Hyeong-Seok Choi

Hyeong-Seok Choi, Juheon Lee, Wansoo Kim et al.

We present a neural analysis and synthesis (NANSY) framework that can manipulate voice, pitch, and speed of an arbitrary speech signal. Most of the previous works have focused on using information bottleneck to disentangle analysis features for controllable synthesis, which usually results in poor reconstruction quality. We address this issue by proposing a novel training strategy based on information perturbation. The idea is to perturb information in the original input signal (e.g., formant, pitch, and frequency response), thereby letting synthesis networks selectively take essential attributes to reconstruct the input signal. Because NANSY does not need any bottleneck structures, it enjoys both high reconstruction quality and controllability. Furthermore, NANSY does not require any labels associated with speech data such as text and speaker information, but rather uses a new set of analysis features, i.e., wav2vec feature and newly proposed pitch feature, Yingram, which allows for fully self-supervised training. Taking advantage of fully self-supervised training, NANSY can be easily extended to a multilingual setting by simply training it with a multilingual dataset. The experiments show that NANSY can achieve significant improvement in performance in several applications such as zero-shot voice conversion, pitch shift, and time-scale modification.

Sungho Lee, Hyeong-Seok Choi, Kyogu Lee

Artificial reverberation (AR) models play a central role in various audio applications. Therefore, estimating the AR model parameters (ARPs) of a reference reverberation is a crucial task. Although a few recent deep-learning-based approaches have shown promising performance, their non-end-to-end training scheme prevents them from fully exploiting the potential of deep neural networks. This motivates the introduction of differentiable artificial reverberation (DAR) models, allowing loss gradients to be back-propagated end-to-end. However, implementing the AR models with their difference equations "as is" in the deep learning framework severely bottlenecks the training speed when executed with a parallel processor like GPU due to their infinite impulse response (IIR) components. We tackle this problem by replacing the IIR filters with finite impulse response (FIR) approximations with the frequency-sampling method. Using this technique, we implement three DAR models -- differentiable Filtered Velvet Noise (FVN), Advanced Filtered Velvet Noise (AFVN), and Delay Network (DN). For each AR model, we train its ARP estimation networks for analysis-synthesis (RIR-to-ARP) and blind estimation (reverberant-speech-to-ARP) task in an end-to-end manner with its DAR model counterpart. Experiment results show that the proposed method achieves consistent performance improvement over the non-end-to-end approaches in both objective metrics and subjective listening test results.

Jaejun Lee, Donmoon Lee, Hyeong-Seok Choi et al.

Ensuring performance robustness for a variety of situations that can occur in real-world environments is one of the challenging tasks in sound event classification. One of the unpredictable and detrimental factors in performance, especially in indoor environments, is reverberation. To alleviate this problem, we propose a conditioning method that provides room impulse response (RIR) information to help the network become less sensitive to environmental information and focus on classifying the desired sound. Experimental results show that the proposed method successfully reduced performance degradation caused by the reverberation of the room. In particular, our proposed method works even with similar RIR that can be inferred from the room type rather than the exact one, which has the advantage of potentially being used in real-world applications.

Hyeong-Seok Choi, Sungjin Park, Jie Hwan Lee et al.

Modern deep learning-based models have seen outstanding performance improvement with speech enhancement tasks. The number of parameters of state-of-the-art models, however, is often too large to be deployed on devices for real-world applications. To this end, we propose Tiny Recurrent U-Net (TRU-Net), a lightweight online inference model that matches the performance of current state-of-the-art models. The size of the quantized version of TRU-Net is 362 kilobytes, which is small enough to be deployed on edge devices. In addition, we combine the small-sized model with a new masking method called phase-aware $β$-sigmoid mask, which enables simultaneous denoising and dereverberation. Results of both objective and subjective evaluations have shown that our model can achieve competitive performance with the current state-of-the-art models on benchmark datasets using fewer parameters by orders of magnitude.

Chang-Bin Jeon, Hyeong-Seok Choi, Kyogu Lee

In this paper, we propose a method of utilizing aligned lyrics as additional information to improve the performance of singing voice separation. We have combined the highway network-based lyrics encoder into Open-unmix separation network and show that the model trained with the aligned lyrics indeed results in a better performance than the model that was not informed. The question now remains whether the increase of performance is actually due to the phonetic contents that lie in the informed aligned lyrics or not. To this end, we investigated the source of performance increase in multifaceted ways by observing the change of performance when incorrect lyrics were given to the model. Experiment results show that the model can use not only just vocal activity information but also the phonetic contents from the aligned lyrics.

Hyeong-Seok Choi, Hoon Heo, Jie Hwan Lee et al.

In this work, we tackle a denoising and dereverberation problem with a single-stage framework. Although denoising and dereverberation may be considered two separate challenging tasks, and thus, two modules are typically required for each task, we show that a single deep network can be shared to solve the two problems. To this end, we propose a new masking method called phase-aware beta-sigmoid mask (PHM), which reuses the estimated magnitude values to estimate the clean phase by respecting the triangle inequality in the complex domain between three signal components such as mixture, source and the rest. Two PHMs are used to deal with direct and reverberant source, which allows controlling the proportion of reverberation in the enhanced speech at inference time. In addition, to improve the speech enhancement performance, we propose a new time-domain loss function and show a reasonable performance gain compared to MSE loss in the complex domain. Finally, to achieve a real-time inference, an optimization strategy for U-Net is proposed which significantly reduces the computational overhead up to 88.9% compared to the naïve version.

