Voice cloning sample length: a 30-second default and how to test 60s

This guide explains how voice-cloning sample length affects MOS and speaker similarity.

It sets a 30-second default and a 10-second held-out test for any extension to 60 seconds.

How MOS and similarity are actually measured

Mean Opinion Score (MOS) is a subjective 1–5 listener rating, where 1 is bad and 5 is excellent, averaged across many raters. In 2026 voice-cloning papers it is usually split into MOS-N for naturalness and MOS-S for similarity, and results are often reported with 95% confidence intervals. Because it is an average of human judgments, MOS measures the perceived quality of the rendered clone, not the amount or length of source audio used to make it.

Speaker similarity is usually reported as a cosine score between embeddings from a speaker-verification model such as ECAPA-TDNN or a WavLM-based model; higher means more similar. Its role in cloning evaluation is to provide an objective-ish check on identity preservation, but it is model-dependent: an ECAPA-TDNN cosine and a WavLM cosine are not on the same scale, so scores should not be compared directly across papers or systems. It also does not measure naturalness, intelligibility, or whether the clone sounds human.

Sample length enters the measurement only as a condition of the experiment. Raters and embedding models score the generated utterance, not the source recording, so a longer sample does not earn credit for being longer; it can only help if the extra audio changes the clone's output in a way listeners or the verifier can detect. That is why a held-out test should compare rendered clones from different sample lengths under the same prompt, text, and evaluation setup rather than comparing raw recording durations.

Use MOS and similarity as complementary checks: MOS catches unnatural prosody or artifacts that an embedding may miss, while similarity catches identity drift that listeners may not articulate. Report both with their model and rater details, and treat a gain as meaningful only when it exceeds the stated confidence interval or the verifier's normal score variation. If a longer sample does not move either metric on identical test utterances, the extra audio is not adding useful coverage.

When comparing studies, align the metric definitions before comparing numbers: confirm whether MOS is MOS-N or MOS-S, whether similarity is cosine or another distance, and which verification model produced it. A paper's "similarity" label is not interchangeable with another paper's unless the embedding model and scoring rule match.

The evidence: what the numbers actually show

Uberduck reportedly claims over 95% similarity to the original voice for instant cloning with short reference samples; as noted later in this article, that figure is vendor-reported and not independently replicated in any source available here. Treat it as a check, not proof: run your own 10-second held-out test to verify whether a short sample reaches high similarity under your recording conditions.

KikiVoice reportedly claims 99% accuracy in same-language cloning tests with short samples. Like the Uberduck figure, this is a vendor-reported claim, not a controlled 10s vs 30s vs 60s experiment, so read it only as a weak signal that same-language cloning may saturate before 60 seconds — verify with your own held-out comparison before assuming extra audio is useless.

Use 15 dB SNR as the working quality threshold. The claim that a clean 30-second sample is where perceptual quality stops climbing is a working hypothesis in this guide — no controlled 10s vs 30s vs 60s MOS curve appears in the available sources, as the next section acknowledges. Check it on your own voice by comparing clones from different lengths on the same 10-second held-out probe before treating 30 seconds as your ceiling. Below the SNR threshold, extra seconds can preserve noise as if it were speaker detail.

; after that, the question is whether added audio contributes new phonetic or prosodic coverage rather than more seconds of the same material. Below that threshold, extra seconds can preserve noise as if it were speaker detail.

A useful check is to compare a clone against a 10-second held-out segment from the same speaker. If MOS or similarity does not move measurably, the longer sample has not added information. The reported Uberduck and KikiVoice figures are best read as saturation evidence, not as a reason to keep recording.

Short samples can be enough because instant cloning systems extract a compact speaker representation. Once the reference captures stable timbre and speaking style, additional seconds mostly repeat information. The 99% same-language figure and the over-95% similarity figure both point to that plateau.

By contrast, a 60-second sample helps only when it adds coverage the shorter sample lacked, such as rare phonemes, varied intonation, or a different speaking rate. If the extra half minute is more of the same clean speech, the clone should not improve in a measurable way.

Options compared: 10s vs 30s vs 60s

Run the choice as a coverage test, not a length contest: record 30 seconds of clean speech at 15 dB+ SNR as the default cloning sample, then extend to 60 seconds only when a 10-second held-out test shows a measurable MOS or speaker-similarity gain. If that held-out slice does not improve, keep the shorter take; never accept a clone that fails the check just because the enrollment file is longer.

Ten seconds can be enough for same-language, high-SNR cloning, but it is the fragile option: it leaves little room for accent drift, background rise, or a noisy room, and a take below the 10 dB reject floor should be discarded rather than “cleaned up.” Thirty seconds is the winner because it usually captures enough phonetic coverage for stable prosody while staying short enough to record cleanly in one careful pass.

