for N-best hypothesis reranking.

Bi-Cheng Yan, Berlin Chen

Mispronunciation detection and diagnosis (MDD) is designed to identify pronunciation errors and provide instructive feedback to guide non-native language learners, which is a core component in computer-assisted pronunciation training (CAPT) systems. However, MDD often suffers from the data-sparsity problem due to that collecting non-native data and the associated annotations is time-consuming and labor-intensive. To address this issue, we explore a fully end-to-end (E2E) neural model for MDD, which processes learners' speech directly based on raw waveforms. Compared to conventional hand-crafted acoustic features, raw waveforms retain more acoustic phenomena and potentially can help neural networks discover better and more customized representations. To this end, our MDD model adopts a co-called SincNet module to take input a raw waveform and covert it to a suitable vector representation sequence. SincNet employs the cardinal sine (sinc) function to implement learnable bandpass filters, drawing inspiration from the convolutional neural network (CNN). By comparison to CNN, SincNet has fewer parameters and is more amenable to human interpretation. Extensive experiments are conducted on the L2-ARCTIC dataset, which is a publicly-available non-native English speech corpus compiled for research on CAPT. We find that the sinc filters of SincNet can be adapted quickly for non-native language learners of different nationalities. Furthermore, our model can achieve comparable mispronunciation detection performance in relation to state-of-the-art E2E MDD models that take input the standard handcrafted acoustic features. Besides that, our model also provides considerable improvements on phone error rate (PER) and diagnosis accuracy.

Yu-Sen Cheng, Chun-Liang Shih, Tien-Hong Lo et al.

In this report, we describe our submission to the VoxCeleb Speaker Recognition Challenge (VoxSRC) 2020. Two approaches are adopted. One is to apply query expansion on speaker verification, which shows significant progress compared to baseline in the study. Another is to use Kaldi extract x-vector and to combine its Probabilistic Linear Discriminant Analysis (PLDA) score with ResNet score.

Wen-Ting Tseng, Tien-Hong Lo, Yung-Chang Hsu et al.

Frequently asked question (FAQ) retrieval, with the purpose of providing information on frequent questions or concerns, has far-reaching applications in many areas, where a collection of question-answer (Q-A) pairs compiled a priori can be employed to retrieve an appropriate answer in response to a user\u2019s query that is likely to reoccur frequently. To this end, predominant approaches to FAQ retrieval typically rank question-answer pairs by considering either the similarity between the query and a question (q-Q), the relevance between the query and the associated answer of a question (q-A), or combining the clues gathered from the q-Q similarity measure and the q-A relevance measure. In this paper, we extend this line of research by combining the clues gathered from the q-Q similarity measure and the q-A relevance measure and meanwhile injecting extra word interaction information, distilled from a generic (open domain) knowledge base, into a contextual language model for inferring the q-A relevance. Furthermore, we also explore to capitalize on domain-specific topically-relevant relations between words in an unsupervised manner, acting as a surrogate to the supervised domain-specific knowledge base information. As such, it enables the model to equip sentence representations with the knowledge about domain-specific and topically-relevant relations among words, thereby providing a better q-A relevance measure. We evaluate variants of our approach on a publicly-available Chinese FAQ dataset, and further apply and contextualize it to a large-scale question-matching task, which aims to search questions from a QA dataset that have a similar intent as an input query. Extensive experimental results on these two datasets confirm the promising performance of the proposed approach in relation to some state-of-the-art ones.

Shi-Yan Weng, Tien-Hong Lo, Berlin Chen

Tremendous amounts of multimedia associated with speech information are driving an urgent need to develop efficient and effective automatic summarization methods. To this end, we have seen rapid progress in applying supervised deep neural network-based methods to extractive speech summarization. More recently, the Bidirectional Encoder Representations from Transformers (BERT) model was proposed and has achieved record-breaking success on many natural language processing (NLP) tasks such as question answering and language understanding. In view of this, we in this paper contextualize and enhance the state-of-the-art BERT-based model for speech summarization, while its contributions are at least three-fold. First, we explore the incorporation of confidence scores into sentence representations to see if such an attempt could help alleviate the negative effects caused by imperfect automatic speech recognition (ASR). Secondly, we also augment the sentence embeddings obtained from BERT with extra structural and linguistic features, such as sentence position and inverse document frequency (IDF) statistics. Finally, we validate the effectiveness of our proposed method on a benchmark dataset, in comparison to several classic and celebrated speech summarization methods.

Tien-Hong Lo, Shi-Yan Weng, Hsiu-Jui Chang et al.

Recently, end-to-end (E2E) automatic speech recognition (ASR) systems have garnered tremendous attention because of their great success and unified modeling paradigms in comparison to conventional hybrid DNN-HMM ASR systems. Despite the widespread adoption of E2E modeling frameworks on ASR, there still is a dearth of work on investigating the E2E frameworks for use in computer-assisted pronunciation learning (CAPT), particularly for Mispronunciation detection (MD). In response, we first present a novel use of hybrid CTCAttention approach to the MD task, taking advantage of the strengths of both CTC and the attention-based model meanwhile getting around the need for phone-level forced alignment. Second, we perform input augmentation with text prompt information to make the resulting E2E model more tailored for the MD task. On the other hand, we adopt two MD decision methods so as to better cooperate with the proposed framework: 1) decision-making based on a recognition confidence measure or 2) simply based on speech recognition results. A series of Mandarin MD experiments demonstrate that our approach not only simplifies the processing pipeline of existing hybrid DNN-HMM systems but also brings about systematic and substantial performance improvements. Furthermore, input augmentation with text prompts seems to hold excellent promise for the E2E-based MD approach.

