Ngoc Thang Vu

Hung-yi Lee, Shang-Wen Li, Ngoc Thang Vu · meta-ai, mit

Deep learning has been the mainstream technique in natural language processing (NLP) area. However, the techniques require many labeled data and are less generalizable across domains. Meta-learning is an arising field in machine learning studying approaches to learn better learning algorithms. Approaches aim at improving algorithms in various aspects, including data efficiency and generalizability. Efficacy of approaches has been shown in many NLP tasks, but there is no systematic survey of these approaches in NLP, which hinders more researchers from joining the field. Our goal with this survey paper is to offer researchers pointers to relevant meta-learning works in NLP and attract more attention from the NLP community to drive future innovation. This paper first introduces the general concepts of meta-learning and the common approaches. Then we summarize task construction settings and application of meta-learning for various NLP problems and review the development of meta-learning in NLP community.

Florian Lux, Julia Koch, Ngoc Thang Vu

While neural methods for text-to-speech (TTS) have shown great advances in modeling multiple speakers, even in zero-shot settings, the amount of data needed for those approaches is generally not feasible for the vast majority of the world's over 6,000 spoken languages. In this work, we bring together the tasks of zero-shot voice cloning and multilingual low-resource TTS. Using the language agnostic meta learning (LAML) procedure and modifications to a TTS encoder, we show that it is possible for a system to learn speaking a new language using just 5 minutes of training data while retaining the ability to infer the voice of even unseen speakers in the newly learned language. We show the success of our proposed approach in terms of intelligibility, naturalness and similarity to target speaker using objective metrics as well as human studies and provide our code and trained models open source.

Florian Lux, Julia Koch, Sarina Meyer et al.

For our contribution to the Blizzard Challenge 2023, we improved on the system we submitted to the Blizzard Challenge 2021. Our approach entails a rule-based text-to-phoneme processing system that includes rule-based disambiguation of homographs in the French language. It then transforms the phonemes to spectrograms as intermediate representations using a fast and efficient non-autoregressive synthesis architecture based on Conformer and Glow. A GAN based neural vocoder that combines recent state-of-the-art approaches converts the spectrogram to the final wave. We carefully designed the data processing, training, and inference procedures for the challenge data. Our system identifier is G. Open source code and demo are available.

Manuel Mager, Rajat Bhatnagar, Graham Neubig et al.

Neural models have drastically advanced state of the art for machine translation (MT) between high-resource languages. Traditionally, these models rely on large amounts of training data, but many language pairs lack these resources. However, an important part of the languages in the world do not have this amount of data. Most languages from the Americas are among them, having a limited amount of parallel and monolingual data, if any. Here, we present an introduction to the interested reader to the basic challenges, concepts, and techniques that involve the creation of MT systems for these languages. Finally, we discuss the recent advances and findings and open questions, product of an increased interest of the NLP community in these languages.

Injy Hamed, Nizar Habash, Slim Abdennadher et al.

We present our work on collecting ArzEn-ST, a code-switched Egyptian Arabic - English Speech Translation Corpus. This corpus is an extension of the ArzEn speech corpus, which was collected through informal interviews with bilingual speakers. In this work, we collect translations in both directions, monolingual Egyptian Arabic and monolingual English, forming a three-way speech translation corpus. We make the translation guidelines and corpus publicly available. We also report results for baseline systems for machine translation and speech translation tasks. We believe this is a valuable resource that can motivate and facilitate further research studying the code-switching phenomenon from a linguistic perspective and can be used to train and evaluate NLP systems.

Sarina Meyer, Pascal Tilli, Pavel Denisov et al.

In order to protect the privacy of speech data, speaker anonymization aims for hiding the identity of a speaker by changing the voice in speech recordings. This typically comes with a privacy-utility trade-off between protection of individuals and usability of the data for downstream applications. One of the challenges in this context is to create non-existent voices that sound as natural as possible. In this work, we propose to tackle this issue by generating speaker embeddings using a generative adversarial network with Wasserstein distance as cost function. By incorporating these artificial embeddings into a speech-to-text-to-speech pipeline, we outperform previous approaches in terms of privacy and utility. According to standard objective metrics and human evaluation, our approach generates intelligible and content-preserving yet privacy-protecting versions of the original recordings.

Injy Hamed, Nizar Habash, Slim Abdennadher et al.

Data sparsity is a main problem hindering the development of code-switching (CS) NLP systems. In this paper, we investigate data augmentation techniques for synthesizing dialectal Arabic-English CS text. We perform lexical replacements using word-aligned parallel corpora where CS points are either randomly chosen or learnt using a sequence-to-sequence model. We compare these approaches against dictionary-based replacements. We assess the quality of the generated sentences through human evaluation and evaluate the effectiveness of data augmentation on machine translation (MT), automatic speech recognition (ASR), and speech translation (ST) tasks. Results show that using a predictive model results in more natural CS sentences compared to the random approach, as reported in human judgements. In the downstream tasks, despite the random approach generating more data, both approaches perform equally (outperforming dictionary-based replacements). Overall, data augmentation achieves 34% improvement in perplexity, 5.2% relative improvement on WER for ASR task, +4.0-5.1 BLEU points on MT task, and +2.1-2.2 BLEU points on ST over a baseline trained on available data without augmentation.

