HiFiSinger: Towards High-Fidelity Neural Singing Voice Synthesis


HiFiSinger_pipline

ArXiv: arXiv:2009.01776

Authors

*Corresponding author

Abstract

High-fidelity singing voices usually require higher sampling rate (e.g., 48kHz, compared with 16kHz or 24kHz in speaking voices) with large range of frequency to convey expression and emotion. However, higher sampling rate causes the wider frequency band and longer waveform sequences and throws challenges for singing modeling in both frequency and time domains in singing voice synthesis (SVS). Conventional SVS systems that adopt moderate sampling rate (e.g., 16kHz or 24kHz) cannot well address the above challenges. In this paper, we develop HiFiSinger, an SVS system towards high-fidelity singing voice using 48kHz sampling rate. HiFiSinger consists of a FastSpeech based neural acoustic model and a Parallel WaveGAN based neural vocoder to ensure fast training and inference and also high voice quality. To tackle the difficulty of singing modeling caused by high sampling rate (wider frequency band and longer waveform), we introduce multi-scale adversarial training in both the acoustic model and vocoder to improve singing modeling. Specifically, 1) To handle the larger range of frequencies caused by higher sampling rate (e.g., 48kHz vs. 24kHz), we propose a novel sub-frequency GAN (SF-GAN) on mel-spectrogram generation, which splits the full 80-dimensional mel-frequency into multiple sub-bands (e.g. low, middle and high frequency bands) and models each sub-band with a separate discriminator. 2) To model longer waveform sequences caused by higher sampling rate, we propose a multi-length GAN (ML-GAN) for waveform generation to model different lengths of waveform sequences with separate discriminators. 3) We also introduce several additional designs and findings in HiFiSinger that are crucial for high-fidelity voices, such as adding F0 (pitch) and V/UV (voiced/unvoiced flag) as acoustic features, choosing an appropriate window/hop size for mel-spectrogram, and increasing the receptive field in vocoder for long vowel modeling in singing voices. Experiment results show that HiFiSinger synthesizes high-fidelity singing voices with much higher quality: 0.32/0.44 MOS gain over 48kHz/24kHz baseline and 0.83 MOS gain over previous SVS systems. Audio samples are available at https://speechresearch.github.io/hifisinger.

Contents

Audio Samples
1.1 Audio Quality
Ablation Studies
2.1 SF-GAN
2.2 ML-GAN
2.3 Pitch input
2.4 Window/hop size
2.5 Pitch Control
2.6 Duration control

Audio Samples

48kHz sampling rate is used unless otherwise stated.

Audio Quality

HiFiSinger_pipline

这么说来很不单纯,你陪我看海zhè me shuō lái hěn bú dān chún, nǐ péi wǒ kàn hǎi

Recording (48kHz) Baseline (24kHz up) XiaoiceSing (48kHz) HiFiSinger (48kHz)
Recording (24kHz) Baseline (24kHz) Baseline (48kHz) HiFiSinger (24kHz)

宁静的夏天,天空中繁星点点níng jìng de xià tiān, tiān kōng zhōng fán xīng diǎn diǎn

Recording (48kHz) Baseline (24kHz up) XiaoiceSing (48kHz) HiFiSinger (48kHz)
Recording (24kHz) Baseline (24kHz) Baseline (48kHz) HiFiSinger (24kHz)

坏的是我发现不知不觉不见到你不是很习惯huài de shì wǒ fā xiàn bú zhī bú jué bú jiàn dào nǐ bú shì hěn xí guàn

Recording (48kHz) Baseline (24kHz up) XiaoiceSing (48kHz) HiFiSinger (48kHz)
Recording (24kHz) Baseline (24kHz) Baseline (48kHz) HiFiSinger (24kHz)

Ablation Studies

SF-GAN

HiFiSinger_pipline

这么说来很不单纯,你陪我看海zhè me shuō lái hěn bú dān chún, nǐ péi wǒ kàn hǎi

HiFiSinger HiFiSinger with 0 SF-GAN HiFiSinger with 1 SF-GAN HiFiSinger with 5 SF-GAN

遇见一个人然后生命全改变,原来不是恋爱才有的情节yù jiàn yī gè rén rán hòu shēng mìng quán gǎi biàn,yuán lái bú shì liàn ài cái yǒu de qíng jié

HiFiSinger HiFiSinger with 0 SF-GAN HiFiSinger with 1 SF-GAN HiFiSinger with 5 SF-GAN

我的小鬼小鬼,逗逗你的眉眼,让你喜欢这世界wǒ de xiǎo guǐ xiǎo guǐ, dòu dòu nǐ de méi yǎn, ràng nǐ xǐ huān zhè shì jiè

HiFiSinger HiFiSinger with 0 SF-GAN HiFiSinger with 1 SF-GAN HiFiSinger with 5 SF-GAN

