Kwanghee Choi

Xuankai Chang, Brian Yan, Kwanghee Choi et al. · cmu, meta-ai

Speech signals, typically sampled at rates in the tens of thousands per second, contain redundancies, evoking inefficiencies in sequence modeling. High-dimensional speech features such as spectrograms are often used as the input for the subsequent model. However, they can still be redundant. Recent investigations proposed the use of discrete speech units derived from self-supervised learning representations, which significantly compresses the size of speech data. Applying various methods, such as de-duplication and subword modeling, can further compress the speech sequence length. Hence, training time is significantly reduced while retaining notable performance. In this study, we undertake a comprehensive and systematic exploration into the application of discrete units within end-to-end speech processing models. Experiments on 12 automatic speech recognition, 3 speech translation, and 1 spoken language understanding corpora demonstrate that discrete units achieve reasonably good results in almost all the settings. We intend to release our configurations and trained models to foster future research efforts.

Masao Someki, Kwanghee Choi, Siddhant Arora et al. · cmu, nvidia

We introduce ESPnet-EZ, an extension of the open-source speech processing toolkit ESPnet, aimed at quick and easy development of speech models. ESPnet-EZ focuses on two major aspects: (i) easy fine-tuning and inference of existing ESPnet models on various tasks and (ii) easy integration with popular deep neural network frameworks such as PyTorch-Lightning, Hugging Face transformers and datasets, and Lhotse. By replacing ESPnet design choices inherited from Kaldi with a Python-only, Bash-free interface, we dramatically reduce the effort required to build, debug, and use a new model. For example, to fine-tune a speech foundation model, ESPnet-EZ, compared to ESPnet, reduces the number of newly written code by 2.7x and the amount of dependent code by 6.7x while dramatically reducing the Bash script dependencies. The codebase of ESPnet-EZ is publicly available.

Kwanghee Choi, Eun Jung Yeo · cmu

Self-supervised models, namely, wav2vec and its variants, have shown promising results in various downstream tasks in the speech domain. However, their inner workings are poorly understood, calling for in-depth analyses on what the model learns. In this paper, we concentrate on the convolutional feature encoder where its latent space is often speculated to represent discrete acoustic units. To analyze the embedding space in a reductive manner, we feed the synthesized audio signals, which is the summation of simple sine waves. Through extensive experiments, we conclude that various information is embedded inside the feature encoder representations: (1) fundamental frequency, (2) formants, and (3) amplitude, packed with (4) sufficient temporal detail. Further, the information incorporated inside the latent representations is analogous to spectrograms but with a fundamental difference: latent representations construct a metric space so that closer representations imply acoustic similarity.

Eun Jung Yeo, Kwanghee Choi, Sunhee Kim et al. · cmu

Automatic assessment of dysarthric speech is essential for sustained treatments and rehabilitation. However, obtaining atypical speech is challenging, often leading to data scarcity issues. To tackle the problem, we propose a novel automatic severity assessment method for dysarthric speech, using the self-supervised model in conjunction with multi-task learning. Wav2vec 2.0 XLS-R is jointly trained for two different tasks: severity classification and auxiliary automatic speech recognition (ASR). For the baseline experiments, we employ hand-crafted acoustic features and machine learning classifiers such as SVM, MLP, and XGBoost. Explored on the Korean dysarthric speech QoLT database, our model outperforms the traditional baseline methods, with a relative percentage increase of 1.25% for F1-score. In addition, the proposed model surpasses the model trained without ASR head, achieving 10.61% relative percentage improvements. Furthermore, we present how multi-task learning affects the severity classification performance by analyzing the latent representations and regularization effect.