Florian Lux, Ngoc Thang Vu

While neural text-to-speech systems perform remarkably well in high-resource scenarios, they cannot be applied to the majority of the over 6,000 spoken languages in the world due to a lack of appropriate training data. In this work, we use embeddings derived from articulatory vectors rather than embeddings derived from phoneme identities to learn phoneme representations that hold across languages. In conjunction with language agnostic meta learning, this enables us to fine-tune a high-quality text-to-speech model on just 30 minutes of data in a previously unseen language spoken by a previously unseen speaker.

Sarina Meyer, Florian Lux, Pavel Denisov et al.

In this work, we propose a speaker anonymization pipeline that leverages high quality automatic speech recognition and synthesis systems to generate speech conditioned on phonetic transcriptions and anonymized speaker embeddings. Using phones as the intermediate representation ensures near complete elimination of speaker identity information from the input while preserving the original phonetic content as much as possible. Our experimental results on LibriSpeech and VCTK corpora reveal two key findings: 1) although automatic speech recognition produces imperfect transcriptions, our neural speech synthesis system can handle such errors, making our system feasible and robust, and 2) combining speaker embeddings from different resources is beneficial and their appropriate normalization is crucial. Overall, our final best system outperforms significantly the baselines provided in the Voice Privacy Challenge 2020 in terms of privacy robustness against a lazy-informed attacker while maintaining high intelligibility and naturalness of the anonymized speech.

Marwa Gaser, Manuel Mager, Injy Hamed et al.

Data sparsity is one of the main challenges posed by code-switching (CS), which is further exacerbated in the case of morphologically rich languages. For the task of machine translation (MT), morphological segmentation has proven successful in alleviating data sparsity in monolingual contexts; however, it has not been investigated for CS settings. In this paper, we study the effectiveness of different segmentation approaches on MT performance, covering morphology-based and frequency-based segmentation techniques. We experiment on MT from code-switched Arabic-English to English. We provide detailed analysis, examining a variety of conditions, such as data size and sentences with different degrees of CS. Empirical results show that morphology-aware segmenters perform the best in segmentation tasks but under-perform in MT. Nevertheless, we find that the choice of the segmentation setup to use for MT is highly dependent on the data size. For extreme low-resource scenarios, a combination of frequency and morphology-based segmentations is shown to perform the best. For more resourced settings, such a combination does not bring significant improvements over the use of frequency-based segmentation.

Daniel Ortega, Chia-Yu Li, Ngoc Thang Vu

This paper presents our latest investigation on modeling backchannel in conversations. Motivated by a proactive backchanneling theory, we aim at developing a system which acts as a proactive listener by inserting backchannels, such as continuers and assessment, to influence speakers. Our model takes into account not only lexical and acoustic cues, but also introduces the simple and novel idea of using listener embeddings to mimic different backchanneling behaviours. Our experimental results on the Switchboard benchmark dataset reveal that acoustic cues are more important than lexical cues in this task and their combination with listener embeddings works best on both, manual transcriptions and automatically generated transcriptions.

Florian Lux, Julia Koch, Ngoc Thang Vu

The cloning of a speaker's voice using an untranscribed reference sample is one of the great advances of modern neural text-to-speech (TTS) methods. Approaches for mimicking the prosody of a transcribed reference audio have also been proposed recently. In this work, we bring these two tasks together for the first time through utterance level normalization in conjunction with an utterance level speaker embedding. We further introduce a lightweight aligner for extracting fine-grained prosodic features, that can be finetuned on individual samples within seconds. We show that it is possible to clone the voice of a speaker as well as the prosody of a spoken reference independently without any degradation in quality and high similarity to both original voice and prosody, as our objective evaluation and human study show. All of our code and trained models are available, alongside static and interactive demos.

Manuel Mager, Arturo Oncevay, Elisabeth Mager et al.

Morphologically-rich polysynthetic languages present a challenge for NLP systems due to data sparsity, and a common strategy to handle this issue is to apply subword segmentation. We investigate a wide variety of supervised and unsupervised morphological segmentation methods for four polysynthetic languages: Nahuatl, Raramuri, Shipibo-Konibo, and Wixarika. Then, we compare the morphologically inspired segmentation methods against Byte-Pair Encodings (BPEs) as inputs for machine translation (MT) when translating to and from Spanish. We show that for all language pairs except for Nahuatl, an unsupervised morphological segmentation algorithm outperforms BPEs consistently and that, although supervised methods achieve better segmentation scores, they under-perform in MT challenges. Finally, we contribute two new morphological segmentation datasets for Raramuri and Shipibo-Konibo, and a parallel corpus for Raramuri--Spanish.

Julia Koch, Florian Lux, Nadja Schauffler et al.

