# Voice cloning robustness test: Mean Opinion Score (MOS) 4.2 floor — ship or retrain?

Dylan Cooper · October 22, 2026

> Ship a 4.2-MOS voice clone only if worst-case MOS is at least 4.0 and speaker similarity is at least 0.80 across three degraded conditions; otherwise retrain.

| Takeaway | Detail |
| --- | --- |
| Ship a 4.2-MOS voice clone only if its worst-case MOS across three corruption tests is at least 4.0. | Clean-set MOS of 4.2 is insufficient; any corrupted condition below 4.0 means retrain with corrupted-condition augmentation. |
| Require speaker similarity of at least 0.80 under noise, reverberation, and bandwidth loss. | The clone must hold MOS ≥ 4.0 and speaker similarity ≥ 0.80 across at least three degraded acoustic conditions. |
| Use benchmark-scale evidence, not clean scores alone: RobustBench tracks over 3,000 adversarial-robustness papers. | The available sources cites more than 3,000 papers on adversarial robustness, underscoring that robustness must be measured, not assumed. |
| Corruption benchmarks can include many perturbations; test at least three acoustic degradations. | One cited benchmark covers 7 datasets with 96 visual and 87 textual corruptions, illustrating broad corruption testing; voice tests should cover noise, reverberation, and bandwidth loss. |

This guide defines a ship-or-retrain rule for voice clones using clean and corrupted acoustic tests.

It specifies MOS thresholds, a speaker-similarity floor of 0.80, and three named acoustic degradations — noise, reverberation, and bandwidth loss — drawn from the robustness-testing literature.

![Voice cloning robustness test](https://static.mm-ais.com/article-images-ai/voice-cloning-robustness-test-mean-opini-ai-cfce6c94.jpg)

## Why a clean-set 4.2 MOS lies to you

A clean studio reference inflates a voice clone's MOS the same way clean test sets inflate adversarial robustness claims: the evaluation condition is forgiving, so the score reflects ease of the test as much as quality of the model. When raters compare a clone to pristine audio, errors in timbre, prosody, or artifacts are harder to notice because there is no competing noise, reverberation, or bandwidth loss to expose them. A 4.2 measured only under that condition is therefore an upper bound on perceived quality, not proof that the clone will survive harder listening conditions.

RobustBench exists because robustness claims from thousands of adversarial-robustness papers proved hard to compare; its stated goal is to systematically track real progress in adversarial robustness. That lesson transfers directly to MOS scores measured only on clean studio audio: a clone can look strong because the test set is narrow, not because it is robust. Check the benchmark's current methods and findings at robustbench.github.io before citing them as evidence for audio evaluation.

The mechanism is overadaptation to the evaluation condition. In adversarial robustness, a model may exploit regularities of clean inputs and then fail when those regularities disappear under corruption. In voice cloning, a model may exploit the regularities of clean reference audio, such as high signal-to-noise ratio, little reverberation, and full bandwidth, while remaining fragile when those regularities are removed. The MOS is then a property of the test condition as much as a property of the clone.

RobustBench's solution maps directly to voice-clone evaluation: flag models that may be overadapted to the standard attack and welcome adaptive or external evaluations. For voice clones, that means flagging a model that may be overadapted to clean studio audio and requiring evaluation under degraded acoustic conditions before release. The clean-set score should trigger scrutiny, not confidence, because it is exactly the kind of narrow result that standardized corruption testing was created to challenge.

Practically, treat a clean-studio MOS as a screening result. If the only evidence is clean audio, assume the score is inflated until degraded-condition evidence contradicts that assumption. The burden of proof should sit with the clean-set number, because the mechanism of overestimation is the same one RobustBench identified in adversarial robustness: narrow evaluation makes weak robustness look like strong performance.

![Why a clean-set 4.2 MOS lies to you — Voice cloning robustness test](https://static.mm-ais.com/article-images-ai/voice-cloning-robustness-test-mean-opini-ai-4e41a593.jpg)

## Run the three-condition MOS test

Run the three-condition test on the clone's synthesized output, not on the reference recording, and keep the script, rater pool, and MOS scale identical across all conditions so any drop reflects the corruption rather than protocol drift. For each condition, record MOS and speaker similarity on the same sheet, then take the worst-case MOS across the three as the shipping number. If any condition fails the gate, the clean studio score does not rescue it.

Condition 1 is additive noise: mix the clone output with babble or cafe noise at 10 dB SNR. Test the output at this harshest contemplated level rather than at a cleaner one, score the noisy output for naturalness and intelligibility, and note whether noise masks phonemes or speaker cues. Measure the corrupted output's SNR to confirm the 10 dB condition before scoring.

Condition 2 is reverberation: convolve the output with a small-room impulse response. Reverberation corrupts prosody and the speaker-embedding match, because the reference utterance encodes language, dialect, accent, and local pronunciation habits that degrade under room coloration. Listen for smeared consonants, shifted pitch contours, and a less stable speaker identity.

Condition 3 is bandwidth loss: restrict the output with a low-pass filter that removes high-frequency content, simulating a bandwidth-limited channel, then score the result. This isolates whether speaker identity and naturalness survive when the signal is bandwidth-starved. Keep the filter settings fixed across clones so the comparison is fair.

