Abstract
GAN vocoders reconstruct waveforms from acoustic features or compressed audio representations with predominantly one-pass generation. They are mature speech-generation infrastructure, but their perceptual quality depends as much on discriminator design, conditioning fidelity, data coverage, and evaluation protocol as on the generator itself.
Current State
Adversarial generation remains a strong practical default, but its speed advantage is easier to establish than universal perceptual superiority. HiFi-GAN demonstrated that a compact adversarial generator can approach autoregressive quality at real-time-capable speeds, while later systems broadened the operating range (HiFi-GAN, BigVGAN). The reviewed evidence still treats the matched quality-versus-latency comparison as emerging because hardware, sampling rate, upstream acoustic generation, and listening protocols are rarely controlled together.
The discriminator is a central part of the vocoder, not merely a training accessory. Period-aware waveform critics and multi-resolution spectral critics expose complementary failures in periodic structure, transients, and spectral detail. This conclusion is strongly supported across waveform generators and codec decoders, although no discriminator set is uniformly best across representations (HiFi-GAN, High Fidelity Neural Audio Compression, Universal Harmonic Discriminator).
Signal structure is increasingly built into the model. Frequency-domain generation, explicit phase modeling, anti-aliasing, harmonic decomposition, and band-aware representations can improve efficiency and suppress artifacts (Vocos, QHARMA-GAN, Aliasing-Free Neural Audio Synthesis). These constraints improve extrapolation only when their phase, voicing, or band assumptions remain valid; explicit phase modeling as a general replacement for adversarial supervision remains emerging (Is GAN Necessary?).
A vocoder must be trained for the distribution it will actually decode. Predicted spectrograms, quantized latents, bandwidth-limited features, and streamed chunks differ systematically from clean reference conditioning. Adaptation to those inputs is strongly supported, but can overfit one acoustic model or codec (Tacotron 2, Spectrogram Patch Codec). Speaker, language, and recording diversity also improve transfer, though data scale and architecture are often confounded.
Codec decoding and streaming have become first-class vocoder settings. Adversarial decoders remain effective at low bitrates, while causal GANs can approach offline quality with limited lookahead or distillation. Flow and diffusion vocoders provide another refinement path, but iterative inference retains a quality–latency trade-off even as one-step and short-trajectory variants narrow it (High Fidelity Neural Audio Compression, Low-Latency Vocoder Analysis, PeriodWave).
Single automatic metrics are not reliable enough to rank vocoders across operating conditions. Spectral distances and learned MOS predictors can disagree with listeners when phase, pitch, bandwidth, compression, or intended signal transformation changes. A late conversational-conversion result adds only weak refining evidence because its scores and protocol are underspecified (Is GAN Necessary?, Bridging the Gap).
Method Landscape
- Adversarial waveform and time–frequency vocoders use one-pass generators with periodic, spectral, harmonic, or projected critics. They offer the clearest latency advantage, but require careful stabilization and matched conditioning.
- Codec and VAE decoders reconstruct detail from compressed continuous or discrete latents. Their outcome is jointly determined by token rate, quantizer structure, and decoder capacity rather than by the vocoder alone.
- Flow and diffusion vocoders iteratively transport noise or degraded audio toward a waveform. They can refine difficult reconstructions, but must be distilled or shortened for interactive use.
- Hybrid signal–neural systems embed phase, source–filter, filterbank, band, or anti-aliasing structure. They can be efficient and data-aware, while introducing explicit estimation bottlenecks.
- Autoregressive decoders remain a quality and historical reference where latency is secondary, rather than the default deployment choice.
Key Trade-offs
- Speed versus controlled comparison. Fast generation does not by itself establish equal end-to-end quality when hardware and upstream models differ.
- Learned flexibility versus analytic structure. Harmonic and phase constraints can improve robustness, but failures in voicing or phase estimation become audible model assumptions.
- Clean training targets versus deployment realism. Training on predicted or quantized conditioning improves robustness to that source but may reduce portability to another front end.
- Compression versus reconstruction detail. Lower token rates reduce sequence and storage cost while placing more perceptual burden on the decoder.
- Offline quality versus streaming latency. Lookahead and chunk size improve continuity but delay the first usable audio and can conceal state-reset artifacts.
Open Questions
- Which discriminator combination transfers reliably across speech, music, sampling rates, codecs, and recording domains?
- When can explicit phase or signal-processing structure replace adversarial supervision under matched listener tests?
- Can one-step flow or diffusion vocoders consistently match GAN latency and quality across datasets?
- How should training balance clean, predicted, quantized, and streamed conditioning without overfitting one upstream model?
- Which evaluation suite captures phase, pitch, bandwidth, transient, and out-of-domain artifacts without misranking listener preference?
Go Deeper
Read GAN Vocoders: In Depth for the research landscape, areas of convergence and uncertainty, deployment implications, and a representative reading path.
Scope
This synthesis covers 82 reviewed papers published from 2016 through the Q4 2025 evidence window. The corpus spans TTS, voice conversion, neural codecs, streaming, restoration, and non-standard acoustic domains, but many comparisons reuse common datasets or alter architecture and data together. Strong support therefore describes convergence within the encoded evidence, not a universal ranking of vocoders.
This page is generated from wiki/_claims/gan-vocoder.yaml (digest date: 2026-10-31). For complete structured claims and paper-level provenance, use the source YAML.