Speech synthesis for poetry is challenging due to specific intonation patterns inherent to poetic speech. In this work, we propose an approach to synthesise poems with almost human like naturalness in order to enable literary scholars to systematically examine hypotheses on the interplay between text, spoken realisation, and the listener's perception of poems. To meet these special requirements for literary studies, we resynthesise poems by cloning prosodic values from a human reference recitation, and afterwards make use of fine-grained prosody control to manipulate the synthetic speech in a human-in-the-loop setting to alter the recitation w.r.t. specific phenomena. We find that finetuning our TTS model on poetry captures poetic intonation patterns to a large extent which is beneficial for prosody cloning and manipulation and verify the success of our approach both in an objective evaluation as well as in human studies.

Daniel Ortega, Sarina Meyer, Antje Schweitzer et al.

We present our latest findings on backchannel modeling novelly motivated by the canonical use of the minimal responses Yeah and Uh-huh in English and their correspondent tokens in German, and the effect of encoding the speaker-listener interaction. Backchanneling theories emphasize the active and continuous role of the listener in the course of the conversation, their effects on the speaker's subsequent talk, and the consequent dynamic speaker-listener interaction. Therefore, we propose a neural-based acoustic backchannel classifier on minimal responses by processing acoustic features from the speaker speech, capturing and imitating listeners' backchanneling behavior, and encoding speaker-listener interaction. Our experimental results on the Switchboard and GECO datasets reveal that in almost all tested scenarios the speaker or listener behavior embeddings help the model make more accurate backchannel predictions. More importantly, a proper interaction encoding strategy, i.e., combining the speaker and listener embeddings, leads to the best performance on both datasets in terms of F1-score.

Florian Lux, Pascal Tilli, Sarina Meyer et al.

Customizing voice and speaking style in a speech synthesis system with intuitive and fine-grained controls is challenging, given that little data with appropriate labels is available. Furthermore, editing an existing human's voice also comes with ethical concerns. In this paper, we propose a method to generate artificial speaker embeddings that cannot be linked to a real human while offering intuitive and fine-grained control over the voice and speaking style of the embeddings, without requiring any labels for speaker or style. The artificial and controllable embeddings can be fed to a speech synthesis system, conditioned on embeddings of real humans during training, without sacrificing privacy during inference.

Chia-Yu Li, Ngoc Thang Vu

We propose a novel method that combines CycleGAN and inter-domain losses for semi-supervised end-to-end automatic speech recognition. Inter-domain loss targets the extraction of an intermediate shared representation of speech and text inputs using a shared network. CycleGAN uses cycle-consistent loss and the identity mapping loss to preserve relevant characteristics of the input feature after converting from one domain to another. As such, both approaches are suitable to train end-to-end models on unpaired speech-text inputs. In this paper, we exploit the advantages from both inter-domain loss and CycleGAN to achieve better shared representation of unpaired speech and text inputs and thus improve the speech-to-text mapping. Our experimental results on the WSJ eval92 and Voxforge (non English) show 8~8.5% character error rate reduction over the baseline, and the results on LibriSpeech test_clean also show noticeable improvement.

Lucas Möller, Pascal Tilli, Ngoc Thang Vu et al.

Dual encoder architectures like Clip models map two types of inputs into a shared embedding space and predict similarities between them. Despite their wide application, it is, however, not understood how these models compare their two inputs. Common first-order feature-attribution methods explain importances of individual features and can, thus, only provide limited insights into dual encoders, whose predictions depend on interactions between features. In this paper, we first derive a second-order method enabling the attribution of predictions by any differentiable dual encoder onto feature-interactions between its inputs. Second, we apply our method to Clip models and show that they learn fine-grained correspondences between parts of captions and regions in images. They match objects across input modes and also account for mismatches. This intrinsic visual-linguistic grounding ability, however, varies heavily between object classes, exhibits pronounced out-of-domain effects and we can identify individual errors as well as systematic failure categories. Code is publicly available: https://github.com/lucasmllr/exCLIP

Florian Lux, Ching-Yi Chen, Ngoc Thang Vu

Embedding paralinguistic properties is a challenging task as there are only a few hours of training data available for domains such as emotional speech. One solution to this problem is to pretrain a general self-supervised speech representation model on large amounts of unlabeled speech. This pretrained model is then finetuned to a specific task. Paralinguistic properties however have notoriously high class variance, making the finetuning ineffective. In this work, we propose a two step approach to this. First we improve the embedding space, then we train an adapter to bridge the gap from the embedding space to a classification task. In order to improve the class invariance we use a combination of contrastive and non-contrastive losses to explicitly optimize for class invariant, yet discriminative features. Our approach consistently outperforms baselines that are finetuned end-to-end on multiple tasks and surpasses a benchmark on state-of-the-art emotion classification.

Injy Hamed, Alia El Bolock, Cornelia Herbert et al.

Code-switching (CS) is a common linguistic phenomenon exhibited by multilingual individuals, where they tend to alternate between languages within one single conversation. CS is a complex phenomenon that not only encompasses linguistic challenges, but also contains a great deal of complexity in terms of its dynamic behaviour across speakers. Given that the factors giving rise to CS vary from one country to the other, as well as from one person to the other, CS is found to be a speaker-dependant behaviour, where the frequency by which the foreign language is embedded differs across speakers. While several researchers have looked into predicting CS behaviour from a linguistic point of view, research is still lacking in the task of predicting user CS behaviour from sociological and psychological perspectives. We provide an empirical user study, where we investigate the correlations between users' CS levels and character traits. We conduct interviews with bilinguals and gather information on their profiles, including their demographics, personality traits, and traveling experiences. We then use machine learning (ML) to predict users' CS levels based on their profiles, where we identify the main influential factors in the modeling process. We experiment with both classification as well as regression tasks. Our results show that the CS behaviour is affected by the relation between speakers, travel experiences as well as Neuroticism and Extraversion personality traits.

