Singing Voice Synthesis System Research Articles

HiddenSinger: High-quality singing voice synthesis via neural audio codec and latent diffusion models

Recently, denoising diffusion models have demonstrated remarkable performance among generative models in various domains. However, in the speech domain, there are limitations in complexity and controllability to apply diffusion models for time-varying audio synthesis. Particularly, a singing voice synthesis (SVS) task, which has begun to emerge as a practical application in the game and entertainment industries, requires high-dimensional samples with long-term acoustic features. To alleviate the challenges posed by model complexity in the SVS task, we propose HiddenSinger, a high-quality SVS system using a neural audio codec and latent diffusion models. To ensure high-fidelity audio, we introduce an audio autoencoder that can encode audio into an audio codec as a compressed representation and reconstruct the high-fidelity audio from the low-dimensional compressed latent vector. Subsequently, we use the latent diffusion models to sample a latent representation from a musical score. In addition, our proposed model is extended to an unsupervised singing voice learning framework, HiddenSinger-U, to train the model using an unlabeled singing voice dataset. Experimental results demonstrate that our model outperforms previous models regarding audio quality. Furthermore, the HiddenSinger-U can synthesize high-quality singing voices of speakers trained solely on unlabeled data.

SinTechSVS: A Singing Technique Controllable Singing Voice Synthesis System

SinTechSVS: A Singing Technique Controllable Singing Voice Synthesis System

DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism

Singing voice synthesis (SVS) systems are built to synthesize high-quality and expressive singing voice, in which the acoustic model generates the acoustic features (e.g., mel-spectrogram) given a music score. Previous singing acoustic models adopt a simple loss (e.g., L1 and L2) or generative adversarial network (GAN) to reconstruct the acoustic features, while they suffer from over-smoothing and unstable training issues respectively, which hinder the naturalness of synthesized singing. In this work, we propose DiffSinger, an acoustic model for SVS based on the diffusion probabilistic model. DiffSinger is a parameterized Markov chain that iteratively converts the noise into mel-spectrogram conditioned on the music score. By implicitly optimizing variational bound, DiffSinger can be stably trained and generate realistic outputs. To further improve the voice quality and speed up inference, we introduce a shallow diffusion mechanism to make better use of the prior knowledge learned by the simple loss. Specifically, DiffSinger starts generation at a shallow step smaller than the total number of diffusion steps, according to the intersection of the diffusion trajectories of the ground-truth mel-spectrogram and the one predicted by a simple mel-spectrogram decoder. Besides, we propose boundary prediction methods to locate the intersection and determine the shallow step adaptively. The evaluations conducted on a Chinese singing dataset demonstrate that DiffSinger outperforms state-of-the-art SVS work. Extensional experiments also prove the generalization of our methods on text-to-speech task (DiffSpeech). Audio samples: https://diffsinger.github.io. Codes: https://github.com/MoonInTheRiver/DiffSinger.

Sinsy: A Deep Neural Network-Based Singing Voice Synthesis System

This paper presents Sinsy, a deep neural network (DNN)-based singing voice synthesis (SVS) system. In recent years, DNNs have been utilized in statistical parametric SVS systems, and DNN-based SVS systems have demonstrated better performance than conventional hidden Markov model-based ones. SVS systems are required to synthesize a singing voice with pitch and timing that strictly follow a given musical score. Additionally, singing expressions that are not described on the musical score, such as vibrato and timing fluctuations, should be reproduced. The proposed system is composed of four modules: a time-lag model, a duration model, an acoustic model, and a vocoder, and singing voices can be synthesized taking these characteristics of singing voices into account. To better model a singing voice, the proposed system incorporates improved approaches to modeling pitch and vibrato and better training criteria into the acoustic model. In addition, we incorporated PeriodNet, a non-autoregressive neural vocoder with robustness for the pitch, into our systems to generate a high-fidelity singing voice waveform. Moreover, we propose automatic pitch correction techniques for DNN-based SVS to synthesize singing voices with correct pitch even if the training data has out-of-tune phrases. Experimental results show our system can synthesize a singing voice with better timing, more natural vibrato, and correct pitch, and it can achieve better mean opinion scores in subjective evaluation tests.

