Jaejin Cho

Jaejin Cho, Raghavendra Pappagari, Piotr Żelasko et al. · nvidia

Considering the abundance of unlabeled speech data and the high labeling costs, unsupervised learning methods can be essential for better system development. One of the most successful methods is contrastive self-supervised methods, which require negative sampling: sampling alternative samples to contrast with the current sample (anchor). However, it is hard to ensure if all the negative samples belong to classes different from the anchor class without labels. This paper applies a non-contrastive self-supervised learning method on an unlabeled speech corpus to learn utterance-level embeddings. We used DIstillation with NO labels (DINO), proposed in computer vision, and adapted it to the speech domain. Unlike the contrastive methods, DINO does not require negative sampling. These embeddings were evaluated on speaker verification and emotion recognition. In speaker verification, the unsupervised DINO embedding with cosine scoring provided 4.38% EER on the VoxCeleb1 test trial. This outperforms the best contrastive self-supervised method by 40% relative in EER. An iterative pseudo-labeling training pipeline, not requiring speaker labels, further improved the EER to 1.89%. In emotion recognition, the DINO embedding performed 60.87, 79.21, and 56.98% in micro-f1 score on IEMOCAP, Crema-D, and MSP-Podcast, respectively. The results imply the generality of the DINO embedding to different speech applications.

Jaejin Cho, Jes'us Villalba, Laureano Moro-Velazquez et al.

In recent studies, self-supervised pre-trained models tend to outperform supervised pre-trained models in transfer learning. In particular, self-supervised learning (SSL) of utterance-level speech representation can be used in speech applications that require discriminative representation of consistent attributes within an utterance: speaker, language, emotion, and age. Existing frame-level self-supervised speech representation, e.g., wav2vec, can be used as utterance-level representation with pooling, but the models are usually large. There are also SSL techniques to learn utterance-level representation. One of the most successful is a contrastive method, which requires negative sampling: selecting alternative samples to contrast with the current sample (anchor). However, this does not ensure that all the negative samples belong to classes different from the anchor class without labels. This paper applies a non-contrastive self-supervised method to learn utterance-level embeddings. We adapted DIstillation with NO labels (DINO) from computer vision to speech. Unlike contrastive methods, DINO does not require negative sampling. We compared DINO to x-vector trained in a supervised manner. When transferred to down-stream tasks (speaker verification, speech emotion recognition (SER), and Alzheimer's disease detection), DINO outperformed x-vector. We studied the influence of several aspects during transfer learning such as dividing the fine-tuning process into steps, chunk lengths, or augmentation. During fine-tuning, tuning the last affine layers first and then the whole network surpassed fine-tuning all at once. Using shorter chunk lengths, although they generate more diverse inputs, did not necessarily improve performance, implying speech segments at least with a specific length are required for better performance per application. Augmentation was helpful in SER.

Rakshith Sharma Srinivasa, Jaejin Cho, Chouchang Yang et al.

This paper considers contrastive training for cross-modal 0-shot transfer wherein a pre-trained model in one modality is used for representation learning in another domain using pairwise data. The learnt models in the latter domain can then be used for a diverse set of tasks in a zero-shot way, similar to ``Contrastive Language-Image Pre-training (CLIP)'' and ``Locked-image Tuning (LiT)'' that have recently gained considerable attention. Most existing works for cross-modal representation alignment (including CLIP and LiT) use the standard contrastive training objective, which employs sets of positive and negative examples to align similar and repel dissimilar training data samples. However, similarity amongst training examples has a more continuous nature, thus calling for a more `non-binary' treatment. To address this, we propose a novel loss function called Continuously Weighted Contrastive Loss (CWCL) that employs a continuous measure of similarity. With CWCL, we seek to align the embedding space of one modality with another. Owing to the continuous nature of similarity in the proposed loss function, these models outperform existing methods for 0-shot transfer across multiple models, datasets and modalities. Particularly, we consider the modality pairs of image-text and speech-text and our models achieve 5-8% (absolute) improvement over previous state-of-the-art methods in 0-shot image classification and 20-30% (absolute) improvement in 0-shot speech-to-intent classification and keyword classification.

Jaejin Cho, Yohan Jun, Xiaoqing Wang et al.

