Demo samples: Multi-stream HiFi-GAN with data-driven waveform decomposition

T. Okamoto, T. Toda and H. Kawai,
"Multi-stream HiFi-GAN with data-driven waveform decomposition,"
in Proc. ASRU, Dec 2021, pp. 610–617.



Proposed Multi-stream HiFi-GAN generator trained using the same discriminators as original HiFi-GAN
Proposed Multi-stream HiFi-GAN can accelerate the inference speed while keeping the synthesis quality



(a): Transposed convolution [as used in original HiFi-GAN], (b): Interpolation + Convolution, (c): Sub-pixel convolution
Multi-speaker models are trained using JVS corpus

Audio Samples

Analysis-synthesis condition

Female (JAF001)
Original Sub-DiffWave (25 iterations) Multi-band MelGAN Multi-stream MelGAN
HiFi-GAN V1 (a) HiFi-GAN V2 (a) Multi-stream HiFi-GAN

Male (JAM017)
Original Sub-DiffWave (25 iterations) Multi-band MelGAN Multi-stream MelGAN
HiFi-GAN V1 (a) HiFi-GAN V2 (a) Multi-stream HiFi-GAN


Text-to-speech condition (Parallel Tacotron 2 with forced alignment)

Female (JAF001)
Sub-DiffWave (25 iterations) HiFi-GAN V1 (a) Milti-stream HiFi-GAN

Male (JAM017)
Sub-DiffWave (25 iterations) HiFi-GAN V1 (a) Milti-stream HiFi-GAN


Multi-speaker condition (analysis-synthesis of unseen speaker)

Female (JAF001)
HiFi-GAN V1 (c) HiFi-GAN V2 (c) Multi-stream HiFi-GAN

Male (JAM017)
HiFi-GAN V1 (c) HiFi-GAN V2 (c) Multi-stream HiFi-GAN