Eun Jung Yeo, Kwanghee Choi, Sunhee Kim et al. · cmu

This paper proposes a cross-lingual classification method for English, Korean, and Tamil, which employs both language-independent features and language-unique features. First, we extract thirty-nine features from diverse speech dimensions such as voice quality, pronunciation, and prosody. Second, feature selections are applied to identify the optimal feature set for each language. A set of shared features and a set of distinctive features are distinguished by comparing the feature selection results of the three languages. Lastly, automatic severity classification is performed, utilizing the two feature sets. Notably, the proposed method removes different features by languages to prevent the negative effect of unique features for other languages. Accordingly, eXtreme Gradient Boosting (XGBoost) algorithm is employed for classification, due to its strength in imputing missing data. In order to validate the effectiveness of our proposed method, two baseline experiments are conducted: experiments using the intersection set of mono-lingual feature sets (Intersection) and experiments using the union set of mono-lingual feature sets (Union). According to the experimental results, our method achieves better performance with a 67.14% F1 score, compared to 64.52% for the Intersection experiment and 66.74% for the Union experiment. Further, the proposed method attains better performances than mono-lingual classifications for all three languages, achieving 17.67%, 2.28%, 7.79% relative percentage increases for English, Korean, and Tamil, respectively. The result specifies that commonly shared features and language-specific features must be considered separately for cross-language dysarthria severity classification.

Shikhar Bharadwaj, Chin-Jou Li, Yoonjae Kim et al.

Phone recognition (PR) serves as the atomic interface for language-agnostic modeling for cross-lingual speech processing and phonetic analysis. Despite prolonged efforts in developing PR systems, current evaluations only measure surface-level transcription accuracy. We introduce PRiSM, the first open-source benchmark designed to expose blind spots in phonetic perception through intrinsic and extrinsic evaluation of PR systems. PRiSM standardizes transcription-based evaluation and assesses downstream utility in clinical, educational, and multilingual settings with transcription and representation probes. We find that diverse language exposure during training is key to PR performance, encoder-CTC models are the most stable, and specialized PR models still outperform Large Audio Language Models. PRiSM releases code, recipes, and datasets to move the field toward multilingual speech models with robust phonetic ability: https://github.com/changelinglab/prism.

Kwanghee Choi, Hyung-Min Park

Multilingual speech data often suffer from long-tailed language distribution, resulting in performance degradation. However, multilingual text data is much easier to obtain, yielding a more useful general language model. Hence, we are motivated to distill the rich knowledge embedded inside a well-trained teacher text model to the student speech model. We propose a novel method called the Distilling a Language model to a Speech model (Distill-L2S), which aligns the latent representations of two different modalities. The subtle differences are handled by the shrinking mechanism, nearest-neighbor interpolation, and a learnable linear projection layer. We demonstrate the effectiveness of our distillation method by applying it to the multilingual automatic speech recognition (ASR) task. We distill the transformer-based cross-lingual language model (InfoXLM) while fine-tuning the large-scale multilingual ASR model (XLSR-wav2vec 2.0) for each language. We show the superiority of our method on 20 low-resource languages of the CommonVoice dataset with less than 100 hours of speech data.

Abner Hernandez, Eunjung Yeo, Kwanghee Choi et al.

The limited availability of dysarthric speech data makes cross-lingual detection an important but challenging problem. A key difficulty is that speech representations often encode language-dependent structure that can confound dysarthria detection. We propose a representation-level language shift (LS) that aligns source-language self-supervised speech representations with the target-language distribution using centroid-based vector adaptation estimated from healthy-control speech. We evaluate the approach on oral DDK recordings from Parkinson's disease speech datasets in Czech, German, and Spanish under both cross-lingual and multilingual settings. LS substantially improves sensitivity and F1 in cross-lingual settings, while yielding smaller but consistent gains in multilingual settings. Representation analysis further shows that LS reduces language identity in the embedding space, supporting the interpretation that LS removes language-dependent structure.

Seong Min Kye, Kwanghee Choi, Hyeongmin Byun et al.