Diffusion MRI is commonly performed using echo-planar imaging (EPI) due to its rapid acquisition time. However, the resolution of diffusion-weighted images is often limited by magnetic field inhomogeneity-related artifacts and blurring induced by T2- and T2*-relaxation effects. To address these limitations, multi-shot EPI (msEPI) combined with parallel imaging techniques is frequently employed. Nevertheless, reconstructing msEPI can be challenging due to phase variation between multiple shots. In this study, we introduce a novel msEPI reconstruction approach called zero-MIRID (zero-shot self-supervised learning of Multi-shot Image Reconstruction for Improved Diffusion MRI). This method jointly reconstructs msEPI data by incorporating deep learning-based image regularization techniques. The network incorporates CNN denoisers in both k- and image-spaces, while leveraging virtual coils to enhance image reconstruction conditioning. By employing a self-supervised learning technique and dividing sampled data into three groups, the proposed approach achieves superior results compared to the state-of-the-art parallel imaging method, as demonstrated in an in-vivo experiment.

Namgyu Han, Seong Dae Yun, Chaeeun Lim et al.

Echo-planar imaging (EPI) remains the cornerstone of diffusion MRI, but it is prone to severe geometric distortions due to its rapid sampling scheme that renders the sequence highly sensitive to $B_{0}$ field inhomogeneities. While deep learning has helped improve MRI reconstruction, integrating robust geometric distortion correction into a self-supervised framework remains an unmet need. To address this, we present FINDER (Field-Integrated Network for Distortion-free EPI Reconstruction), a novel zero-shot, scan-specific framework that reformulates reconstruction as a joint optimization of the underlying image and the $B_{0}$ field map. Specifically, we employ a physics-guided unrolled network that integrates dual-domain denoisers and virtual coil extensions to enforce robust data consistency. This is coupled with an Implicit Neural Representation (INR) conditioned on spatial coordinates and latent image features to model the off-resonance field as a continuous, differentiable function. Employing an alternating minimization strategy, FINDER synergistically updates the reconstruction network and the field map, effectively disentangling susceptibility-induced geometric distortions from anatomical structures. Experimental results demonstrate that FINDER achieves superior geometric fidelity and image quality compared to state-of-the-art baselines, offering a robust solution for high-quality diffusion imaging.

Ching-Hua Lee, Chouchang Yang, Jaejin Cho et al.

Denoising diffusion probabilistic models (DDPMs) can be utilized to recover a clean signal from its degraded observation(s) by conditioning the model on the degraded signal. The degraded signals are themselves contaminated versions of the clean signals; due to this correlation, they may encompass certain useful information about the target clean data distribution. However, existing adoption of the standard Gaussian as the prior distribution in turn discards such information when shaping the prior, resulting in sub-optimal performance. In this paper, we propose to improve conditional DDPMs for signal restoration by leveraging a more informative prior that is jointly learned with the diffusion model. The proposed framework, called RestoreGrad, seamlessly integrates DDPMs into the variational autoencoder (VAE) framework, taking advantage of the correlation between the degraded and clean signals to encode a better diffusion prior. On speech and image restoration tasks, we show that RestoreGrad demonstrates faster convergence (5-10 times fewer training steps) to achieve better quality of restored signals over existing DDPM baselines and improved robustness to using fewer sampling steps in inference time (2-2.5 times fewer), advocating the advantages of leveraging jointly learned prior for efficiency improvements in the diffusion process.

Xinrui Jiang, Yohan Jun, Jaejin Cho et al.

Typical quantitative MRI (qMRI) methods estimate parameter maps after image reconstructing, which is prone to biases and error propagation. We propose a Nonlinear Conjugate Gradient (NLCG) optimizer for model-based T2/T1 estimation, which incorporates U-Net regularization trained in a scan-specific manner. This end-to-end method directly estimates qMRI maps from undersampled k-space data using mono-exponential signal modeling with zero-shot scan-specific neural network regularization to enable high fidelity T1 and T2 mapping. T2 and T1 mapping results demonstrate the ability of the proposed NLCG-Net to improve estimation quality compared to subspace reconstruction at high accelerations.

