Jaeyoung Kim

Hyunji Lee, Jaeyoung Kim, Hoyeon Chang et al. · deepmind

The generative retrieval model depends solely on the information encoded in its model parameters without external memory, its information capacity is limited and fixed. To overcome the limitation, we propose Nonparametric Decoding (Np Decoding) which can be applied to existing generative retrieval models. Np Decoding uses nonparametric contextualized vocab embeddings (external memory) rather than vanilla vocab embeddings as decoder vocab embeddings. By leveraging the contextualized vocab embeddings, the generative retrieval model is able to utilize both the parametric and nonparametric space. Evaluation over 9 datasets (8 single-hop and 1 multi-hop) in the document retrieval task shows that applying Np Decoding to generative retrieval models significantly improves the performance. We also show that Np Decoding is data- and parameter-efficient, and shows high performance in the zero-shot setting.

Soheil Khorram, Jaeyoung Kim, Anshuman Tripathi et al.

This paper introduces contrastive siamese (c-siam) network, an architecture for leveraging unlabeled acoustic data in speech recognition. c-siam is the first network that extracts high-level linguistic information from speech by matching outputs of two identical transformer encoders. It contains augmented and target branches which are trained by: (1) masking inputs and matching outputs with a contrastive loss, (2) incorporating a stop gradient operation on the target branch, (3) using an extra learnable transformation on the augmented branch, (4) introducing new temporal augment functions to prevent the shortcut learning problem. We use the Libri-light 60k unsupervised data and the LibriSpeech 100hrs/960hrs supervised data to compare c-siam and other best-performing systems. Our experiments show that c-siam provides 20% relative word error rate improvement over wav2vec baselines. A c-siam network with 450M parameters achieves competitive results compared to the state-of-the-art networks with 600M parameters.

Jaeyoung Kim, Dongbin Na, Sungchul Choi et al.

While pre-trained language models (PLMs) have become a de-facto standard promoting the accuracy of text classification tasks, recent studies find that PLMs often predict over-confidently. Although various calibration methods have been proposed, such as ensemble learning and data augmentation, most of the methods have been verified in computer vision benchmarks rather than in PLM-based text classification tasks. In this paper, we present an empirical study on confidence calibration for PLMs, addressing three categories, including confidence penalty losses, data augmentations, and ensemble methods. We find that the ensemble model overfitted to the training set shows sub-par calibration performance and also observe that PLMs trained with confidence penalty loss have a trade-off between calibration and accuracy. Building on these observations, we propose the Calibrated PLM (CALL), a combination of calibration techniques. The CALL complements the drawbacks that may occur when utilizing a calibration method individually and boosts both classification and calibration accuracy. Design choices in CALL's training procedures are extensively studied, and we provide a detailed analysis of how calibration techniques affect the calibration performance of PLMs.

Jaeyoung Kim, Kyuheon Jung, Dongbin Na et al.

For real-world language applications, detecting an out-of-distribution (OOD) sample is helpful to alert users or reject such unreliable samples. However, modern over-parameterized language models often produce overconfident predictions for both in-distribution (ID) and OOD samples. In particular, language models suffer from OOD samples with a similar semantic representation to ID samples since these OOD samples lie near the ID manifold. A rejection network can be trained with ID and diverse outlier samples to detect test OOD samples, but explicitly collecting auxiliary OOD datasets brings an additional burden for data collection. In this paper, we propose a simple but effective method called Pseudo Outlier Exposure (POE) that constructs a surrogate OOD dataset by sequentially masking tokens related to ID classes. The surrogate OOD sample introduced by POE shows a similar representation to ID data, which is most effective in training a rejection network. Our method does not require any external OOD data and can be easily implemented within off-the-shelf Transformers. A comprehensive comparison with state-of-the-art algorithms demonstrates POE's competitiveness on several text classification benchmarks.

Jaeyoung Kim, Seo Taek Kong, Dongbin Na et al.