Active learning (AL) aims to select the most useful data samples from an unlabeled data pool and annotate them to expand the labeled dataset under a limited budget. Especially, uncertainty-based methods choose the most uncertain samples, which are known to be effective in improving model performance. However, AL literature often overlooks training dynamics (TD), defined as the ever-changing model behavior during optimization via stochastic gradient descent, even though other areas of literature have empirically shown that TD provides important clues for measuring the sample uncertainty. In this paper, we propose a novel AL method, Training Dynamics for Active Learning (TiDAL), which leverages the TD to quantify uncertainties of unlabeled data. Since tracking the TD of all the large-scale unlabeled data is impractical, TiDAL utilizes an additional prediction module that learns the TD of labeled data. To further justify the design of TiDAL, we provide theoretical and empirical evidence to argue the usefulness of leveraging TD for AL. Experimental results show that our TiDAL achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to state-of-the-art AL methods, which estimate data uncertainty using only static information after model training.

Jeongkyun Park, Jung-Wook Hwang, Kwanghee Choi et al.

Inspired by humans comprehending speech in a multi-modal manner, various audio-visual datasets have been constructed. However, most existing datasets focus on English, induce dependencies with various prediction models during dataset preparation, and have only a small number of multi-view videos. To mitigate the limitations, we recently developed the Open Large-scale Korean Audio-Visual Speech (OLKAVS) dataset, which is the largest among publicly available audio-visual speech datasets. The dataset contains 1,150 hours of transcribed audio from 1,107 Korean speakers in a studio setup with nine different viewpoints and various noise situations. We also provide the pre-trained baseline models for two tasks, audio-visual speech recognition and lip reading. We conducted experiments based on the models to verify the effectiveness of multi-modal and multi-view training over uni-modal and frontal-view-only training. We expect the OLKAVS dataset to facilitate multi-modal research in broader areas such as Korean speech recognition, speaker recognition, pronunciation level classification, and mouth motion analysis.

Donghyun Son, Byounggyu Lew, Kwanghee Choi et al.

Social media platforms struggle to protect users from harmful content through content moderation. These platforms have recently leveraged machine learning models to cope with the vast amount of user-generated content daily. Since moderation policies vary depending on countries and types of products, it is common to train and deploy the models per policy. However, this approach is highly inefficient, especially when the policies change, requiring dataset re-labeling and model re-training on the shifted data distribution. To alleviate this cost inefficiency, social media platforms often employ third-party content moderation services that provide prediction scores of multiple subtasks, such as predicting the existence of underage personnel, rude gestures, or weapons, instead of directly providing final moderation decisions. However, making a reliable automated moderation decision from the prediction scores of the multiple subtasks for a specific target policy has not been widely explored yet. In this study, we formulate real-world scenarios of content moderation and introduce a simple yet effective threshold optimization method that searches the optimal thresholds of the multiple subtasks to make a reliable moderation decision in a cost-effective way. Extensive experiments demonstrate that our approach shows better performance in content moderation compared to existing threshold optimization methods and heuristics.

Yifan Peng, Jinchuan Tian, William Chen et al. · nvidia

Recent studies have highlighted the importance of fully open foundation models. The Open Whisper-style Speech Model (OWSM) is an initial step towards reproducing OpenAI Whisper using public data and open-source toolkits. However, previous versions of OWSM (v1 to v3) are still based on standard Transformer, which might lead to inferior performance compared to state-of-the-art speech encoder architectures. This work aims to improve the performance and efficiency of OWSM without additional data. We present a series of E-Branchformer-based models named OWSM v3.1, ranging from 100M to 1B parameters. OWSM v3.1 outperforms its predecessor, OWSM v3, in most evaluation benchmarks, while showing an improved inference speed of up to 25%. We further reveal the emergent ability of OWSM v3.1 in zero-shot contextual biasing speech recognition. We also provide a model trained on a subset of data with low license restrictions. We will publicly release the code, pre-trained models, and training logs.