Jaejin Cho, Borjan Gagoski, Taehyung Kim et al.

Purpose: To propose a wave-encoded model-based deep learning (wave-MoDL) strategy for highly accelerated 3D imaging and joint multi-contrast image reconstruction, and further extend this to enable rapid quantitative imaging using an interleaved look-locker acquisition sequence with T2 preparation pulse (3D-QALAS). Method: Recently introduced MoDL technique successfully incorporates convolutional neural network (CNN)-based regularizers into physics-based parallel imaging reconstruction using a small number of network parameters. Wave-CAIPI is an emerging parallel imaging method that accelerates the imaging speed by employing sinusoidal gradients in the phase- and slice-encoding directions during the readout to take better advantage of 3D coil sensitivity profiles. In wave-MoDL, we propose to combine the wave-encoding strategy with unrolled network constraints to accelerate the acquisition speed while enforcing wave-encoded data consistency. We further extend wave-MoDL to reconstruct multi-contrast data with controlled aliasing in parallel imaging (CAIPI) sampling patterns to leverage similarity between multiple images to improve the reconstruction quality. Result: Wave-MoDL enables a 47-second MPRAGE acquisition at 1 mm resolution at 16-fold acceleration. For quantitative imaging, wave-MoDL permits a 2-minute acquisition for T1, T2, and proton density mapping at 1 mm resolution at 12-fold acceleration, from which contrast weighted images can be synthesized as well. Conclusion: Wave-MoDL allows rapid MR acquisition and high-fidelity image reconstruction and may facilitate clinical and neuroscientific applications by incorporating unrolled neural networks into wave-CAIPI reconstruction.

Yamin Arefeen, Onur Beker, Jaejin Cho et al.

Purpose: To develop a scan-specific model that estimates and corrects k-space errors made when reconstructing accelerated Magnetic Resonance Imaging (MRI) data. Methods: Scan-Specific Artifact Reduction in k-space (SPARK) trains a convolutional-neural-network to estimate and correct k-space errors made by an input reconstruction technique by back-propagating from the mean-squared-error loss between an auto-calibration signal (ACS) and the input technique's reconstructed ACS. First, SPARK is applied to GRAPPA and demonstrates improved robustness over other scan-specific models, such as RAKI and residual-RAKI. Subsequent experiments demonstrate that SPARK synergizes with residual-RAKI to improve reconstruction performance. SPARK also improves reconstruction quality when applied to advanced acquisition and reconstruction techniques like 2D virtual coil (VC-) GRAPPA, 2D LORAKS, 3D GRAPPA without an integrated ACS region, and 2D/3D wave-encoded images. Results: SPARK yields 1.5x - 2x RMSE reduction when applied to GRAPPA and improves robustness to ACS size for various acceleration rates in comparison to other scan-specific techniques. When applied to advanced reconstruction techniques such as residual-RAKI, 2D VC-GRAPPA and LORAKS, SPARK achieves up to 20% RMSE improvement. SPARK with 3D GRAPPA also improves performance by ~2x and perceived image quality without a fully sampled ACS region. Finally, SPARK synergizes with non-cartesian 2D and 3D wave-encoding imaging by reducing RMSE between 20-25% and providing qualitative improvements. Conclusion: SPARK synergizes with physics-based acquisition and reconstruction techniques to improve accelerated MRI by training scan-specific models to estimate and correct reconstruction errors in k-space.

Jaejin Cho, Piotr Zelasko, Jesus Villalba et al.

Zero-shot multi-speaker Text-to-Speech (TTS) generates target speaker voices given an input text and the corresponding speaker embedding. In this work, we investigate the effectiveness of the TTS reconstruction objective to improve representation learning for speaker verification. We jointly trained end-to-end Tacotron 2 TTS and speaker embedding networks in a self-supervised fashion. We hypothesize that the embeddings will contain minimal phonetic information since the TTS decoder will obtain that information from the textual input. TTS reconstruction can also be combined with speaker classification to enhance these embeddings further. Once trained, the speaker encoder computes representations for the speaker verification task, while the rest of the TTS blocks are discarded. We investigated training TTS from either manual or ASR-generated transcripts. The latter allows us to train embeddings on datasets without manual transcripts. We compared ASR transcripts and Kaldi phone alignments as TTS inputs, showing that the latter performed better due to their finer resolution. Unsupervised TTS embeddings improved EER by 2.06\% absolute with regard to i-vectors for the LibriTTS dataset. TTS with speaker classification loss improved EER by 0.28\% and 0.73\% absolutely from a model using only speaker classification loss in LibriTTS and Voxceleb1 respectively.