Hyeong-Seok Choi, Changdae Park, Kyogu Lee

This work seeks the possibility of generating the human face from voice solely based on the audio-visual data without any human-labeled annotations. To this end, we propose a multi-modal learning framework that links the inference stage and generation stage. First, the inference networks are trained to match the speaker identity between the two different modalities. Then the trained inference networks cooperate with the generation network by giving conditional information about the voice. The proposed method exploits the recent development of GANs techniques and generates the human face directly from the speech waveform making our system fully end-to-end. We analyze the extent to which the network can naturally disentangle two latent factors that contribute to the generation of a face image - one that comes directly from a speech signal and the other that is not related to it - and explore whether the network can learn to generate natural human face image distribution by modeling these factors. Experimental results show that the proposed network can not only match the relationship between the human face and speech, but can also generate the high-quality human face sample conditioned on its speech. Finally, the correlation between the generated face and the corresponding speech is quantitatively measured to analyze the relationship between the two modalities.

Juheon Lee, Hyeong-Seok Choi, Junghyun Koo et al.

In this study, we define the identity of the singer with two independent concepts - timbre and singing style - and propose a multi-singer singing synthesis system that can model them separately. To this end, we extend our single-singer model into a multi-singer model in the following ways: first, we design a singer identity encoder that can adequately reflect the identity of a singer. Second, we use encoded singer identity to condition the two independent decoders that model timbre and singing style, respectively. Through a user study with the listening tests, we experimentally verify that the proposed framework is capable of generating a natural singing voice of high quality while independently controlling the timbre and singing style. Also, by using the method of changing singing styles while fixing the timbre, we suggest that our proposed network can produce a more expressive singing voice.

Jie Hwan Lee, Hyeong-Seok Choi, Kyogu Lee

In recent years, music source separation has been one of the most intensively studied research areas in music information retrieval. Improvements in deep learning lead to a big progress in music source separation performance. However, most of the previous studies are restricted to separating a few limited number of sources, such as vocals, drums, bass, and other. In this study, we propose a network for audio query-based music source separation that can explicitly encode the source information from a query signal regardless of the number and/or kind of target signals. The proposed method consists of a Query-net and a Separator: given a query and a mixture, the Query-net encodes the query into the latent space, and the Separator estimates masks conditioned by the latent vector, which is then applied to the mixture for separation. The Separator can also generate masks using the latent vector from the training samples, allowing separation in the absence of a query. We evaluate our method on the MUSDB18 dataset, and experimental results show that the proposed method can separate multiple sources with a single network. In addition, through further investigation of the latent space we demonstrate that our method can generate continuous outputs via latent vector interpolation.

Juheon Lee, Hyeong-Seok Choi, Chang-Bin Jeon et al.

In this paper, we propose an end-to-end Korean singing voice synthesis system from lyrics and a symbolic melody using the following three novel approaches: 1) phonetic enhancement masking, 2) local conditioning of text and pitch to the super-resolution network, and 3) conditional adversarial training. The proposed system consists of two main modules; a mel-synthesis network that generates a mel-spectrogram from the given input information, and a super-resolution network that upsamples the generated mel-spectrogram into a linear-spectrogram. In the mel-synthesis network, phonetic enhancement masking is applied to generate implicit formant masks solely from the input text, which enables a more accurate phonetic control of singing voice. In addition, we show that two other proposed methods -- local conditioning of text and pitch, and conditional adversarial training -- are crucial for a realistic generation of the human singing voice in the super-resolution process. Finally, both quantitative and qualitative evaluations are conducted, confirming the validity of all proposed methods.

Hyeong-Seok Choi, Jang-Hyun Kim, Jaesung Huh et al.

Most deep learning-based models for speech enhancement have mainly focused on estimating the magnitude of spectrogram while reusing the phase from noisy speech for reconstruction. This is due to the difficulty of estimating the phase of clean speech. To improve speech enhancement performance, we tackle the phase estimation problem in three ways. First, we propose Deep Complex U-Net, an advanced U-Net structured model incorporating well-defined complex-valued building blocks to deal with complex-valued spectrograms. Second, we propose a polar coordinate-wise complex-valued masking method to reflect the distribution of complex ideal ratio masks. Third, we define a novel loss function, weighted source-to-distortion ratio (wSDR) loss, which is designed to directly correlate with a quantitative evaluation measure. Our model was evaluated on a mixture of the Voice Bank corpus and DEMAND database, which has been widely used by many deep learning models for speech enhancement. Ablation experiments were conducted on the mixed dataset showing that all three proposed approaches are empirically valid. Experimental results show that the proposed method achieves state-of-the-art performance in all metrics, outperforming previous approaches by a large margin.