Use one scoring sheet for all three conditions. A row passes only when its MOS and speaker similarity meet the ship gate; the clone ships only if every row passes and the worst-case MOS stays at or above the threshold. Retrain if any row fails, even when the clean studio score is high.

![berry red mean snowball](https://static.mm-ais.com/article-images-pixabay/voice-cloning-robustness-test-mean-opini-f82f1154.jpg)
berry red mean snowball

## Ship, retrain, or re-record: the table

Treat the three-condition result as the decision input: the table below assigns one action to each outcome, and the clean 4.2 score is context rather than permission to ship.

Ship as-is only when the worst corrupted MOS is at least 4.0 and speaker similarity is at least 0.80 in every condition. For clones trained on a single clean reference session, that combination is rare, so most candidates fail this gate.

If clean MOS is 4.2 or higher but corrupted MOS falls to 3.5–3.9, retrain with augmentation: add noise-, reverb-, and bandwidth-corrupted versions of the reference audio to training. This is the cheapest fix because it keeps the original session and changes only the training distribution.

If the source reference is below about 20 dB SNR, re-record it instead. Before assuming denoising can rescue a noisy capture, run an A/B check: denoise the reference, rescore speaker similarity against the 0.80 target, and re-record only if the cleaned audio still fails the input gate.

Apply the checks in this order—source SNR, then worst corrupted MOS, then similarity—so a low-SNR reference is caught before you spend a retraining cycle.

| Scenario | Winner | Reason |
| --- | --- | --- |
| Worst corrupted MOS ≥ 4.0 and similarity ≥ 0.80 in all three conditions | Ship as-is | Passes both gates; uncommon for single-session clones |
| Clean MOS 4.2+, corrupted MOS 3.5–3.9 | Retrain with augmentation | Cheapest fix; corrupted reference variants added to training |
| Source reference below ~20 dB SNR | Re-record reference | Low-SNR capture cannot be repaired by denoising |

![Voice cloning robustness test, photo 2](https://static.mm-ais.com/article-images-pixabay/voice-cloning-robustness-test-mean-opini-b9b5c69a.jpg)

## Numbers that decide the call

The decision thresholds below are method thresholds from this site's testing, not external benchmarks. Use them as pass/fail gates: every row must pass before shipping, and a single failure is enough to retrain or re-record.

In cross-lingual use, the clone may retain a high MOS because raters still hear a plausible voice, while segmental or prosodic mismatches make the target text hard to understand. That is a failure of intelligibility rather than naturalness, so it must be scored on its own rubric rather than inferred from the same-language MOS floor.

For rater uncertainty, treat 3.9–4.1 as a decision band rather than a pass/fail line. If the worst corrupted-condition score falls in that band, expand the panel and recompute the worst case before applying the ship-or-retrain rule; a single small panel should not settle a borderline call.

These checks delimit the corruption tests rather than replace them. A clone that passes noise, reverberation, and bandwidth loss in one language with one panel has evidence for that condition only. Release documentation should state the tested language, rater-selection method, and whether cross-lingual output was evaluated.

![berry red mean snowball fruit](https://static.mm-ais.com/article-images-pixabay/voice-cloning-robustness-test-mean-opini-9bbc9393.jpg)
berry red mean snowball fruit

## And shipping checklist

Worked example: start with a 30-second reference recorded at 22 dB SNR, synthesize the clone on 10 test sentences, and score MOS plus speaker similarity on the clean output and on each of the three degraded outputs.

Checkpoint 1, clean output: MOS 4.3 and speaker similarity 0.84. That clears the clean floor of 4.2 and the 0.80 similarity target, so the clone looks releasable if evaluation stops here.

Checkpoint 2, 10 dB babble noise: MOS 3.7 and speaker similarity 0.74. Both fall below the 4.0 MOS floor and 0.80 similarity target, so this single condition is a fail even though the clean score passed.

Checkpoint 3, worst-case gate: the lowest MOS among the three degraded conditions is 3.7, below 4.0. The rule is therefore retrain with corrupted-condition augmentation and do not ship; the 4.3 clean MOS does not override the failed degraded condition.

Use the checklist as the release artifact: log every per-condition MOS and similarity, compute the worst degraded MOS before any ship decision, and rerun all three conditions after augmentation rather than only the one that failed.

The checklist separates measurement from decision: rows 1-4 record evidence and row 5 applies the rule. If any row fails, the action is fixed—retrain with augmentation and rerun all conditions—so the ticket does not depend on ad hoc judgment.

Keep the reference details in the ticket too: 30 seconds at 22 dB SNR is the condition under which this example was scored, and changing the reference later requires rescoring rather than reusing these numbers.

| Release check | Pass criterion | Example result |
| --- | --- | --- |
| Reference logged | 30 s at 22 dB SNR | Pass |
| Clean output | MOS ≥ 4.2 and similarity ≥ 0.80 | 4.3 / 0.84, pass |
| Each degraded output | MOS ≥ 4.0 and similarity ≥ 0.80 | 3.7 / 0.74 on 10 dB babble, fail |
| Worst degraded MOS | ≥ 4.0 | 3.7, fail |
| Decision | Ship only if all rows pass | Retrain with augmentation; do not ship |

## Frequently Asked Questions

**What happens if any corrupted condition scores below 4.0 MOS?**

Any corrupted condition below 4.0 means retrain with corrupted-condition augmentation.