Chien-yu Huang, Wei-Chih Chen, Shu-wen Yang et al. · cmu, mit

Multimodal foundation models, such as Gemini and ChatGPT, have revolutionized human-machine interactions by seamlessly integrating various forms of data. Developing a universal spoken language model that comprehends a wide range of natural language instructions is critical for bridging communication gaps and facilitating more intuitive interactions. However, the absence of a comprehensive evaluation benchmark poses a significant challenge. We present Dynamic-SUPERB Phase-2, an open and evolving benchmark for the comprehensive evaluation of instruction-based universal speech models. Building upon the first generation, this second version incorporates 125 new tasks contributed collaboratively by the global research community, expanding the benchmark to a total of 180 tasks, making it the largest benchmark for speech and audio evaluation. While the first generation of Dynamic-SUPERB was limited to classification tasks, Dynamic-SUPERB Phase-2 broadens its evaluation capabilities by introducing a wide array of novel and diverse tasks, including regression and sequence generation, across speech, music, and environmental audio. Evaluation results show that no model performed well universally. SALMONN-13B excelled in English ASR and Qwen2-Audio-7B-Instruct showed high accuracy in emotion recognition, but current models still require further innovations to handle a broader range of tasks. We open-source all task data and the evaluation pipeline at https://github.com/dynamic-superb/dynamic-superb.

Soojung Hong, Kwanghee Choi

As maintaining road networks is labor-intensive, many automatic road extraction approaches have been introduced to solve this real-world problem, fueled by the abundance of large-scale high-resolution satellite imagery and advances in computer vision. However, their performance is limited for fully automating the road map extraction in real-world services. Hence, many services employ the two-step human-in-the-loop system to post-process the extracted road maps: error localization and automatic mending for faulty road maps. Our paper exclusively focuses on the latter step, introducing a novel image inpainting approach for fixing road maps with complex road geometries without custom-made heuristics, yielding a method that is readily applicable to any road geometry extraction model. We demonstrate the effectiveness of our method on various real-world road geometries, such as straight and curvy roads, T-junctions, and intersections.

Jeongkyun Park, Kwanghee Choi, Hyunjun Heo et al.

With the advent of general-purpose speech representations from large-scale self-supervised models, applying a single model to multiple downstream tasks is becoming a de-facto approach. However, the pooling problem remains; the length of speech representations is inherently variable. The naive average pooling is often used, even though it ignores the characteristics of speech, such as differently lengthed phonemes. Hence, we design a novel pooling method to squash acoustically similar representations via vector quantization, which does not require additional training, unlike attention-based pooling. Further, we evaluate various unsupervised pooling methods on various self-supervised models. We gather diverse methods scattered around speech and text to evaluate on various tasks: keyword spotting, speaker identification, intent classification, and emotion recognition. Finally, we quantitatively and qualitatively analyze our method, comparing it with supervised pooling methods.

Brian Yan, Injy Hamed, Shuichiro Shimizu et al. · cmu

We present CS-FLEURS, a new dataset for developing and evaluating code-switched speech recognition and translation systems beyond high-resourced languages. CS-FLEURS consists of 4 test sets which cover in total 113 unique code-switched language pairs across 52 languages: 1) a 14 X-English language pair set with real voices reading synthetically generated code-switched sentences, 2) a 16 X-English language pair set with generative text-to-speech 3) a 60 {Arabic, Mandarin, Hindi, Spanish}-X language pair set with the generative text-to-speech, and 4) a 45 X-English lower-resourced language pair test set with concatenative text-to-speech. Besides the four test sets, CS-FLEURS also provides a training set with 128 hours of generative text-to-speech data across 16 X-English language pairs. Our hope is that CS-FLEURS helps to broaden the scope of future code-switched speech research. Dataset link: https://huggingface.co/datasets/byan/cs-fleurs.

Kwanghee Choi, Eunjung Yeo, Cheol Jun Cho et al.