Hendrik Schuff, Heike Adel, Peng Qi et al.

While much research focused on producing explanations, it is still unclear how the produced explanations' quality can be evaluated in a meaningful way. Today's predominant approach is to quantify explanations using proxy scores which compare explanations to (human-annotated) gold explanations. This approach assumes that explanations which reach higher proxy scores will also provide a greater benefit to human users. In this paper, we present problems of this approach. Concretely, we (i) formulate desired characteristics of explanation quality, (ii) describe how current evaluation practices violate them, and (iii) support our argumentation with initial evidence from a crowdsourcing case study in which we investigate the explanation quality of state-of-the-art explainable question answering systems. We find that proxy scores correlate poorly with human quality ratings and, additionally, become less expressive the more often they are used (i.e. following Goodhart's law). Finally, we propose guidelines to enable a meaningful evaluation of explanations to drive the development of systems that provide tangible benefits to human users.

Sarina Meyer, Florian Lux, Ngoc Thang Vu

In speaker anonymization, speech recordings are modified in a way that the identity of the speaker remains hidden. While this technology could help to protect the privacy of individuals around the globe, current research restricts this by focusing almost exclusively on English data. In this study, we extend a state-of-the-art anonymization system to nine languages by transforming language-dependent components to their multilingual counterparts. Experiments testing the robustness of the anonymized speech against privacy attacks and speech deterioration show an overall success of this system for all languages. The results suggest that speaker embeddings trained on English data can be applied across languages, and that the anonymization performance for a language is mainly affected by the quality of the speech synthesis component used for it.

Dirk Väth, Lindsey Vanderlyn, Ngoc Thang Vu

Conversational interfaces provide a flexible and easy way for users to seek information that may otherwise be difficult or inconvenient to obtain. However, existing interfaces generally fall into one of two categories: FAQs, where users must have a concrete question in order to retrieve a general answer, or dialogs, where users must follow a predefined path but may receive a personalized answer. In this paper, we introduce Conversational Tree Search (CTS) as a new task that bridges the gap between FAQ-style information retrieval and task-oriented dialog, allowing domain-experts to define dialog trees which can then be converted to an efficient dialog policy that learns only to ask the questions necessary to navigate a user to their goal. We collect a dataset for the travel reimbursement domain and demonstrate a baseline as well as a novel deep Reinforcement Learning architecture for this task. Our results show that the new architecture combines the positive aspects of both the FAQ and dialog system used in the baseline and achieves higher goal completion while skipping unnecessary questions.

Maximilian Kimmich, Andrea Bartezzaghi, Jasmina Bogojeska et al.

Neural approaches have become very popular in Question Answering (QA), however, they require a large amount of annotated data. In this work, we propose a novel approach that combines data augmentation via question-answer generation with Active Learning to improve performance in low-resource settings, where the target domains are diverse in terms of difficulty and similarity to the source domain. We also investigate Active Learning for question answering in different stages, overall reducing the annotation effort of humans. For this purpose, we consider target domains in realistic settings, with an extremely low amount of annotated samples but with many unlabeled documents, which we assume can be obtained with little effort. Additionally, we assume a sufficient amount of labeled data from the source domain being available. We perform extensive experiments to find the best setup for incorporating domain experts. Our findings show that our novel approach, where humans are incorporated in a data generation approach, boosts performance in the low-resource, domain-specific setting, allowing for low-labeling-effort question answering systems in new, specialized domains. They further demonstrate how human annotation affects the performance of QA depending on the stage it is performed.

Tushar Dhyani, Florian Lux, Michele Mancusi et al.

Traditional speech enhancement methods often oversimplify the task of restoration by focusing on a single type of distortion. Generative models that handle multiple distortions frequently struggle with phone reconstruction and high-frequency harmonics, leading to breathing and gasping artifacts that reduce the intelligibility of reconstructed speech. These models are also computationally demanding, and many solutions are restricted to producing outputs in the wide-band frequency range, which limits their suitability for professional applications. To address these challenges, we propose Hi-ResLDM, a novel generative model based on latent diffusion designed to remove multiple distortions and restore speech recordings to studio quality, sampled at 48kHz. We benchmark Hi-ResLDM against state-of-the-art methods that leverage GAN and Conditional Flow Matching (CFM) components, demonstrating superior performance in regenerating high-frequency-band details. Hi-ResLDM not only excels in non-instrusive metrics but is also consistently preferred in human evaluation and performs competitively on intrusive evaluations, making it ideal for high-resolution speech restoration.

Daniel Grießhaber, Maximilian Kimmich, Johannes Maucher et al.