Tien-Hong Lo, Fu-An Chao, Shi-Yan Weng et al.

This paper describes the NTNU ASR system participating in the Interspeech 2020 Non-Native Children's Speech ASR Challenge supported by the SIG-CHILD group of ISCA. This ASR shared task is made much more challenging due to the coexisting diversity of non-native and children speaking characteristics. In the setting of closed-track evaluation, all participants were restricted to develop their systems merely based on the speech and text corpora provided by the organizer. To work around this under-resourced issue, we built our ASR system on top of CNN-TDNNF-based acoustic models, meanwhile harnessing the synergistic power of various data augmentation strategies, including both utterance- and word-level speed perturbation and spectrogram augmentation, alongside a simple yet effective data-cleansing approach. All variants of our ASR system employed an RNN-based language model to rescore the first-pass recognition hypotheses, which was trained solely on the text dataset released by the organizer. Our system with the best configuration came out in second place, resulting in a word error rate (WER) of 17.59 %, while those of the top-performing, second runner-up and official baseline systems are 15.67%, 18.71%, 35.09%, respectively.

Ian Tenney, Patrick Xia, Berlin Chen et al.

Contextualized representation models such as ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2018) have recently achieved state-of-the-art results on a diverse array of downstream NLP tasks. Building on recent token-level probing work, we introduce a novel edge probing task design and construct a broad suite of sub-sentence tasks derived from the traditional structured NLP pipeline. We probe word-level contextual representations from four recent models and investigate how they encode sentence structure across a range of syntactic, semantic, local, and long-range phenomena. We find that existing models trained on language modeling and translation produce strong representations for syntactic phenomena, but only offer comparably small improvements on semantic tasks over a non-contextual baseline.

Alex Wang, Jan Hula, Patrick Xia et al.

Natural language understanding has recently seen a surge of progress with the use of sentence encoders like ELMo (Peters et al., 2018a) and BERT (Devlin et al., 2019) which are pretrained on variants of language modeling. We conduct the first large-scale systematic study of candidate pretraining tasks, comparing 19 different tasks both as alternatives and complements to language modeling. Our primary results support the use language modeling, especially when combined with pretraining on additional labeled-data tasks. However, our results are mixed across pretraining tasks and show some concerning trends: In ELMo's pretrain-then-freeze paradigm, random baselines are worryingly strong and results vary strikingly across target tasks. In addition, fine-tuning BERT on an intermediate task often negatively impacts downstream transfer. In a more positive trend, we see modest gains from multitask training, suggesting the development of more sophisticated multitask and transfer learning techniques as an avenue for further research.

Kuan-Yu Chen, Shih-Hung Liu, Berlin Chen et al.

Word embedding methods revolve around learning continuous distributed vector representations of words with neural networks, which can capture semantic and/or syntactic cues, and in turn be used to induce similarity measures among words, sentences and documents in context. Celebrated methods can be categorized as prediction-based and count-based methods according to the training objectives and model architectures. Their pros and cons have been extensively analyzed and evaluated in recent studies, but there is relatively less work continuing the line of research to develop an enhanced learning method that brings together the advantages of the two model families. In addition, the interpretation of the learned word representations still remains somewhat opaque. Motivated by the observations and considering the pressing need, this paper presents a novel method for learning the word representations, which not only inherits the advantages of classic word embedding methods but also offers a clearer and more rigorous interpretation of the learned word representations. Built upon the proposed word embedding method, we further formulate a translation-based language modeling framework for the extractive speech summarization task. A series of empirical evaluations demonstrate the effectiveness of the proposed word representation learning and language modeling techniques in extractive speech summarization.

Kuan-Yu Chen, Shih-Hung Liu, Berlin Chen et al.

Extractive summarization aims at selecting a set of indicative sentences from a source document as a summary that can express the major theme of the document. A general consensus on extractive summarization is that both relevance and coverage are critical issues to address. The existing methods designed to model coverage can be characterized by either reducing redundancy or increasing diversity in the summary. Maximal margin relevance (MMR) is a widely-cited method since it takes both relevance and redundancy into account when generating a summary for a given document. In addition to MMR, there is only a dearth of research concentrating on reducing redundancy or increasing diversity for the spoken document summarization task, as far as we are aware. Motivated by these observations, two major contributions are presented in this paper. First, in contrast to MMR, which considers coverage by reducing redundancy, we propose two novel coverage-based methods, which directly increase diversity. With the proposed methods, a set of representative sentences, which not only are relevant to the given document but also cover most of the important sub-themes of the document, can be selected automatically. Second, we make a step forward to plug in several document/sentence representation methods into the proposed framework to further enhance the summarization performance. A series of empirical evaluations demonstrate the effectiveness of our proposed methods.