Self-supervised speech models (S3Ms) are known to encode rich phonetic information, yet how this information is structured remains underexplored. We conduct a comprehensive study across 96 languages to analyze the underlying structure of S3M representations, with particular attention to phonological vectors. We first show that there exist linear directions within the model's representation space that correspond to phonological features. We further demonstrate that the scale of these phonological vectors correlate to the degree of acoustic realization of their corresponding phonological features in a continuous manner. For example, the difference between [d] and [t] yields a voicing vector: adding this vector to [p] produces [b], while scaling it results in a continuum of voicing. Together, these findings indicate that S3Ms encode speech using phonologically interpretable and compositional vectors, demonstrating phonological vector arithmetic. All code and interactive demos are available at https://github.com/juice500ml/phonetic-arithmetic .

Kwanghee Choi, Minyoung Choe, Hyelee Lee

Neural architecture search (NAS) has fostered various fields of machine learning. Despite its prominent dedications, many have criticized the intrinsic limitations of high computational cost. We aim to ameliorate this by proposing a pretraining scheme that can be generally applied to controller-based NAS. Our method, locality-based self-supervised classification task, leverages the structural similarity of network architectures to obtain good architecture representations. We incorporate our method into neural architecture optimization (NAO) to analyze the pretrained embeddings and its effectiveness and highlight that adding metric learning loss brings a favorable impact on NAS. Our code is available at \url{https://github.com/Multi-Objective-NAS/self-supervised-nas}.

Kwanghee Choi, Eunjung Yeo, Kalvin Chang et al. · cmu

Allophony refers to the variation in the phonetic realization of a phoneme based on its phonetic environment. Modeling allophones is crucial for atypical pronunciation assessment, which involves distinguishing atypical from typical pronunciations. However, recent phoneme classifier-based approaches often simplify this by treating various realizations as a single phoneme, bypassing the complexity of modeling allophonic variation. Motivated by the acoustic modeling capabilities of frozen self-supervised speech model (S3M) features, we propose MixGoP, a novel approach that leverages Gaussian mixture models to model phoneme distributions with multiple subclusters. Our experiments show that MixGoP achieves state-of-the-art performance across four out of five datasets, including dysarthric and non-native speech. Our analysis further suggests that S3M features capture allophonic variation more effectively than MFCCs and Mel spectrograms, highlighting the benefits of integrating MixGoP with S3M features.

Kwanghee Choi, Jee-weon Jung, Shinji Watanabe

With the success of self-supervised representations, researchers seek a better understanding of the information encapsulated within a representation. Among various interpretability methods, we focus on classification-based linear probing. We aim to foster a solid understanding and provide guidelines for linear probing by constructing a novel mathematical framework leveraging information theory. First, we connect probing with the variational bounds of mutual information (MI) to relax the probe design, equating linear probing with fine-tuning. Then, we investigate empirical behaviors and practices of probing through our mathematical framework. We analyze the layer-wise performance curve being convex, which seemingly violates the data processing inequality. However, we show that the intermediate representations can have the biggest MI estimate because of the tradeoff between better separability and decreasing MI. We further suggest that the margin of linearly separable representations can be a criterion for measuring the "goodness of representation." We also compare accuracy with MI as the measuring criteria. Finally, we empirically validate our claims by observing the self-supervised speech models on retaining word and phoneme information.

Tomohiko Nakamura, Kwanghee Choi, Keigo Hojo et al.

Self-supervised speech models (S3Ms) have become a common tool for the speech processing community, leveraging representations for downstream tasks. Clustering S3M representations yields discrete speech units (DSUs), which serve as compact representations for speech signals. DSUs are typically obtained by k-means clustering. Using DSUs often leads to strong performance in various tasks, including automatic speech recognition (ASR). However, even with the high dimensionality and redundancy of S3M representations, preprocessing S3M representations for better clustering remains unexplored, even though it can affect the quality of DSUs. In this paper, we investigate the potential of linear preprocessing methods for extracting DSUs. We evaluate standardization, principal component analysis, whitening, and independent component analysis (ICA) on DSU-based ASR benchmarks and demonstrate their effectiveness as preprocessing for k-means. We also conduct extensive analyses of their behavior, such as orthogonality or interpretability of individual components of ICA.

Kwanghee Choi, Masao Someki, Emma Strubell et al.

