Voice vocoder speed vs quality: HiFi-GAN vs BigVGAN, 48 kHz for +0.2 MOS

The mechanism check is therefore direct: if a comparison keeps the HiFi-GAN generator and changes only activation and anti-aliasing, the quality difference is attributable to those two changes; if parameter count also changes, the comparison is no longer architecture-only. Hold mel inputs, hardware, and batch size constant so capacity does not masquerade as an activation or anti-aliasing gain.

What the published numbers actually show

The published evidence converges on one capacity figure rather than on a sample-rate MOS gain. The BigVGAN abstract states that the model is trained at the largest scale up to 112M parameters, calling that unprecedented in the literature, while the alphaXiv summary contrasts it with typical 14M-parameter models. Taken together, those two sources give a traceable 8x capacity ratio, 112 divided by 14, and they agree on the same two endpoints.

The headline claim is scoped to generalization, not to sample rate. The abstract says BigVGAN is trained only on clean LibriTTS speech, yet achieves state-of-the-art performance for zero-shot conditions: unseen speakers, languages, recording environments, singing voices, music, and instrumental audio. That is an out-of-distribution claim tied to training data, not a claim about a sample-rate MOS increment.

Because the zero-shot list is explicit, the reader can check whether a reported advantage applies to their own inputs. If your use case is in-distribution speech, the published numbers do not by themselves show that the larger model is the better serving choice; they show that it was evaluated across the listed out-of-distribution categories. The capacity figure and the generalization scope should therefore be recorded separately.

The cross-source agreement matters: alphaXiv repeats the 112M versus 14M contrast, so the capacity figure is not an isolated sentence in the abstract. The zero-shot scope is likewise stated in the abstract. In this evidence set, those are the stable, traceable claims to carry forward, and they are the claims a reviewer can verify without rerunning training.

A practical reading is to keep three buckets separate: capacity (112M versus 14M), generalization scope (the zero-shot list), and any cost or MOS claim. Only the first two are directly supported by the cited sources here; the third requires your own controlled test on identical mel inputs, hardware, and batch size.

HiFi-GAN vs BigVGAN: which wins where

BigVGAN-base is the default pick for general serving when one vocoder must cover multiple speakers, languages, recording conditions, or non-speech audio, but only after it clears the 0.2 MOS margin with non-overlapping 95% confidence intervals on your own data, per the gate defined in the worked test. The 112M checkpoint is not the default; reserve it for out-of-distribution serving, where the zero-shot universal-vocoder claim from arXiv:2206.04658 is actually being exercised.

HiFi-GAN at its typical ~14M parameters remains the winner for in-domain, clean, single-speaker, latency-bound work. Keep it as both the production baseline and the control in your MOS test. If BigVGAN does not clear the required margin on your utterances, HiFi-GAN is the shipping choice, and the burden of proof has not been met.

The 112M HiFi-GAN ablation is a research and demo control, not a product. The demo identifies it as the same-capacity HiFi-GAN with the architectural improvements disabled, so its purpose is to separate architecture from capacity. It should never be treated as a shipping option, and it should not be cited as proof that a 112M HiFi-GAN is a practical deployment target.

Apply the gate as specified in the worked test below: identical mel inputs, identical hardware, batch size 1, at least 30 utterances, and non-overlapping 95% confidence intervals around a 0.2 MOS difference. If any element is missing, the comparison is not decision-grade, and the default remains HiFi-GAN.

Costs: parameters, RTF, and 48 kHz tax

Serving cost is where the architectural scale of BigVGAN becomes concrete, and this section does the serving-cost arithmetic: the 8x parameter ratio and the derived samples-per-second ratio between 48 kHz and 22.05 kHz. As reported in the BigVGAN paper (arXiv:2206.04658) and summarized by alphaXiv, the largest BigVGAN model is trained at up to 112M parameters, compared with a typical 14M-parameter HiFi-GAN baseline, yielding 112M / 14M = 8x the parameter count. Because that added capacity sits in the upsampling stack, it increases memory bandwidth and compute per generated second of audio; treat this as a measurement target rather than assuming a fixed slowdown on your hardware.

Sample rate adds a separate, derived cost that is independent of model size. Generating 48 kHz audio requires more output samples per second of speech than common lower-rate settings, so the final upsample layers perform proportionally more work. The ratios below are arithmetic derivations, not benchmark results, and should be combined with your measured runtime.

To price a deployment, report real-time factor (RTF) as audio seconds generated divided by wall-clock seconds, measured on your target CPU and GPU at batch size 1. Use identical mel inputs and identical hardware when comparing configurations, and record both steady-state throughput and any initialization overhead separately. An RTF above 1.0 indicates faster-than-real-time generation on that device, but the decision-relevant number is the RTF margin you retain under your expected load and latency budget.

Evaluate the 8x parameter increase and the 48 kHz sample-rate tax together, not in isolation. A 112M-parameter model at 48 kHz compounds model capacity with 2.18x the sample output of 22.05 kHz, so memory pressure, thermal behavior, and tail latency can diverge from small-batch demos. If your application does not require 48 kHz output, measure whether a lower rate meets the quality bar on your own utterances before accepting the additional per-second work.

Before committing to a higher-cost vocoder, run the same cost audit required by the quality check: test on 30 or more of your own utterances with identical mel inputs, identical hardware, and batch size 1. Pair the measured RTF and resource usage with your listening results so the serving-cost arithmetic and the perceptual evidence come from the same controlled setup.

What the evidence does not establish

No source in the cited evidence set measures a MOS delta at 48 kHz, so the "+0.2 MOS" premise is unverified rather than established. The cited material does trace an architectural difference—periodic activation plus anti-aliased representation—at up to 112M parameters versus a typical 14M HiFi-GAN (alphaXiv summary of arXiv:2206.04658), but that is a capacity and inductive-bias claim, not a sample-rate MOS measurement.