PeriodNet: A Non-Autoregressive Raw Waveform Generative Model With a Structure Separating Periodic and Aperiodic Components

This paper presents PeriodNet, a non-autoregressive (non-AR) waveform generative model with a new model structure for modeling periodic and aperiodic components in speech waveforms. Non-AR raw waveform generative models have enabled the fast generation of high-quality waveforms. However, the variations of waveforms that these models can reconstruct are limited by training data. In addition, typical non-AR models reconstruct a speech waveform from a single Gaussian input despite the mixture of periodic and aperiodic signals in speech. These may significantly affect the waveform generation process in some applications such as singing voice synthesis systems, which require reproducing accurate pitch and natural sounds with less periodicity, including husky and breath sounds. PeriodNet uses a parallel or series model structure to model a speech waveform to tackle these problems. Two sub-generators connected in parallel or in series take an explicit periodic and aperiodic signal (sine wave and Gaussian noise) as an input. Since PeriodNet models periodic and aperiodic components by focusing on whether these input signals are autocorrelated or not, it does not require external periodic/aperiodic decomposition during training. Experimental results show that our proposed structure improves the naturalness of generated waveforms. We also show that speech waveforms with a pitch outside of the training data range can be generated with more naturalness.

Spectral and pitch modeling with hybrid approach to singing voice synthesis using hidden semi-Markov model and deep neural network

We propose a corpus-based singing voice synthesis system combining the hidden Markov model (HMM) and the Deep Neural Network (DNN). Recently, in the area of text-to-speech synthesis, it was reported that the DNN-based speech synthesis method showed better speech quality than the HMM-based one. However, when we introduced the DNN to statistical singing voice synthesis, it did not improve the synthetic singing voice quality. Thus, we introduced the DNN in the singing voice synthesis in a different way. Instead of modeling the speech spectra, we exploited the DNN to model the difference between the spectra of natural singing voice and synthetic singing voice from the HMM. To do that, we used the DNN to map the input musical information such as lyrics, tones, durations into the difference of output acoustic features between the natural and synthetic singing voice. This allows us to reconstruct the spectral fine structures in singing voice generated by HMMs. Our results proved that the proposed method improved the quality of synthetic singing voice compared to the conventional methods.

Vowel duration dependent hidden Markov model for automatic lyrics recognition

Recently, due to the spread of music distribution service, a large amount of music is available on the Internet. Accordingly, it is generally increasing the demand of music information retrieval (MIR). In the field of MIR research, there are several researches to extract meaningful information from music audio signals. However, automatic lyrics recognition is still a challenging problem because the variation of singing voice is much larger than that of speaking voice and a large database of singing voice is not available. In the relevant study, lyrics recognition was performed by extending the framework of speech recognition using hidden Markov model (HMM). However, accuracy rate was not sufficient. To recognize singing voice precisely, one promising approach is utilizing musical features. This study considers the task of recognizing syllable from a cappella singing voice. To respond to the variation of the length of a phoneme, we construct the duration dependent HMM. A large database of singing voice is essential for training the acoustic model. We use synthetic singing voice by HMM based singing voice synthesis system to solve the lack of the database of a cappella singing voice. We confirmed the effectiveness of our method.

Singing voice synthesizer with non-filter waveform generation using spectral phase reconstruction

This paper discusses a singing voice synthesis system based on non-filter waveform generation using spectral phase reconstruction as an alternative method to replace the conventional source-filter model. Source-filter model has been considered as an essential technique in the long history of speech synthesis as it simulates the human process of speech production consisting of excitation and resonances, even after hidden Markov model was introduced in the 90s toward statistically trainable speech synthesis. Filter (particularly, recursive filter), however, may cause serious problems in undesired amplitude and long time decay when sharp resonances in recursive filter match harmonics of F0 of the excitation. As the ultimate purpose of usage of filter is to transfer the spectrogram designed in the TTS system to the listener’s hearing organs, an alternative solution can be brought from “phase reconstruction” without using filters. We propose a spectral phase reconstruction, instead of using filter, to generate...