Evolutionary prompt optimization has demonstrated effectiveness in refining prompts for LLMs. However, existing approaches lack robust operators and efficient evaluation mechanisms. In this work, we propose several key improvements to evolutionary prompt optimization that can partially generalize to prompt optimization in general: 1) decomposing evolution into distinct steps to enhance the evolution and its control, 2) introducing an LLM-based judge to verify the evolutions, 3) integrating human feedback to refine the evolutionary operator, and 4) developing more efficient evaluation strategies that maintain performance while reducing computational overhead. Our approach improves both optimization quality and efficiency. We release our code, enabling prompt optimization on new tasks and facilitating further research in this area.

Injy Hamed, Nizar Habash, Ngoc Thang Vu

Code-switching (CSW) text generation has been receiving increasing attention as a solution to address data scarcity. In light of this growing interest, we need more comprehensive studies comparing different augmentation approaches. In this work, we compare three popular approaches: lexical replacements, linguistic theories, and back-translation (BT), in the context of Egyptian Arabic-English CSW. We assess the effectiveness of the approaches on machine translation and the quality of augmentations through human evaluation. We show that BT and CSW predictive-based lexical replacement, being trained on CSW parallel data, perform best on both tasks. Linguistic theories and random lexical replacement prove to be effective in the lack of CSW parallel data, where both approaches achieve similar results.

Lindsey Vanderlyn, Dirk Väth, Ngoc Thang Vu

Mental models play an important role in whether user interaction with intelligent systems, such as dialog systems is successful or not. Adaptive dialog systems present the opportunity to align a dialog agent's behavior with heterogeneous user expectations. However, there has been little research into what mental models users form when interacting with a task-oriented dialog system, how these models affect users' interactions, or what role system adaptation can play in this process, making it challenging to avoid damage to human-AI partnership. In this work, we collect a new publicly available dataset for exploring user mental models about information seeking dialog systems. We demonstrate that users have a variety of conflicting mental models about such systems, the validity of which directly impacts the success of their interactions and perceived usability of system. Furthermore, we show that adapting a dialog agent's behavior to better align with users' mental models, even when done implicitly, can improve perceived usability, dialog efficiency, and success. To this end, we argue that implicit adaptation can be a valid strategy for task-oriented dialog systems, so long as developers first have a solid understanding of users' mental models.

Pavel Denisov, Ngoc Thang Vu

A number of methods have been proposed for End-to-End Spoken Language Understanding (E2E-SLU) using pretrained models, however their evaluation often lacks multilingual setup and tasks that require prediction of lexical fillers, such as slot filling. In this work, we propose a unified method that integrates multilingual pretrained speech and text models and performs E2E-SLU on six datasets in four languages in a generative manner, including the prediction of lexical fillers. We investigate how the proposed method can be improved by pretraining on widely available speech recognition data using several training objectives. Pretraining on 7000 hours of multilingual data allows us to outperform the state-of-the-art ultimately on two SLU datasets and partly on two more SLU datasets. Finally, we examine the cross-lingual capabilities of the proposed model and improve on the best known result on the PortMEDIA-Language dataset by almost half, achieving a Concept/Value Error Rate of 23.65%.

Yixuan Xiao, Florian Lux, Alejandro Pérez-González-de-Martos et al.

Since Text-to-Speech systems typically don't produce waveforms directly, recent spoof detection studies use resynthesized waveforms from vocoders and neural audio codecs to simulate an attacker. Unlike vocoders, which are specifically designed for speech synthesis, neural audio codecs were originally developed for compressing audio for storage and transmission. However, their ability to discretize speech also sparked interest in language-modeling-based speech synthesis. Owing to this dual functionality, codec resynthesized data may be labeled as either bonafide or spoof. So far, very little research has addressed this issue. In this study, we present a challenging extension of the ASVspoof 5 dataset constructed for this purpose. We examine how different labeling choices affect detection performance and provide insights into labeling strategies.

Dirk Väth, Lindsey Vanderlyn, Ngoc Thang Vu

Conversational Tree Search (Väth et al., 2023) is a recent approach to controllable dialog systems, where domain experts shape the behavior of a Reinforcement Learning agent through a dialog tree. The agent learns to efficiently navigate this tree, while adapting to information needs, e.g., domain familiarity, of different users. However, the need for additional training data hinders deployment in new domains. To address this, we explore approaches to generate this data directly from dialog trees. We improve the original approach, and show that agents trained on synthetic data can achieve comparable dialog success to models trained on human data, both when using a commercial Large Language Model for generation, or when using a smaller open-source model, running on a single GPU. We further demonstrate the scalability of our approach by collecting and testing on two new datasets: ONBOARD, a new domain helping foreign residents moving to a new city, and the medical domain DIAGNOSE, a subset of Wikipedia articles related to scalp and head symptoms. Finally, we perform human testing, where no statistically significant differences were found in either objective or subjective measures between models trained on human and generated data.

Yassine El Kheir, Arnab Das, Yixuan Xiao et al.

