Neural Vocoders
Networks that turn spectrograms or acoustic features into audio waveforms quickly and with high fidelity.
Tier Grades
Tier A: Frontier AI labs
feasible
NVIDIA trained BigVGAN at up to 112M parameters.
Tier B: Indonesian AI labs
feasible
HiFi-GAN runs 167.9 times faster than real time on one V100, so training and serving fit a national lab.
Tier C: Institution units such as ITB
feasible
HiFi-GAN came from a 3-author team and trains on public speech, which suits a unit.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
The small HiFi-GAN runs 13.4 times faster than real time on CPU, so it can ship in an on-premise proof of concept.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeHiFi-GAN generates 22.05 kHz audio 167.9 times faster than real time on one V100 and 13.4 times on CPU; BigVGAN trains a GAN vocoder of up to 112M parameters.
- DataBigVGAN is trained only on clean speech (LibriTTS) and generalises to unseen speakers, languages and music.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 3 authors on HiFi-GAN; 5 on BigVGAN.
- Time horizonMonths.
- Shelf-life signalsCodec-token and flow-matching TTS that decode waveforms without a separate vocoder
- Pipeline roleTier A and Tier C publish vocoders; Tier B pairs them with Indonesian acoustic models; Tier D deploys them inside TTS pipelines.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
BigVGANspeech synthesisexemplar
HiFi-GANspeech synthesisexemplar
WaveNetspeech synthesisexemplarVITSspeech synthesisproduced by
DiffWavespeech synthesisproduced byFastPitchspeech synthesisproduced by
FastSpeech 2speech synthesisproduced byGlow-TTSspeech synthesisproduced by
Parallel WaveGANspeech synthesisproduced by
MelGANspeech synthesisproduced byWaveGlowspeech synthesisproduced byWaveRNNspeech synthesisproduced by
Tacotron 2speech synthesisproduced byLinks
R&D TopicNeural Vocoders
ModelsHiFi-GANKakao Enterprise, speech synthesisBigVGANNVIDIA, speech synthesisWaveNetGoogle DeepMind, speech synthesisWaveGlowNVIDIA, speech synthesis+9 more
HiFi-GANKakao Enterprise, speech synthesisBigVGANNVIDIA, speech synthesisWaveNetGoogle DeepMind, speech synthesisWaveGlowNVIDIA, speech synthesis+9 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 13
FastPitchNVIDIA, speech synthesisproduced byBigVGANNVIDIA, speech synthesisexemplarDiffWaveBaidu, speech synthesisproduced byFastSpeech 2Microsoft AI, speech synthesisproduced byGlow-TTSKakao Enterprise, speech synthesisproduced byHiFi-GANKakao Enterprise, speech synthesisexemplarMelGANDescript, speech synthesisproduced byParallel WaveGANNaver, speech synthesisproduced byTacotron 2Google, speech synthesisproduced byVITSKakao Enterprise, speech synthesisproduced byWaveGlowNVIDIA, speech synthesisproduced byWaveNetGoogle DeepMind, speech synthesisexemplarPapers
- WaveNet: A Generative Model for Raw AudioAaron van den Oord et al., arXiv preprint, 2016 founding
- Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesisHubert Siuzdak, arXiv preprint, 2023 state of the art
- BigVGAN: A Universal Neural Vocoder with Large-Scale TrainingSang-gil Lee et al., ICLR 2023 (per arXiv comment), 2022 state of the art
- HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech SynthesisJungil Kong et al., NeurIPS 2020 (per arXiv comment), 2020 state of the art
- WaveGlow: A Flow-based Generative Network for Speech SynthesisRyan Prenger et al., arXiv preprint, 2018 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2010.05646 https://arxiv.org/abs/2010.05646
- arxiv.org/abs/2206.04658 https://arxiv.org/abs/2206.04658
- arxiv.org/abs/1609.03499 https://arxiv.org/abs/1609.03499
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Neural Vocoders", https://aurion.semai.id/topics/neural-vocoders/, records of October 2026.