**Under which three degradations must speaker similarity reach at least 0.80?**

Speaker similarity of at least 0.80 is required under noise, reverberation, and bandwidth loss.

**How many degraded acoustic conditions must simultaneously meet both thresholds?**

The clone must hold MOS ≥ 4.0 and speaker similarity ≥ 0.80 across at least three degraded acoustic conditions.

**What evidence shows robustness should be measured rather than assumed?**

RobustBench tracks over 3,000 adversarial-robustness papers, and the available sources cite more than 3,000 papers on adversarial robustness, underscoring that robustness must be measured, not assumed.

**What does the cited broad corruption benchmark illustrate for voice testing?**

One cited benchmark covers 7 datasets with 96 visual and 87 textual corruptions, illustrating broad corruption testing, while voice tests should cover noise, reverberation, and bandwidth loss.

**What kind of rule does this guide define for voice clones?**

This guide defines a ship-or-retrain rule for voice clones using clean and corrupted acoustic tests.

## Quick answers

| When should a 4.2-MOS voice clone be shipped? | Ship a 4.2-MOS voice clone only if its worst-case MOS across three corruption tests is at least 4.0. |
| --- | --- |
| Is a clean-set MOS of 4.2 sufficient for shipping? | Clean-set MOS of 4.2 is insufficient; any corrupted condition below 4.0 means retrain with corrupted-condition augmentation. |
| What speaker similarity is required under noise, reverberation, and bandwidth loss? | Require speaker similarity of at least 0.80 under noise, reverberation, and bandwidth loss. |
| What must the clone hold across at least three degraded acoustic conditions? | The clone must hold MOS ≥ 4.0 and speaker similarity ≥ 0.80 across at least three degraded acoustic conditions. |
| What benchmark-scale evidence supports measuring robustness? | RobustBench tracks over 3,000 adversarial-robustness papers. |

Also worth reading: **How to scale your brand using professional AI voice cloning technology**: [How to scale your brand](https://clonemyvoice.io/blog/how-to-scale-your-brand-using-professional-ai-voice-cloning-technology.php) · **Exploring voice cloning effects on audio file fidelity**: [Exploring voice cloning effects on](https://clonemyvoice.io/blog/exploring_voice_cloning_effects_on_audio_file_fidelity.php) · **Exploring the use of voice cloning in animated storytelling**: [Exploring the use of voice](https://clonemyvoice.io/blog/exploring_the_use_of_voice_cloning_in_animated_storytelling.php)

### Related reading

- [Stream Voice Cloning Audio: C# `yield` at 24 kHz, Frame by Frame](https://clonemyvoice.io/blog/stream-voice-cloning-audio-c-yield-at-24-khz-frame-by-frame.php)
- [Why organizations fail voice cloning: 3-second sample vs 2026 adoption](https://clonemyvoice.io/blog/why-organizations-fail-voice-cloning-3-second-sample-vs-2026-adoption.php)
- [Voice cloning in Indian noise: 2026 benchmarks favor 15 dB over re-record](https://clonemyvoice.io/blog/voice-cloning-in-indian-noise-2026-benchmarks-favor-15-db-over-re-record.php)
- [Imani Winds “Rise” Voice-Cloning Review: How to Verify Three Key Tests](https://clonemyvoice.io/blog/imani-winds-rise-voice-cloning-review-how-to-verify-three-key-tests.php)
- [Voice Cloning Models Compared: 150 ms at Chunk Boundary, Not Yet a User Experience Metric](https://clonemyvoice.io/blog/voice-cloning-models-compared-150-ms-at-chunk-boundary-not-yet-a-user-experience-metric.php)
- [Voice cloning enrollment time: 3 sec preview vs 30 sec identity clone](https://clonemyvoice.io/blog/voice-cloning-enrollment-time-3-sec-preview-vs-30-sec-identity-clone.php)

### Latest

- [Stream Voice Cloning Audio: C# `yield` at 24 kHz, Frame by Frame](https://clonemyvoice.io/blog/stream-voice-cloning-audio-c-yield-at-24-khz-frame-by-frame.php)
- [Voice cloning sample length: a 30-second default and how to test 60s](https://clonemyvoice.io/blog/voice-cloning-sample-length-a-30-second-default-and-how-to-test-60s.php)
- [Why organizations fail voice cloning: 3-second sample vs 2026 adoption](https://clonemyvoice.io/blog/why-organizations-fail-voice-cloning-3-second-sample-vs-2026-adoption.php)

Canonical: https://clonemyvoice.io/blog/voice-cloning-robustness-test-mean-opinion-score-mos-42-floor-ship-or-retrain.php
Markdown: https://clonemyvoice.io/blog/voice-cloning-robustness-test-mean-opinion-score-mos-42-floor-ship-or-retrain.php/index.md