A HMM-based mandarin chinese singing voice synthesis system

We propose a mandarin Chinese singing voice synthesis system, in which hidden Markov model (HMM)-based speech synthesis technique is used. A mandarin Chinese singing voice corpus is recorded and musical contextual features are well designed for training. F0 and spectrum of singing voice are simultaneously modeled with context-dependent HMMs. There is a new problem, F0 of singing voice is always sparse because of large amount of context, i.e., tempo and pitch of note, key, time signature and etc. So the features hardly ever appeared in the training data cannot be well obtained. To address this problem, difference between F0 of singing voice and that of musical score (DF0) is modeled by a single Viterbi training. To overcome the over-smoothing of the generated F0 contour, syllable level F0 model based on discrete cosine transforms (DCT) is applied, F0 contour is generated by integrating two-level statistical models. The experimental results demonstrate that the proposed system outperforms the baseline system in both objective and subjective evaluations. The proposed system can generate a more natural F0 contour. Furthermore, the syllable level F0 model can make singing voice more expressive.

HMM 기반 TTS와 MusicXML을 이용한 노래음 합성

노래음 합성이란 주어진 가사와 악보를 이용하여 컴퓨터에서 노래음을 생성하는 것이다. 텍스트/음성 변환기에 널리 사용된 HMM 기반 음성합성기는 최근 노래음 합성에도 적용되고 있다. 그러나 기존의 구현방법에는 대용량의 노래음 데이터베이스 수집과 학습이 필요하여 구현에 어려움이 있다. 또한 기존의 상용 노래음 합성시스템은 피아노 롤 방식의 악보 표현방식을 사용하고 있어 일반인에게는 익숙하지 않으므로 읽기 쉬운 표준 악보형식의 사용자 인터페이스를 지원하여 노래 학습의 편의성을 향상시킬 필요가 있다. 이 문제를 해결하기 위하여 본 논문에서는 기존 낭독형 음성합성기의 HMM 모델을 이용하고 노래음에 적합한 피치값과 지속시간 제어방법을 적용하여 HMM 모델 파라미터 값을 변화시킴으로서 노래음을 생성하는 방법을 제안한다. 그리고 음표와 가사를 입력하기 위한 MusicXML 기반의 악보편집기를 전단으로, HMM 기반의 텍스트/음성 변환 합성기를 합성기 후단으로서 사용하여 노래음 합성시스템을 구현하는 방법을 제안한다. 본 논문에서 제안하는 방법을 이용하여 합성된 노래음을 평가하였으며 평가결과 활용 가능성을 확인하였다. Singing voice synthesis is the generation of a song using a computer given its lyrics and musical notes. Hidden Markov models (HMM) have been proved to be the models of choice for text to speech synthesis. HMMs have also been used for singing voice synthesis research, however, a huge database is needed for the training of HMMs for singing voice synthesis. And commercially available singing voice synthesis systems which use the piano roll music notation, needs to adopt the easy to read standard music notation which make it suitable for singing learning applications. To overcome this problem, we use a speech database for training context dependent HMMs, to be used for singing voice synthesis. Pitch and duration control methods have been devised to modify the parameters of the HMMs trained on speech, to be used as the synthesis units for the singing voice. This work describes a singing voice synthesis system which uses a MusicXML based music score editor as the front-end interface for entry of the notes and lyrics to be synthesized and a hidden Markov model based text to speech synthesis system as the back-end synthesizer. A perceptual test shows the feasibility of our proposed system.

Spectral modification technique in conversion of musical notes or tempos for singing voice synthesis system

The STRAIGHT time‐frequency representations (spectrograms) of singing voice signals in various musical notes and various tempos are observed to develop a high‐quality synthesis system of singing voice. The spectrogram of STRAIGHT, which is a very high‐quality analysis‐synthesis system, can represent the vocal tract information accurately. A conversion system of a musical note or a tempo of an input singing voice signal has been implemented based on the observation. As a result of the observation, the frequency warping of the STRAIGHT spectrogram based on a DP matching algorithm has been introduced into the system. It was found that the method using a differential of a smoothed spectrum as a spectral distance measure in the frequency warping produces subjectively better quality than that using a smoothed spectrum directly. It was also found that the method without spectral modification, i.e., only with pitch/tempo modification in the conversion, produces better quality than that using a differential of a smoothed spectrum. This can be caused by the destruction of naturalness in the method using a differential of a smoothed spectrum.