Discrete speech units (DSUs) are derived from clustering the features of self-supervised speech models (S3Ms). DSUs offer significant advantages for on-device streaming speech applications due to their rich phonetic information, high transmission efficiency, and seamless integration with large language models. However, conventional DSU-based approaches are impractical as they require full-length speech input and computationally expensive S3Ms. In this work, we reduce both the attention window and the model size while preserving the effectiveness of DSUs. Our results demonstrate that we can reduce floating-point operations (FLOPs) by 50% with only a relative increase of 6.5% in character error rate (CER) on the ML-SUPERB 1h dataset. These findings highlight the potential of DSUs for real-time speech processing in resource-constrained environments.

Kwanghee Choi, Eunjung Yeo, Cheol Jun Cho et al.

Transformer-based self-supervised speech models (S3Ms) are often described as contextualized, yet what this entails remains unclear. Here, we focus on how a single frame-level S3M representation can encode phones and their surrounding context. Prior work has shown that S3Ms represent phones compositionally; for example, phonological vectors such as voicing, bilabiality, and nasality vectors are superposed in the S3M representation of [m]. We extend this view by proposing that phonological information from a sequence of neighboring phones is also compositionally encoded in a single frame, such that vectors corresponding to previous, current, and next phones are superposed within a single frame-level representation. We show that this structure has several properties, including orthogonality between relative positions, and emergence of implicit phonetic boundaries. Together, our findings advance our understanding of context-dependent S3M representations.

Haitong Sun, Stephen McIntosh, Kwanghee Choi et al.

Speech representations from self-supervised speech models (S3Ms) are known to be sensitive to phonemic contrasts, but their sensitivity to prosodic contrasts has not been directly measured. The ABX discrimination task has been used to measure phonemic contrast in S3M representations via minimal pairs. We introduce prosodic ABX, an extension of this framework to evaluate prosodic contrast with only a handful of examples and no explicit labels. Also, we build and release a dataset of English and Japanese minimal pairs and use it along with a Mandarin dataset to evaluate contrast in English stress, Japanese pitch accent, and Mandarin tone. Finally, we show that model and layer rankings are often preserved across several experimental conditions, making it practical for low-resource settings.

Chin-Jou Li, Eunjung Yeo, Kwanghee Choi et al. · cmu

Automatic speech recognition (ASR) for dysarthric speech remains challenging due to data scarcity, particularly in non-English languages. To address this, we fine-tune a voice conversion model on English dysarthric speech (UASpeech) to encode both speaker characteristics and prosodic distortions, then apply it to convert healthy non-English speech (FLEURS) into non-English dysarthric-like speech. The generated data is then used to fine-tune a multilingual ASR model, Massively Multilingual Speech (MMS), for improved dysarthric speech recognition. Evaluation on PC-GITA (Spanish), EasyCall (Italian), and SSNCE (Tamil) demonstrates that VC with both speaker and prosody conversion significantly outperforms the off-the-shelf MMS performance and conventional augmentation techniques such as speed and tempo perturbation. Objective and subjective analyses of the generated data further confirm that the generated speech simulates dysarthric characteristics.

Shikhar Bharadwaj, Chin-Jou Li, Kwanghee Choi et al.

Phone recognition (PR) is a key enabler of multilingual and low-resource speech processing tasks, yet robust performance remains elusive. Highly performant English-focused models do not generalize across languages, while multilingual models underutilize pretrained representations. It also remains unclear how data scale, architecture, and training objective contribute to multilingual PR. We present PhoneticXEUS -- trained on large-scale multilingual data and achieving state-of-the-art performance on both multilingual (17.7% PFER) and accented English speech (10.6% PFER). Through controlled ablations with evaluations across 100+ languages under a unified scheme, we empirically establish our training recipe and quantify the impact of SSL representations, data scale, and loss objectives. In addition, we analyze error patterns across language families, accented speech, and articulatory features. All data and code are released openly.