Out-of-distribution (OOD) detection can be used in deep learning-based applications to reject outlier samples from being unreliably classified by deep neural networks. Learning to classify between OOD and in-distribution samples is difficult because data comprising the former is extremely diverse. It has been observed that an auxiliary OOD dataset is most effective in training a "rejection" network when its samples are semantically similar to in-distribution images. We first deduce that OOD images are perceived by a deep neural network to be semantically similar to in-distribution samples when they share a common background, as deep networks are observed to incorrectly classify such images with high confidence. We then propose a simple yet effective Key In-distribution feature Replacement BY inpainting (KIRBY) procedure that constructs a surrogate OOD dataset by replacing class-discriminative features of in-distribution samples with marginal background features. The procedure can be implemented using off-the-shelf vision algorithms, where each step within the algorithm is shown to make the surrogate data increasingly similar to in-distribution data. Design choices in each step are studied extensively, and an exhaustive comparison with state-of-the-art algorithms demonstrates KIRBY's competitiveness on various benchmarks.

Jaeyoung Kim, Han Lu, Soheil Khorram et al.

Modern automatic speech recognition (ASR) systems are typically trained on more than tens of thousands hours of speech data, which is one of the main factors for their great success. However, the distribution of such data is typically biased towards common accents or typical speech patterns. As a result, those systems often poorly perform on atypical accented speech. In this paper, we present accent clustering and mining schemes for fair speech recognition systems which can perform equally well on under-represented accented speech. For accent recognition, we applied three schemes to overcome limited size of supervised accent data: supervised or unsupervised pre-training, distributionally robust optimization (DRO) and unsupervised clustering. Three schemes can significantly improve the accent recognition model especially for unbalanced and small accented speech. Fine-tuning ASR on the mined Indian accent speech using the proposed supervised or unsupervised clustering schemes showed 10.0% and 5.3% relative improvements compared to fine-tuning on the randomly sampled speech, respectively.

Jaeyoung Kim, Han Lu, Anshuman Tripathi et al.

Reducing prediction delay for streaming end-to-end ASR models with minimal performance regression is a challenging problem. Constrained alignment is a well-known existing approach that penalizes predicted word boundaries using external low-latency acoustic models. On the contrary, recently proposed FastEmit is a sequence-level delay regularization scheme encouraging vocabulary tokens over blanks without any reference alignments. Although all these schemes are successful in reducing delay, ASR word error rate (WER) often severely degrades after applying these delay constraining schemes. In this paper, we propose a novel delay constraining method, named self alignment. Self alignment does not require external alignment models. Instead, it utilizes Viterbi forced-alignments from the trained model to find the lower latency alignment direction. From LibriSpeech evaluation, self alignment outperformed existing schemes: 25% and 56% less delay compared to FastEmit and constrained alignment at the similar word error rate. For Voice Search evaluation,12% and 25% delay reductions were achieved compared to FastEmit and constrained alignment with more than 2% WER improvements.

Anshuman Tripathi, Jaeyoung Kim, Qian Zhang et al.