Speech deepfake detection is a well-established research field with different models, datasets, and training strategies. However, the lack of standardized implementations and evaluation protocols limits reproducibility, benchmarking, and comparison across studies. In this work, we present DeepFense, a comprehensive, open-source PyTorch toolkit integrating the latest architectures, loss functions, and augmentation pipelines, alongside over 100 recipes. Using DeepFense, we conducted a large-scale evaluation of more than 400 models. Our findings reveal that while carefully curated training data improves cross-domain generalization, the choice of pre-trained front-end feature extractor dominates overall performance variance. Crucially, we show severe biases in high-performing models regarding audio quality, speaker gender, and language. DeepFense is expected to facilitate real-world deployment with the necessary tools to address equitable training data selection and front-end fine-tuning.

Chia-Yu Li, Ngoc Thang Vu

Training a semi-supervised end-to-end speech recognition system using noisy student training has significantly improved performance. However, this approach requires a substantial amount of paired speech-text and unlabeled speech, which is costly for low-resource languages. Therefore, this paper considers a more extreme case of semi-supervised end-to-end automatic speech recognition where there are limited paired speech-text, unlabeled speech (less than five hours), and abundant external text. Firstly, we observe improved performance by training the model using our previous work on semi-supervised learning "CycleGAN and inter-domain losses" solely with external text. Secondly, we enhance "CycleGAN and inter-domain losses" by incorporating automatic hyperparameter tuning, calling it "enhanced CycleGAN inter-domain losses." Thirdly, we integrate it into the noisy student training approach pipeline for low-resource scenarios. Our experimental results, conducted on six non-English languages from Voxforge and Common Voice, show a 20% word error rate reduction compared to the baseline teacher model and a 10% word error rate reduction compared to the baseline best student model, highlighting the significant improvements achieved through our proposed method.

Pascal Tilli, Ngoc Thang Vu

The large success of deep learning based methods in Visual Question Answering (VQA) has concurrently increased the demand for explainable methods. Most methods in Explainable Artificial Intelligence (XAI) focus on generating post-hoc explanations rather than taking an intrinsic approach, the latter characterizing an interpretable model. In this work, we introduce an interpretable approach for graph-based VQA and demonstrate competitive performance on the GQA dataset. This approach bridges the gap between interpretability and performance. Our model is designed to intrinsically produce a subgraph during the question-answering process as its explanation, providing insight into the decision making. To evaluate the quality of these generated subgraphs, we compare them against established post-hoc explainability methods for graph neural networks, and perform a human evaluation. Moreover, we present quantitative metrics that correlate with the evaluations of human assessors, acting as automatic metrics for the generated explanatory subgraphs. Our implementation is available at https://github.com/DigitalPhonetics/Intrinsic-Subgraph-Generation-for-VQA.

Siddhant Arora, Siddharth Dalmia, Pavel Denisov et al.

As Automatic Speech Processing (ASR) systems are getting better, there is an increasing interest of using the ASR output to do downstream Natural Language Processing (NLP) tasks. However, there are few open source toolkits that can be used to generate reproducible results on different Spoken Language Understanding (SLU) benchmarks. Hence, there is a need to build an open source standard that can be used to have a faster start into SLU research. We present ESPnet-SLU, which is designed for quick development of spoken language understanding in a single framework. ESPnet-SLU is a project inside end-to-end speech processing toolkit, ESPnet, which is a widely used open-source standard for various speech processing tasks like ASR, Text to Speech (TTS) and Speech Translation (ST). We enhance the toolkit to provide implementations for various SLU benchmarks that enable researchers to seamlessly mix-and-match different ASR and NLU models. We also provide pretrained models with intensively tuned hyper-parameters that can match or even outperform the current state-of-the-art performances. The toolkit is publicly available at https://github.com/espnet/espnet.

Chia-Yu Li, Daniel Ortega, Dirk Väth et al.

We present ADVISER - an open-source, multi-domain dialog system toolkit that enables the development of multi-modal (incorporating speech, text and vision), socially-engaged (e.g. emotion recognition, engagement level prediction and backchanneling) conversational agents. The final Python-based implementation of our toolkit is flexible, easy to use, and easy to extend not only for technically experienced users, such as machine learning researchers, but also for less technically experienced users, such as linguists or cognitive scientists, thereby providing a flexible platform for collaborative research. Link to open-source code: https://github.com/DigitalPhonetics/adviser

Pavel Denisov, Ngoc Thang Vu

We present the IMS-Speech, a web based tool for German and English speech transcription aiming to facilitate research in various disciplines which require accesses to lexical information in spoken language materials. This tool is based on modern open source software stack, advanced speech recognition methods and public data resources and is freely available for academic researchers. The utilized models are built to be generic in order to provide transcriptions of competitive accuracy on a diverse set of tasks and conditions.

Maximilian Schmidt, Andrea Bartezzaghi, Ngoc Thang Vu

Although language models (LMs) have boosted the performance of Question Answering, they still need plenty of data. Data annotation, in contrast, is a time-consuming process. This especially applies to Question Answering, where possibly large documents have to be parsed and annotated with questions and their corresponding answers. Furthermore, Question Answering models often only work well for the domain they were trained on. Since annotation is costly, we argue that domain-agnostic knowledge from LMs, such as linguistic understanding, is sufficient to create a well-curated dataset. With this motivation, we show that using large language models can improve Question Answering performance on various datasets in the few-shot setting compared to state-of-the-art approaches. For this, we perform data generation leveraging the Prompting framework, suggesting that language models contain valuable task-agnostic knowledge that can be used beyond the common pre-training/fine-tuning scheme. As a result, we consistently outperform previous approaches on few-shot Question Answering.

