Challenge of Singing Voice Synthesis Using Only Text-To-Speech Corpus with FIRNet Source-Filter Neural Vocoder

T. Okamoto, Y. Ohtani, S. Shimizu, T. Toda, and H. Kawai, "Challenge of singing voice synthesis using only text-to-speech corpus with FIRNet source-filter neural vocoder," in Proc. Interspeech, Sept. 2024, pp. 1870–1874. [ISCA Archive]

[Musical scores used in experiments (csv)]

---

---

Text-to-speech (TTS)

Text (in Japanese): 水をマレーシアから買わなくてはならないのです．

Original

CFS2 + HiFi-GAN (melspc)	CFS2 + HiFi-GAN (WORLD)	PESC + HiFi-GAN (WORLD)

CFS2 + FIRNet (WORLD)	PESC + FIRNet (WORLD)

Singing voice synthesis (SVS)

Lyrics (in Japanese): でんでんむしむしかたつむり

Original

CFS2 + HiFi-GAN (melspc)	CFS2 + HiFi-GAN (WORLD)	PESC + HiFi-GAN (WORLD)	PESC + HiFi-GAN + Input fo shift (WORLD)

CFS2 + FIRNet (WORLD)	PESC + FIRNet (WORLD)	PESC + FIRNet + Input fo shift (WORLD)

PESC + FIRNet + Input fo shift + Direct fo input (WORLD) Not included in experiments

T x 0.5: HiFi-GAN (WORLD)	T x 0.5: FIRNet (WORLD)

fo x 0.5: HiFi-GAN (WORLD)	fo x 0.5: FIRNet (WORLD)

fo x 2.0: HiFi-GAN (WORLD)	fo x 2.0: FIRNet (WORLD)

Lyrics (in Japanese): げんこつやまのたぬきさん

Original

CFS2 + HiFi-GAN (melspc)	CFS2 + HiFi-GAN (WORLD)	PESC + HiFi-GAN (WORLD)	PESC + HiFi-GAN + Input fo shift (WORLD)

CFS2 + FIRNet (WORLD)	PESC + FIRNet (WORLD)	PESC + FIRNet + Input fo shift (WORLD)

PESC + FIRNet + Input fo shift + Direct fo input (WORLD) Not included in experiments

T x 0.5: HiFi-GAN (WORLD)	T x 0.5: FIRNet (WORLD)

fo x 0.5: HiFi-GAN (WORLD)	fo x 0.5: FIRNet (WORLD)

fo x 2.0: HiFi-GAN (WORLD)	fo x 2.0: FIRNet (WORLD)