cchiou-amba/HiFiGAN

HiFi-GAN v1 Neural Vocoder (Ambarella N1-655 CVflow v3 NVP)

HiFi-GAN v1 is a generative adversarial network for high-fidelity neural vocoding that converts 80-bin log-Mel spectrograms ([1, 80, 50]) into 22.05 kHz time-domain audio waveforms ([1, 12800]) using 50-frame overlap tiling (core=20, ctx=15) on the Ambarella N1-655 CVflow v3 NVP.

Original paper: HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis

Model Configuration:

• Reference implementation: https://github.com/jik876/hifi-gan
• Original Weight: https://huggingface.co/csukuangfj/sherpa-onnx-hifigan
• Dataset: LJSpeech-1.1 / Multi-Speaker VCTK / AISHELL-3 (22.05 kHz)
• License: MIT (LICENSE)
• Support Cooper version:
  • CNNGen: [3.9.0 / n1-655x] (cavalry_gen 3.0.4)

Upstream Attribution & Modification Notice

• Upstream Copyright: Copyright (c) 2020 Jungil Kong (jik876/hifi-gan), licensed under the MIT License; ONNX export provided by csukuangfj/sherpa-onnx-hifigan (sherpa-onnx contributors).
• Modification Notice: Modified by Ambarella International LLC (2026). Fixed dynamic Mel frame axis to a 50-frame overlap-tiling window ([1, 80, 50] -> [1, 12800], hop=256) and compiled via Ambarella Cooper SDK CNNGen 3.9.0 (cavalry_gen 3.0.4) into CVflow v3 N1-655 (vdg_arch = 0xB) Cavalry DAG binary format (n1-655_hifigan_vocoder_amba_optimized.bin).

NVP Port & Runtime Specification

• Target SoC: Ambarella N1-655 (vdg_arch = 0xB, CVflow v3 cavalry_gen 3.0.4)
• Input port: mel [1, 80, 50] IEEE FP16 (run with 50-frame overlap tiling: core=20 frames, ctx=15 context frames on each side)
• Output port: audio [1, 12800] IEEE FP16 per tile (crop middle 20 * 256 = 5120 samples per tile @ 22.05 kHz)

Model	Device	Compression	Model Link
HiFiGAN	N1-655	Amba_optimized	Model_Link