Injy Hamed, Caroline Sabty, Slim Abdennadher et al.

Language in the Arab world presents a complex diglossic and multilingual setting, involving the use of Modern Standard Arabic, various dialects and sub-dialects, as well as multiple European languages. This diverse linguistic landscape has given rise to code-switching, both within Arabic varieties and between Arabic and foreign languages. The widespread occurrence of code-switching across the region makes it vital to address these linguistic needs when developing language technologies. In this paper, we provide a review of the current literature in the field of code-switched Arabic NLP, offering a broad perspective on ongoing efforts, challenges, research gaps, and recommendations for future research directions.

Wei Zhou, Heike Adel, Hendrik Schuff et al.

Attribution scores indicate the importance of different input parts and can, thus, explain model behaviour. Currently, prompt-based models are gaining popularity, i.a., due to their easier adaptability in low-resource settings. However, the quality of attribution scores extracted from prompt-based models has not been investigated yet. In this work, we address this topic by analyzing attribution scores extracted from prompt-based models w.r.t. plausibility and faithfulness and comparing them with attribution scores extracted from fine-tuned models and large language models. In contrast to previous work, we introduce training size as another dimension into the analysis. We find that using the prompting paradigm (with either encoder-based or decoder-based models) yields more plausible explanations than fine-tuning the models in low-resource settings and Shapley Value Sampling consistently outperforms attention and Integrated Gradients in terms of leading to more plausible and faithful explanations.

Pavel Denisov, Ngoc Thang Vu

Recent advancements in language modeling have led to the emergence of Large Language Models (LLMs) capable of various natural language processing tasks. Despite their success in text-based tasks, applying LLMs to the speech domain remains limited and challenging. This paper presents BLOOMZMMS, a novel model that integrates a multilingual LLM with a multilingual speech encoder, aiming to harness the capabilities of LLMs for speech recognition and beyond. Utilizing a multi-instructional training approach, we demonstrate the transferability of linguistic knowledge from the text to the speech modality. Our experiments, conducted on 1900 hours of transcribed data from 139 languages, establish that a multilingual speech representation can be effectively learned and aligned with a multilingual LLM. While this learned representation initially shows limitations in task generalization, we address this issue by generating synthetic targets in a multi-instructional style. Our zero-shot evaluation results confirm the robustness of our approach across multiple tasks, including speech translation and multilingual spoken language understanding, thereby opening new avenues for applying LLMs in the speech domain.

Injy Hamed, Ngoc Thang Vu, Nizar Habash

Code-switching, the act of alternating between languages, emerged as a prevalent global phenomenon that needs to be addressed for building user-friendly language technologies. A main bottleneck in this pursuit is data scarcity, motivating research in the direction of code-switched data augmentation. However, current literature lacks comprehensive studies that enable us to understand the relation between the quality of synthetic data and improvements on NLP tasks. We extend previous research conducted in this direction on machine translation (MT) with results on automatic speech recognition (ASR) and cascaded speech translation (ST) to test generalizability of findings. Our experiments involve a wide range of augmentation techniques, covering lexical replacements, linguistic theories, and back-translation. Based on the results of MT, ASR, and ST, we draw conclusions and insights regarding the efficacy of various augmentation techniques and the impact of quality on performance.

Pascal Tilli, Ngoc Thang Vu

Explainable artificial intelligence (XAI) aims to make machine learning models more transparent. While many approaches focus on generating explanations post-hoc, interpretable approaches, which generate the explanations intrinsically alongside the predictions, are relatively rare. In this work, we integrate different discrete subset sampling methods into a graph-based visual question answering system to compare their effectiveness in generating interpretable explanatory subgraphs intrinsically. We evaluate the methods on the GQA dataset and show that the integrated methods effectively mitigate the performance trade-off between interpretability and answer accuracy, while also achieving strong co-occurrences between answer and question tokens. Furthermore, we conduct a human evaluation to assess the interpretability of the generated subgraphs using a comparative setting with the extended Bradley-Terry model, showing that the answer and question token co-occurrence metrics strongly correlate with human preferences. Our source code is publicly available.

Paul Mayer, Florian Lux, Alejandro Pérez-González-de-Martos et al.

While generative methods have progressed rapidly in recent years, generating expressive prosody for an utterance remains a challenging task in text-to-speech synthesis. This is particularly true for systems that model prosody explicitly through parameters such as pitch, energy, and duration, which is commonly done for the sake of interpretability and controllability. In this work, we investigate the effectiveness of stochastic methods for this task, including Normalizing Flows, Conditional Flow Matching, and Rectified Flows. We compare these methods to a traditional deterministic baseline, as well as to real human realizations. Our extensive subjective and objective evaluations demonstrate that stochastic methods produce natural prosody on par with human speakers by capturing the variability inherent in human speech. Further, they open up additional controllability options by allowing the sampling temperature to be tuned.