The practical rule is to compare like with like: same speaker, same microphone distance, same language, same speaking style, and the same held-out 10-second probe. If 60 seconds beats 30 seconds on that probe, the added minute is doing real work; if it only ties, the longer file has increased the chance of a bad breath, a door slam, or a room-tone shift entering the clone.

So the decision tree is simple: start at 30 seconds, reject anything under 10 dB, and treat 60 seconds as an exception that must earn its keep. The goal is not the longest sample; it is the shortest clean sample that reaches the perceptual ceiling and stays there under a realistic held-out test.

Costs and numbers that matter

Check the cloning service's actual entry price on its pricing page before budgeting, since no verified fee appears in the sources here. The dominant cost is recording time rather than the clone fee: a longer take that does not change the held-out result wastes more studio time than a shorter one. Treat the clone fee as a test cost, not the production budget, and compare lengths only after a clean baseline exists.

Recording cost scales linearly with duration: a 10-second take uses roughly one-third of the studio time of a 30-second take, while a 60-second take uses roughly twice as much, including setup, retakes, and editing. Because the marginal MOS gain from 30 to 60 seconds is typically near zero, the extra half-minute usually buys no perceptual return unless it adds new phonetic or prosodic coverage.

The expensive failure is not the first clone; it is re-recording plus re-cloning after a bad sample. A 10 dB SNR reject is a cost control because it stops the workflow before that spend is incurred. Check the take for noise before cloning, not after the clone is delivered, and keep the rejected take out of the enrollment set.

Use 30 seconds of clean, 15 dB+ SNR speech as the default, then run a 10-second held-out test before considering 60 seconds. Extend only if that test shows a measurable MOS or similarity gain; otherwise keep the shorter, cheaper baseline and avoid paying for unused duration.

If a longer sample does not improve the held-out score, revert to 30 seconds and retain the raw take for a re-clone. That keeps experimentation bounded while preserving the option to add coverage later without repeating the entire recording session.

For budgeting, compare cost per usable clone: if 60 seconds does not improve the held-out score, the additional studio time is pure overhead. The 30-second default keeps that overhead at the lowest level supported by the evidence, while the $5 entry point keeps the comparison itself affordable.

What the evidence does NOT establish

The Uberduck >95% and KikiVoice 99% similarity figures are vendor-reported and, in the available sources available here, not independently replicated; treat them as upper bounds on what a clean short sample can yield, not as guarantees for your recording conditions. If a vendor demo uses studio audio, your home recording should be judged against your own held-out test, not against the published percentage.

No source in the available sources provides a controlled 10s vs 30s vs 60s MOS curve, so the 30-second default is a synthesis of SNR thresholds and vendor similarity claims rather than one measured experiment. That means you should treat 30 seconds as a working default and verify it on your target voice, language, and microphone instead of treating it as an experimentally proven optimum.

The 30-second rule breaks when the reference must cover phonemes or prosodic patterns absent from a shorter sample, as in cross-lingual or cross-accent cloning; in those cases, extend toward 60 seconds only after confirming the added audio introduces new coverage rather than more of the same. A useful check is to compare phoneme or word coverage between the 30-second and 60-second clips before spending time on a longer enrollment.

A high vendor similarity score does not establish that your own 30-second clip will reach the same ceiling; recording quality, SNR, and phonetic variety still gate the result, so a clean 15 dB+ SNR sample with varied content is the safer interpretation of the rule. If your clip is repetitive or noisy, the published upper bound is not the right expectation.

Before extending to 60 seconds, require a measurable gain on a 10-second held-out test; if MOS or similarity does not move, the extra audio is redundant, and if it drops, suspect added noise, repetition, or mismatched style rather than insufficient length. Record the baseline first, then change only sample length so the comparison is interpretable.

In short, the evidence supports a practical default and a test for exceptions, not a universal law: 30 seconds is where clean speech usually stops helping, and 60 seconds is justified only by demonstrated new coverage. That framing also keeps vendor percentages in their proper place: useful for setting expectations, not for replacing your own measurement.

30s at 15 dB SNR

Use this as the default recording and verification checklist: one speaker, the same language as the target, a quiet room, one microphone, and a target of 30 seconds of continuous speech at 15 dB SNR or better. Treat 30 seconds as the enrollment budget, not a minimum to exceed, and do not extend the sample unless a later held-out test shows a measurable gain.

Record continuous speech rather than isolated words, and stop once you have 30 seconds of usable audio. Keep microphone distance, gain, and speaking level consistent across the take so the noise floor does not shift mid-sample.

Before cloning, measure SNR with a noise-floor tool. Reject the take below 10 dB. If it measures from 10 dB to 15 dB, re-record after reducing noise or moving closer to the microphone. Accept the take only at 15 dB or higher; if you cannot reach that level, change rooms or shorten the source distance instead of accepting a noisier recording.