The gap matters because a larger model can look better for reasons unrelated to 48 kHz vocoding: more parameters, different training data, or a different evaluation set. The supplied ablation model, HiFi-GAN (112M), keeps the larger capacity while disabling the architectural improvements, which is the control you need if you want to attribute a gain to activation and anti-aliasing rather than to size alone (bigvgan-demo.github.io).

The boundary cases show where the capacity argument helps and where it does not:

In the studio-clean case, the extra capacity has no out-of-distribution structure to exploit, so the smaller HiFi-GAN at the deployed rate is the rational default; in the unseen-language or music case, the question shifts from sample rate to whether BigVGAN's zero-shot claim holds on your own data.

The official NVIDIA/BigVGAN GitHub repository lists 22 kHz BigVGAN models with a band-limited setup (fmax=8000), not a 48 kHz MOS comparison, so the repository's own model notes do not close the 48 kHz gap either.

The honest reading is that 48 kHz "+0.2 MOS" remains an unverified premise: keep it out of procurement or deployment decisions until your own paired test, as specified in the worked test below, clears both the 0.2 MOS margin and the non-overlapping 95% confidence interval.

Worked test: 30 utterances, 3 checkpoints

Fix the inputs before any listening. Assemble 30 utterances drawn as 6 speakers × 5 utterances, each at least 10 seconds long, either native 48 kHz or resampled once with a documented filter. Generate one mel spectrogram per utterance from a single acoustic model and feed that mel byte-identically to both vocoders. Run every render on the same GPU, at batch size 1, with no denoising or post-processing, so any audible difference can be attributed to the vocoder rather than the pipeline.

Checkpoint 1 is a hard gate because it is cheap to run and expensive to ignore. Record RTF and peak VRAM under the same serving conditions you intend to ship; a model that fails here cannot justify a listening study. The 2.18× sample-rate multiplier at 48 kHz versus 22.05 kHz is a derived expectation, not a substitute for your own measurement, but it sets the scale of the trade-off you are testing.

Checkpoint 2 turns preference into evidence. The 600 blind ratings from 6 speakers × 5 utterances give you a balanced, speaker-diverse panel while keeping the workload bounded. Because every listener hears the same byte-identical mel renderings from both systems, the resulting intervals reflect vocoder differences rather than content or synthesis variance.

Checkpoint 3 prices the cost of a wrong call in both directions. A false adopt commits you to higher inference cost and operational complexity for a gain your users cannot reliably hear; a false reject leaves audible quality on the table when BigVGAN would have cleared the 0.2 MOS margin on your own audio. Running all three checkpoints on your material makes that trade-off explicit instead of inferential.

Decision rules you can run today

If your voices are in-domain, clean, and single-speaker, deploy HiFi-GAN at your actual serving rate and stop there: 112M parameters have nothing to generalize, and a 0.2 MOS gap that does not survive your own 30-utterance test with non-overlapping 95% confidence intervals is not a deployment reason.

If you serve unseen speakers, languages, noisy speech, singing, music, or instrumental audio, use BigVGAN, but read the checkpoint's fmax in the NVIDIA/BigVGAN repository before you state any sample rate; the repo documents 22 kHz models with a band-limited fmax of 8000, so a checkpoint's configured limit, not the model name, is what your rate claim must cite.

If a vendor or paper quotes "+0.2 MOS at 48 kHz," require a matched-learning-rate, matched-parameter table with confidence intervals before you accept it. No such table exists in this evidence set; SpecDiff-GAN's fixed lr=1e-4 comparison table is the nearest available reference, and it is not a matched-parameter table.

As established above, capacity is a separate claim from rate: the 112M versus typical 14M comparison tells you how much model you are buying, not that 48 kHz vocoding adds 0.2 MOS. Treat any rate-specific MOS quote as unverified until it is tied to the same mel inputs, same hardware, and same batch size as your deployment.

Frequently Asked Questions

When does a vocoder comparison stop being architecture-only?

The mechanism check is therefore direct: if a comparison keeps the HiFi-GAN generator and changes only activation and anti-aliasing, the quality difference is attributable to those two changes; if parameter count also changes, the comparison is no longer architecture-only.

What must be held constant to prevent capacity from masquerading as an activation or anti-aliasing gain?

Hold mel inputs, hardware, and batch size constant so capacity does not masquerade as an activation or anti-aliasing gain.

What capacity ratio do the BigVGAN abstract and alphaXiv summary jointly support?

Taken together, those two sources give a traceable 8x capacity ratio, 112 divided by 14, and they agree on the same two endpoints.

What training-data limitation accompanies BigVGAN's zero-shot performance claim?

The abstract says BigVGAN is trained only on clean LibriTTS speech, yet achieves state-of-the-art performance for zero-shot conditions: unseen speakers, languages, recording environments, singing voices, music, and instrumental audio.

Is the headline claim about sample-rate MOS or generalization?

The headline claim is scoped to generalization, not to sample rate.

What kind of claim is BigVGAN's zero-shot performance?

That is an out-of-distribution claim tied to training data, not a claim about a sample-rate MOS increment.

Quick answers

Also worth reading: Voice cloning noisy audio: 10s vs 60s Mean Opinion Score (MOS) and Word Error Rate (WER): Voice cloning noisy audio: 10s · Voice cloning robustness test: Mean Opinion Score (MOS) 4.2 floor — ship or retrain?: Voice cloning robustness test: Mean · Exploring voice cloning effects on audio file fidelity: Exploring voice cloning effects on