Thomas Bott, Florian Lux, Ngoc Thang Vu

In recent years, prompting has quickly become one of the standard ways of steering the outputs of generative machine learning models, due to its intuitive use of natural language. In this work, we propose a system conditioned on embeddings derived from an emotionally rich text that serves as prompt. Thereby, a joint representation of speaker and prompt embeddings is integrated at several points within a transformer-based architecture. Our approach is trained on merged emotional speech and text datasets and varies prompts in each training iteration to increase the generalization capabilities of the model. Objective and subjective evaluation results demonstrate the ability of the conditioned synthesis system to accurately transfer the emotions present in a prompt to speech. At the same time, precise tractability of speaker identities as well as overall high speech quality and intelligibility are maintained.

Florian Lux, Sarina Meyer, Lyonel Behringer et al.

In this work, we take on the challenging task of building a single text-to-speech synthesis system that is capable of generating speech in over 7000 languages, many of which lack sufficient data for traditional TTS development. By leveraging a novel integration of massively multilingual pretraining and meta learning to approximate language representations, our approach enables zero-shot speech synthesis in languages without any available data. We validate our system's performance through objective measures and human evaluation across a diverse linguistic landscape. By releasing our code and models publicly, we aim to empower communities with limited linguistic resources and foster further innovation in the field of speech technology.

Manuel Mager, Elisabeth Mager, Katharina Kann et al.

In recent years machine translation has become very successful for high-resource language pairs. This has also sparked new interest in research on the automatic translation of low-resource languages, including Indigenous languages. However, the latter are deeply related to the ethnic and cultural groups that speak (or used to speak) them. The data collection, modeling and deploying machine translation systems thus result in new ethical questions that must be addressed. Motivated by this, we first survey the existing literature on ethical considerations for the documentation, translation, and general natural language processing for Indigenous languages. Afterward, we conduct and analyze an interview study to shed light on the positions of community leaders, teachers, and language activists regarding ethical concerns for the automatic translation of their languages. Our results show that the inclusion, at different degrees, of native speakers and community members is vital to performing better and more ethical research on Indigenous languages.

Alon Jacovi, Hendrik Schuff, Heike Adel et al.

Word-level saliency explanations ("heat maps over words") are often used to communicate feature-attribution in text-based models. Recent studies found that superficial factors such as word length can distort human interpretation of the communicated saliency scores. We conduct a user study to investigate how the marking of a word's neighboring words affect the explainee's perception of the word's importance in the context of a saliency explanation. We find that neighboring words have significant effects on the word's importance rating. Concretely, we identify that the influence changes based on neighboring direction (left vs. right) and a-priori linguistic and computational measures of phrases and collocations (vs. unrelated neighboring words). Our results question whether text-based saliency explanations should be continued to be communicated at word level, and inform future research on alternative saliency explanation methods.

Hendrik Schuff, Alon Jacovi, Heike Adel et al.

While a lot of research in explainable AI focuses on producing effective explanations, less work is devoted to the question of how people understand and interpret the explanation. In this work, we focus on this question through a study of saliency-based explanations over textual data. Feature-attribution explanations of text models aim to communicate which parts of the input text were more influential than others towards the model decision. Many current explanation methods, such as gradient-based or Shapley value-based methods, provide measures of importance which are well-understood mathematically. But how does a person receiving the explanation (the explainee) comprehend it? And does their understanding match what the explanation attempted to communicate? We empirically investigate the effect of various factors of the input, the feature-attribution explanation, and visualization procedure, on laypeople's interpretation of the explanation. We query crowdworkers for their interpretation on tasks in English and German, and fit a GAMM model to their responses considering the factors of interest. We find that people often mis-interpret the explanations: superficial and unrelated factors, such as word length, influence the explainees' importance assignment despite the explanation communicating importance directly. We then show that some of this distortion can be attenuated: we propose a method to adjust saliencies based on model estimates of over- and under-perception, and explore bar charts as an alternative to heatmap saliency visualization. We find that both approaches can attenuate the distorting effect of specific factors, leading to better-calibrated understanding of the explanation.

Chia-Yu Li, Ngoc Thang Vu

Code-Switching (CS) is a common linguistic phenomenon in multilingual communities that consists of switching between languages while speaking. This paper presents our investigations on end-to-end speech recognition for Mandarin-English CS speech. We analyse different CS specific issues such as the properties mismatches between languages in a CS language pair, the unpredictable nature of switching points, and the data scarcity problem. We exploit and improve the state-of-the-art end-to-end system by merging nonlinguistic symbols, by integrating language identification using hierarchical softmax, by modeling sub-word units, by artificially lowering the speaking rate, and by augmenting data using speed perturbed technique and several monolingual datasets to improve the final performance not only on CS speech but also on monolingual benchmarks in order to make the system more applicable on real life settings. Finally, we explore the effect of different language model integration methods on the performance of the proposed model. Our experimental results reveal that all the proposed techniques improve the recognition performance. The best combined system improves the baseline system by up to 35% relatively in terms of mixed error rate and delivers acceptable performance on monolingual benchmarks.