ML-GAN

HiFiSinger_pipline

谁在最需要的时候轻轻拍着我肩膀shuí zài zuì xū yào de shí hòu qīng qīng pāi zhe wǒ jiān bǎng

HiFiSinger HiFiSinger without ML-GAN

见证你成长让我感到充满力量jiàn zhèng nǐ chéng zhǎng ràng wǒ gǎn dào chōng mǎn lì liàng

HiFiSinger HiFiSinger without ML-GAN

你知道它的花语签上名,我继续一个人远行nǐ zhī dào tā de huā yǔ qiān shàng míng, wǒ jì xù yī gè rén yuǎn xíng

HiFiSinger HiFiSinger without ML-GAN

Pitch and Voiced/Unvoiced input

HiFiSinger_pipline

这么说来很不单纯,你陪我看海zhè me shuō lái hěn bú dān chún, nǐ péi wǒ kàn hǎi

HiFiSinger HiFiSinger without F0 and V/UV input

遇见一个人然后生命全改变,原来不是恋爱才有的情节yù jiàn yī gè rén rán hòu shēng mìng quán gǎi biàn, yuán lái bú shì liàn ài cái yǒu de qíng jié

HiFiSinger HiFiSinger without F0 and V/UV input

我的小鬼小鬼,逗逗你的眉眼,让你喜欢这世界wǒ de xiǎo guǐ xiǎo guǐ, dòu dòu nǐ de méi yǎn, ràng nǐ xǐ huān zhè shì jiè

HiFiSinger HiFiSinger without F0 and V/UV input

window/hop size

这么说来很不单纯,你陪我看海zhè me shuō lái hěn bú dān chún, nǐ péi wǒ kàn hǎi

HiFiSinger HiFiSinger with 12ms window HiFiSinger with 50ms window

遇见一个人然后生命全改变,原来不是恋爱才有的情节yù jiàn yī gè rén rán hòu shēng mìng quán gǎi biàn, yuán lái bú shì liàn ài cái yǒu de qíng jié

HiFiSinger HiFiSinger with 12ms window HiFiSinger with 50ms window

我的小鬼小鬼,逗逗你的眉眼,让你喜欢这世界wǒ de xiǎo guǐ xiǎo guǐ, dòu dòu nǐ de méi yǎn, ràng nǐ xǐ huān zhè shì jiè

HiFiSinger HiFiSinger with 12ms window HiFiSinger with 50ms window

Pitch Control

HiFiSinger_pipline

左心房,暖暖的好饱满zuǒ xīn fáng, nuǎn nuǎn de hǎo bǎo mǎn

Normal scale Down 4 semitones Up 4 semitones

我想说其实你很好,你自己却不知道wǒ xiǎng shuō qí shí nǐ hěn hǎo, nǐ zì jǐ què bú zhī dào

Normal scale Down 4 semitones Up 4 semitones

在朋友里面就数你最特别,总让我觉得很亲很铁zài péng yǒu lǐ miàn jiù shù nǐ zuì tè bié, zǒng ràng wǒ jué dé hěn qīn hěn tiě

Normal scale Down 4 semitones Up 4 semitones

Duration Control

HiFiSinger_pipline

因为我,完全信任你yīn wéi wǒ, wán quán xìn rèn nǐ

0.75x Speed 1.00x Speed 1.25x Speed

我想说其实你很好,你自己却不知道wǒ xiǎng shuō qí shí nǐ hěn hǎo, nǐ zì jǐ què bú zhī dào

0.75x Speed 1.00x Speed 1.25x Speed

杜鹃啼血声,芙蓉花蜀国尽缤纷dù juān tí xuè shēng, fú róng huā shǔ guó jìn bīn fēn

0.75x Speed 1.00x Speed 1.25x Speed

Some speech research conducted at Microsoft Research Asia
NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality
FastSpeech: Fast, Robust and Controllable Text to Speech
FastSpeech 2: Fast and High-Quality End-to-End Text-to-Speech
MultiSpeech: Multi-Speaker Text to Speech with Transformer
Almost Unsupervised Text to Speech and Automatic Speech Recognition
LRSpeech: Extremely Low-Resource Speech Synthesis and Recognition
UWSpeech: Speech to Speech Translation for Unwritten Languages
Denoising Text to Speech with Frame-Level Noise Modeling
AdaSpeech: Adaptive Text to Speech for Custom Voice
AdaSpeech 2: Adaptive Text to Speech with Untranscribed Data
AdaSpeech 3: Adaptive Text to Speech for Spontaneous Style
AdaSpeech 4: Adaptive Text to Speech in Zero-Shot Scenarios
DeepSinger: Singing Voice Synthesis with Data Mined From the Web
HiFiSinger: Towards High-Fidelity Neural Singing Voice Synthesis