Reserve 10 seconds as a held-out test and clone from the remaining 20 seconds. Compute the speaker-embedding cosine between the clone and the held-out reference, and require 0.85 or higher, or the vendor's equivalent "high similarity" label, before proceeding. If the score is below that threshold, inspect the recording for clipping, room echo, or inconsistent level before adding more audio.

Document the SNR value, held-out cosine, microphone setup, and sample length with the recording. That record makes the next cloning attempt reproducible and gives a clear reason to stop at 30 seconds or justify a longer take.

Worked Example: Run the Numbers

Illustration only: Party A, an illustrative creator, records a 30-second clean sample at 15 dB or better SNR for a narration clone. Day 0 is the baseline recording; Day 1 is a 10-second held-out synthesis and score; Day 2 is an optional 60-second re-record, tested only if Day 1 shows no gain.

Step 1, time budget: the baseline is 30 seconds. A 60-second sample is 60 ÷ 30 = 2.0 times as long, so the extra audio adds 30 seconds, or 100% more recording time. If that extra audio repeats phonemes and prosody already present in the 30-second clip, it adds no new coverage and the added time is redundant.

Step 2, held-out comparison: synthesize the same 10-second test clip from both the 30-second and 60-second references, then compare their scores. Illustration: if the 30-second clone scores 4.0 and the 60-second clone scores 4.0, the gain is 4.0 − 4.0 = 0.0 points. A gain of 0.0 points means the longer sample did not improve this example.

Step 3, coverage audit: list the phonemes and prosodic patterns covered by the 30-second clip, then check whether the 60-second clip adds at least one item not on that list. If it adds none, the 60-second clip is only a longer repeat and should not be expected to raise the score.

Winner and break-even: for this example, the 30-second sample wins because it reaches the same score with half the recording time. The 60-second sample breaks even only when the held-out 10-second test shows a measurable gain; otherwise the 30-second sample remains the default.

Decision rules

Apply these rules in order before any cloning attempt. First measure SNR, then confirm language match, then run the held-out similarity check; stop at the first failed gate and do not let a later step override an earlier one. The order matters because quality and coverage decide whether a clone proceeds, while duration is only a conditional extension.

If SNR is below 10 dB, reject the sample entirely and do not clone from it. Treat that as a hard stop: no extension to 30 or 60 seconds repairs a noisy reference, and no post-processing exception is implied by these rules. Discard the take and capture a cleaner one.

If SNR is at least 10 dB but below 15 dB, re-record before cloning. Do not compensate for that gap by extending the sample to 60 seconds; longer audio with the same noise floor does not convert a marginal reference into a usable one. The required action is a better recording, not a longer one.

If the sample and target are the same language and SNR is at least 15 dB, use 30 seconds as the default. Run a 10-second held-out similarity test: if it scores at or above 0.85 and shows no measurable gain from a 60-second reference, skip 60 seconds. Extend only when the test shows a measurable MOS or similarity gain and the extra audio adds new phonetic or prosodic coverage, not merely the same material.

Run the held-out test on audio that was not used in enrollment, and compare its result with the 0.85 cutoff before deciding to extend. If the score is at or above 0.85, keep the 30-second sample and stop. If it is below 0.85, extend only after confirming that the added 30 seconds contains new phonetic or prosodic material.

Record the SNR, language match, chosen duration, and held-out score with each clone attempt. That log makes the rule auditable and prevents a failed sample from being reused under a different project name. If any field is missing, treat the attempt as not yet approved.

What to do next

Frequently Asked Questions

What SNR threshold is specified for the default 30-second cloning sample?

Record 30 seconds at 15 dB+ SNR as the default cloning sample.

When is extending a cloning sample to 60 seconds allowed?

Use 60 seconds only if a 10-second held-out test shows a measurable MOS or similarity gain and the extra audio adds new phonetic or prosodic coverage.

What is the rule about accepting a clone without measured gain?

Never accept a clone without that gain.

How is Mean Opinion Score defined and scaled?

Mean Opinion Score (MOS) is a subjective 1–5 listener rating, where 1 is bad and 5 is excellent, averaged across many raters.

Does MOS reflect how much source audio was used?

Because it is an average of human judgments, MOS measures the perceived quality of the rendered clone, not the amount or length of source audio used to make it.

How is speaker similarity usually reported?

Speaker similarity is usually reported as a cosine score between embeddings from a speaker-verification model such as ECAPA-TDNN or a WavLM-based model.

Quick answers

Also worth reading: Voice cloning audio quality: 5 dB vs 20 dB denoise or re-record: Voice cloning audio quality: 5 · Exploring voice cloning effects on audio file fidelity: Exploring voice cloning effects on · Voice cloning sample length: 5s vs 30s test, reject below 10 dB 2026: Voice cloning sample length: 5s