Chin-Jou Li, Kalvin Chang, Shikhar Bharadwaj et al. · cmu

Recent advances in spoken language processing have led to substantial progress in phonetic tasks such as automatic speech recognition (ASR), phone recognition (PR), grapheme-to-phoneme conversion (G2P), and phoneme-to-grapheme conversion (P2G). Despite their conceptual similarity, these tasks have largely been studied in isolation, each relying on task-specific architectures and datasets. In this paper, we introduce POWSM (Phonetic Open Whisper-style Speech Model), the first unified framework capable of jointly performing multiple phone-related tasks. POWSM enables seamless conversion between audio, text (graphemes), and phones, opening up new possibilities for universal and low-resource speech processing. Our model outperforms or matches specialized PR models of similar size (Wav2Vec2Phoneme and ZIPA) while jointly supporting G2P, P2G, and ASR. Our training data, code and models are released to foster open science.

Jinchuan Tian, Yifan Peng, William Chen et al.

The Open Whisper-style Speech Model (OWSM) series was introduced to achieve full transparency in building advanced speech-to-text (S2T) foundation models. To this end, OWSM models are trained on 25 public speech datasets, which are heterogeneous in multiple ways. In this study, we advance the OWSM series by introducing OWSM v3.2, which improves on prior models by investigating and addressing the impacts of this data heterogeneity. Our study begins with a detailed analysis of each dataset, from which we derive two key strategies: data filtering with proxy task to enhance data quality, and the incorporation of punctuation and true-casing using an open large language model (LLM). With all other configurations staying the same, OWSM v3.2 improves performance over the OWSM v3.1 baseline while using 15% less training data.

Kwanghee Choi, Ankita Pasad, Tomohiko Nakamura et al.

Self-supervised speech models (S3Ms) have become an effective backbone for speech applications. Various analyses suggest that S3Ms encode linguistic properties. In this work, we seek a more fine-grained analysis of the word-level linguistic properties encoded in S3Ms. Specifically, we curate a novel dataset of near homophone (phonetically similar) and synonym (semantically similar) word pairs and measure the similarities between S3M word representation pairs. Our study reveals that S3M representations consistently and significantly exhibit more phonetic than semantic similarity. Further, we question whether widely used intent classification datasets such as Fluent Speech Commands and Snips Smartlights are adequate for measuring semantic abilities. Our simple baseline, using only the word identity, surpasses S3M-based models. This corroborates our findings and suggests that high scores on these datasets do not necessarily guarantee the presence of semantic content.

Taiqi He, Kwanghee Choi, Lindia Tjuatja et al.

Thousands of the world's languages are in danger of extinction--a tremendous threat to cultural identities and human language diversity. Interlinear Glossed Text (IGT) is a form of linguistic annotation that can support documentation and resource creation for these languages' communities. IGT typically consists of (1) transcriptions, (2) morphological segmentation, (3) glosses, and (4) free translations to a majority language. We propose Wav2Gloss: a task in which these four annotation components are extracted automatically from speech, and introduce the first dataset to this end, Fieldwork: a corpus of speech with all these annotations, derived from the work of field linguists, covering 37 languages, with standard formatting, and train/dev/test splits. We provide various baselines to lay the groundwork for future research on IGT generation from speech, such as end-to-end versus cascaded, monolingual versus multilingual, and single-task versus multi-task approaches.

Eun Jung Yeo, Kwanghee Choi, Sunhee Kim et al.

This paper proposes an improved Goodness of Pronunciation (GoP) that utilizes Uncertainty Quantification (UQ) for automatic speech intelligibility assessment for dysarthric speech. Current GoP methods rely heavily on neural network-driven overconfident predictions, which is unsuitable for assessing dysarthric speech due to its significant acoustic differences from healthy speech. To alleviate the problem, UQ techniques were used on GoP by 1) normalizing the phoneme prediction (entropy, margin, maxlogit, logit-margin) and 2) modifying the scoring function (scaling, prior normalization). As a result, prior-normalized maxlogit GoP achieves the best performance, with a relative increase of 5.66%, 3.91%, and 23.65% compared to the baseline GoP for English, Korean, and Tamil, respectively. Furthermore, phoneme analysis is conducted to identify which phoneme scores significantly correlate with intelligibility scores in each language.