Jaejin Cho, Shinji Watanabe, Takaaki Hori et al.

In this paper, we explore several new schemes to train a seq2seq model to integrate a pre-trained LM. Our proposed fusion methods focus on the memory cell state and the hidden state in the seq2seq decoder long short-term memory (LSTM), and the memory cell state is updated by the LM unlike the prior studies. This means the memory retained by the main seq2seq would be adjusted by the external LM. These fusion methods have several variants depending on the architecture of this memory cell update and the use of memory cell and hidden states which directly affects the final label inference. We performed the experiments to show the effectiveness of the proposed methods in a mono-lingual ASR setup on the Librispeech corpus and in a transfer learning setup from a multilingual ASR (MLASR) base model to a low-resourced language. In Librispeech, our best model improved WER by 3.7%, 2.4% for test clean, test other relatively to the shallow fusion baseline, with multi-level decoding. In transfer learning from an MLASR base model to the IARPA Babel Swahili model, the best scheme improved the transferred model on eval set by 9.9%, 9.8% in CER, WER relatively to the 2-stage transfer baseline.

Hirofumi Inaguma, Jaejin Cho, Murali Karthick Baskar et al.

This work explores better adaptation methods to low-resource languages using an external language model (LM) under the framework of transfer learning. We first build a language-independent ASR system in a unified sequence-to-sequence (S2S) architecture with a shared vocabulary among all languages. During adaptation, we perform LM fusion transfer, where an external LM is integrated into the decoder network of the attention-based S2S model in the whole adaptation stage, to effectively incorporate linguistic context of the target language. We also investigate various seed models for transfer learning. Experimental evaluations using the IARPA BABEL data set show that LM fusion transfer improves performances on all target five languages compared with simple transfer learning when the external text data is available. Our final system drastically reduces the performance gap from the hybrid systems.

Jaejin Cho, Murali Karthick Baskar, Ruizhi Li et al.

Sequence-to-sequence (seq2seq) approach for low-resource ASR is a relatively new direction in speech research. The approach benefits by performing model training without using lexicon and alignments. However, this poses a new problem of requiring more data compared to conventional DNN-HMM systems. In this work, we attempt to use data from 10 BABEL languages to build a multi-lingual seq2seq model as a prior model, and then port them towards 4 other BABEL languages using transfer learning approach. We also explore different architectures for improving the prior multilingual seq2seq model. The paper also discusses the effect of integrating a recurrent neural network language model (RNNLM) with a seq2seq model during decoding. Experimental results show that the transfer learning approach from the multilingual model shows substantial gains over monolingual models across all 4 BABEL languages. Incorporating an RNNLM also brings significant improvements in terms of %WER, and achieves recognition performance comparable to the models trained with twice more training data.

Takaaki Hori, Jaejin Cho, Shinji Watanabe

This paper investigates the impact of word-based RNN language models (RNN-LMs) on the performance of end-to-end automatic speech recognition (ASR). In our prior work, we have proposed a multi-level LM, in which character-based and word-based RNN-LMs are combined in hybrid CTC/attention-based ASR. Although this multi-level approach achieves significant error reduction in the Wall Street Journal (WSJ) task, two different LMs need to be trained and used for decoding, which increase the computational cost and memory usage. In this paper, we further propose a novel word-based RNN-LM, which allows us to decode with only the word-based LM, where it provides look-ahead word probabilities to predict next characters instead of the character-based LM, leading competitive accuracy with less computation compared to the multi-level LM. We demonstrate the efficacy of the word-based RNN-LMs using a larger corpus, LibriSpeech, in addition to WSJ we used in the prior work. Furthermore, we show that the proposed model achieves 5.1 %WER for WSJ Eval'92 test set when the vocabulary size is increased, which is the best WER reported for end-to-end ASR systems on this benchmark.