In this paper we present a Transformer-Transducer model architecture and a training technique to unify streaming and non-streaming speech recognition models into one model. The model is composed of a stack of transformer layers for audio encoding with no lookahead or right context and an additional stack of transformer layers on top trained with variable right context. In inference time, the context length for the variable context layers can be changed to trade off the latency and the accuracy of the model. We also show that we can run this model in a Y-model architecture with the top layers running in parallel in low latency and high latency modes. This allows us to have streaming speech recognition results with limited latency and delayed speech recognition results with large improvements in accuracy (20% relative improvement for voice-search task). We show that with limited right context (1-2 seconds of audio) and small additional latency (50-100 milliseconds) at the end of decoding, we can achieve similar accuracy with models using unlimited audio right context. We also present optimizations for audio and label encoders to speed up the inference in streaming and non-streaming speech decoding.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Although supervised learning based on a deep neural network has recently achieved substantial improvement on speech enhancement, the existing schemes have either of two critical issues: spectrum or metric mismatches. The spectrum mismatch is a well known issue that any spectrum modification after short-time Fourier transform (STFT), in general, cannot be fully recovered after inverse short-time Fourier transform (ISTFT). The metric mismatch is that a conventional mean square error (MSE) loss function is typically sub-optimal to maximize perceptual speech measure such as signal-to-distortion ratio (SDR), perceptual evaluation of speech quality (PESQ) and short-time objective intelligibility (STOI). This paper presents a new end-to-end denoising framework. First, the network optimization is performed on the time-domain signals after ISTFT to avoid the spectrum mismatch. Second, three loss functions based on SDR, PESQ and STOI are proposed to minimize the metric mismatch. The experimental result showed the proposed denoising scheme significantly improved SDR, PESQ and STOI performance over the existing methods. Moreover, the proposed scheme also provided good generalization performance over generative denoising models on the perceptual speech metrics not used as a loss function during training.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Transformer neural networks (TNN) demonstrated state-of-art performance on many natural language processing (NLP) tasks, replacing recurrent neural networks (RNNs), such as LSTMs or GRUs. However, TNNs did not perform well in speech enhancement, whose contextual nature is different than NLP tasks, like machine translation. Self-attention is a core building block of the Transformer, which not only enables parallelization of sequence computation, but also provides the constant path length between symbols that is essential to learning long-range dependencies. In this paper, we propose a Transformer with Gaussian-weighted self-attention (T-GSA), whose attention weights are attenuated according to the distance between target and context symbols. The experimental results show that the proposed T-GSA has significantly improved speech-enhancement performance, compared to the Transformer and RNNs.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Supervised learning based on a deep neural network recently has achieved substantial improvement on speech enhancement. Denoising networks learn mapping from noisy speech to clean one directly, or to a spectrum mask which is the ratio between clean and noisy spectra. In either case, the network is optimized by minimizing mean square error (MSE) between ground-truth labels and time-domain or spectrum output. However, existing schemes have either of two critical issues: spectrum and metric mismatches. The spectrum mismatch is a well known issue that any spectrum modification after short-time Fourier transform (STFT), in general, cannot be fully recovered after inverse short-time Fourier transform (ISTFT). The metric mismatch is that a conventional MSE metric is sub-optimal to maximize our target metrics, signal-to-distortion ratio (SDR) and perceptual evaluation of speech quality (PESQ). This paper presents a new end-to-end denoising framework with the goal of joint SDR and PESQ optimization. First, the network optimization is performed on the time-domain signals after ISTFT to avoid spectrum mismatch. Second, two loss functions which have improved correlations with SDR and PESQ metrics are proposed to minimize metric mismatch. The experimental result showed that the proposed denoising scheme significantly improved both SDR and PESQ performance over the existing methods.

Jaeyoung Kim, Sion Jang, Sungchul Choi et al.

This paper presents an empirical exploration of the use of capsule networks for text classification. While it has been shown that capsule networks are effective for image classification, their validity in the domain of text has not been explored. In this paper, we show that capsule networks indeed have the potential for text classification and that they have several advantages over convolutional neural networks. We further suggest a simple routing method that effectively reduces the computational complexity of dynamic routing. We utilized seven benchmark datasets to demonstrate that capsule networks, along with the proposed routing method provide comparable results.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Despite the remarkable progress achieved on automatic speech recognition, recognizing far-field speeches mixed with various noise sources is still a challenging task. In this paper, we introduce novel student-teacher transfer learning, BridgeNet which can provide a solution to improve distant speech recognition. There are two key features in BridgeNet. First, BridgeNet extends traditional student-teacher frameworks by providing multiple hints from a teacher network. Hints are not limited to the soft labels from a teacher network. Teacher's intermediate feature representations can better guide a student network to learn how to denoise or dereverberate noisy input. Second, the proposed recursive architecture in the BridgeNet can iteratively improve denoising and recognition performance. The experimental results of BridgeNet showed significant improvements in tackling the distant speech recognition problem, where it achieved up to 13.24% relative WER reductions on AMI corpus compared to a baseline neural network without teacher's hints.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

In this paper, a novel architecture for a deep recurrent neural network, residual LSTM is introduced. A plain LSTM has an internal memory cell that can learn long term dependencies of sequential data. It also provides a temporal shortcut path to avoid vanishing or exploding gradients in the temporal domain. The residual LSTM provides an additional spatial shortcut path from lower layers for efficient training of deep networks with multiple LSTM layers. Compared with the previous work, highway LSTM, residual LSTM separates a spatial shortcut path with temporal one by using output layers, which can help to avoid a conflict between spatial and temporal-domain gradient flows. Furthermore, residual LSTM reuses the output projection matrix and the output gate of LSTM to control the spatial information flow instead of additional gate networks, which effectively reduces more than 10% of network parameters. An experiment for distant speech recognition on the AMI SDM corpus shows that 10-layer plain and highway LSTM networks presented 13.7% and 6.2% increase in WER over 3-layer aselines, respectively. On the contrary, 10-layer residual LSTM networks provided the lowest WER 41.0%, which corresponds to 3.3% and 2.8% WER reduction over plain and highway LSTM networks, respectively.