Seong Min Kye, Kwanghee Choi, Joonyoung Yi et al.

Recent studies on learning with noisy labels have shown remarkable performance by exploiting a small clean dataset. In particular, model agnostic meta-learning-based label correction methods further improve performance by correcting noisy labels on the fly. However, there is no safeguard on the label miscorrection, resulting in unavoidable performance degradation. Moreover, every training step requires at least three back-propagations, significantly slowing down the training speed. To mitigate these issues, we propose a robust and efficient method that learns a label transition matrix on the fly. Employing the transition matrix makes the classifier skeptical about all the corrected samples, which alleviates the miscorrection issue. We also introduce a two-head architecture to efficiently estimate the label transition matrix every iteration within a single back-propagation, so that the estimated matrix closely follows the shifting noise distribution induced by label correction. Extensive experiments demonstrate that our approach shows the best performance in training efficiency while having comparable or better accuracy than existing methods.

Kwanghee Choi, Martin Kersner, Jacob Morton et al.

Improving the performance of on-device audio classification models remains a challenge given the computational limits of the mobile environment. Many studies leverage knowledge distillation to boost predictive performance by transferring the knowledge from large models to on-device models. However, most lack a mechanism to distill the essence of the temporal information, which is crucial to audio classification tasks, or similar architecture is often required. In this paper, we propose a new knowledge distillation method designed to incorporate the temporal knowledge embedded in attention weights of large transformer-based models into on-device models. Our distillation method is applicable to various types of architectures, including the non-attention-based architectures such as CNNs or RNNs, while retaining the original network architecture during inference. Through extensive experiments on both an audio event detection dataset and a noisy keyword spotting dataset, we show that our proposed method improves the predictive performance across diverse on-device architectures.

Youngkyu Hong, Seungju Han, Kwanghee Choi et al.

The current evaluation protocol of long-tailed visual recognition trains the classification model on the long-tailed source label distribution and evaluates its performance on the uniform target label distribution. Such protocol has questionable practicality since the target may also be long-tailed. Therefore, we formulate long-tailed visual recognition as a label shift problem where the target and source label distributions are different. One of the significant hurdles in dealing with the label shift problem is the entanglement between the source label distribution and the model prediction. In this paper, we focus on disentangling the source label distribution from the model prediction. We first introduce a simple but overlooked baseline method that matches the target label distribution by post-processing the model prediction trained by the cross-entropy loss and the Softmax function. Although this method surpasses state-of-the-art methods on benchmark datasets, it can be further improved by directly disentangling the source label distribution from the model prediction in the training phase. Thus, we propose a novel method, LAbel distribution DisEntangling (LADE) loss based on the optimal bound of Donsker-Varadhan representation. LADE achieves state-of-the-art performance on benchmark datasets such as CIFAR-100-LT, Places-LT, ImageNet-LT, and iNaturalist 2018. Moreover, LADE outperforms existing methods on various shifted target label distributions, showing the general adaptability of our proposed method.

Kwanghee Choi, Siyeong Lee

Notable progress has been made in numerous fields of machine learning based on neural network-driven mutual information (MI) bounds. However, utilizing the conventional MI-based losses is often challenging due to their practical and mathematical limitations. In this work, we first identify the symptoms behind their instability: (1) the neural network not converging even after the loss seemed to converge, and (2) saturating neural network outputs causing the loss to diverge. We mitigate both issues by adding a novel regularization term to the existing losses. We theoretically and experimentally demonstrate that added regularization stabilizes training. Finally, we present a novel benchmark that evaluates MI-based losses on both the MI estimation power and its capability on the downstream tasks, closely following the pre-existing supervised and contrastive learning settings. We evaluate six different MI-based losses and their regularized counterparts on multiple benchmarks to show that our approach is simple yet effective.