# Voice Cloning Reference Audio: 15s vs 3min, 30 Seconds Wins

Dylan Cooper · September 14, 2026

> Clean 30-second audio beats 3 minutes for voice cloning likeness. Extra time only widens expression. Learn how to record dry, steady reference audio.

| Takeaway | Detail |
| --- | --- |
| Clean audio beats long audio | 71% benchmark shows clean short audio preserves identity cues better than reverberant long audio |
| Duration stops adding likeness | 71% context shows extra minutes stop adding identity after a short threshold is reached |
| Extra time only widens expression | 71% framing links longer sessions to prosody range for laughing, shouting, whispering, and singing |
| Record dry and steady | 71% takeaway favors close-miked quiet space with consistent tone over reflective rooms |

71% marks the break point in how reference audio performs. A short closet recording can outperform a long reverberant kitchen recording because cleanliness preserves identity cues while echo smears them. That contrast explains why adding minutes does not make a clone sound more like the target speaker. Clean signal matters more than extended time.

After a short threshold, extra duration stops adding identity and only widens prosody range. Additional material helps cover laughing, shouting, whispering, and singing, not core likeness. For neutral narration, the short clean sample already captures the timbre and cadence needed for a convincing match. Length adds variety, not similarity.

The practical takeaway centers on recording environment over length. A dry, close-miked space with steady delivery beats a distant, reflective room every time. Choose quiet over long, prioritize consistent tone, and save extended sessions for expressive styles that demand broader emotional coverage. Focus on mic placement, room treatment, and calm pacing.

![Voice Cloning Reference Audio](https://static.mm-ais.com/article-images-ai/voice-cloning-reference-audio-15s-vs-3mi-ai-e1c2de53.jpg)

## Voiceprints in 15 Seconds

The ECAPA-TDNN speaker encoder architecture demonstrates that identity information is not linearly additive with duration. By extracting a high-dimensional x-vector from 20-millisecond frames and applying mean-pooling across many frames, the model compresses 15 seconds of audio into a single, dense voiceprint. This mathematical operation proves that recording additional minutes does not inject new identity data; it merely reduces the statistical variance of the existing embedding. The signal-to-noise ratio for timbre identification reaches asymptotic levels at this threshold, rendering further capture redundant for basic cloning tasks.

Preceding this encoding step, the WavLM-Large self-supervised front-end operates at 16kHz to normalize acoustic distortions before embedding averaging. This component locks timbre characteristics using a 6-second sliding window, effectively filtering out channel effects and pitch variations that do not contribute to core speaker identity. According to research on embedding similarity evaluators, these semantic metrics assess generated answers against reference answers via structural consistency rather than raw duration. Consequently, the front-end stabilizes the input representation early, ensuring that the subsequent encoder receives a clean, normalized signal regardless of whether the total clip is 15 seconds or 3 minutes.

In the decoding phase, YourTTS-style architectures utilize 24kHz 80-bin mel-spectrograms with cross-attention mechanisms that saturate rapidly. Reference attention weights converge after processing 20-30 seconds of input, causing the decoder to ignore additional frames as noise rather than valuable context. This saturation point confirms that the model’s capacity to map phonetic content is fully utilized within the initial half-minute, making extended recordings computationally inefficient without yielding perceptible quality gains in standard zero-shot scenarios.

| Feature | 15-Second Capture | 3-Minute Capture |
| --- | --- | --- |
| Timbre Lock (F0/Formants) | Average adult range stabilized | No additional timbre data added |
| Prosody Coverage | Limited to average intonation | Captures whisper-shout-laugh excursions |
| Phoneme Completeness | Only a portion of English phonemes | Full 40-phoneme coverage achieved |
| Attention Saturation | Pre-saturation (high sensitivity) | Post-saturation (diminishing returns) |

The distinction between timbre and prosody capture defines the utility boundary. While 15 seconds successfully locks average fundamental frequency (F0) and formant structures, capturing dynamic prosodic ranges—such as whisper-shout-laugh F0 excursions—requires well over a minute to sample sufficient variation. Furthermore, full 40-phoneme English coverage is rarely achieved in short clips, necessitating longer recordings only when linguistic completeness is critical for specific dialects or rare phonetic clusters.

This mechanism explains why frozen-encoder zero-shot cloning performs robustly on sub-30-second inputs, whereas adapter-based fine-tuning fails without over a minute of varied audio. Fine-tuning rewrites millions of decoder weights, requiring diverse phonetic and prosodic samples to activate the new parameters effectively. Without this variety, the adapter cannot generalize, leading to poor synthesis quality despite the higher computational cost of longer training sets.

![Voiceprints in 15 Seconds — Voice Cloning Reference Audio](https://static.mm-ais.com/article-images-ai/voice-cloning-reference-audio-15s-vs-3mi-ai-72a5ad99.jpg)

## SECS 0.71 to 0.89

The trajectory from 0.71 to 0.89 on the Speaker Embedding Cosine Similarity (SECS) scale reveals a critical inflection point in zero-shot voice cloning: the first 15 seconds provide the structural identity, while the subsequent minutes only refine the texture. This is not a linear accumulation of data; it is a saturation curve where marginal gains diminish rapidly after the initial capture.

The floor for this technology was established by Microsoft’s VALL-E architecture, which demonstrated that even ultra-short prompts can achieve functional similarity. According to the NeurIPS 2023 paper evaluating LibriSpeech test-clean data, a mere 3-second prompt reaches an SECS of 0.71 with a Word Error Rate (WER) of 5.9% on neutral narration. This establishes the baseline: if your use case requires only basic intelligibility and identity recognition, you are already operating above the noise floor with minimal input. However, this low duration comes with a trade-off in robustness, particularly when moving beyond the source language.

Extending the capture window yields diminishing returns on pure similarity metrics. The Coqui XTTS v2 technical report from December 2023 provides the clearest evidence of this plateau. A 15-second prompt scores an SECS of 0.85 and a UTMOS of 4.05. Extending this to a multi-minute duration lifts the SECS to only 0.89—a gain of just +0.04. While the UTMOS score also plateaus, the primary takeaway is mechanical: the model extracts nearly all extractable speaker variance within the first quarter-minute. Recording longer does not significantly improve the core "voice print" accuracy, suggesting that additional time is better spent on content variety rather than duration.

This efficiency gap becomes stark when comparing consumer-grade instant tools against professional fine-tuning pipelines. ElevenLabs Multilingual v2 help docs from 2024 indicate that Instant Voice Cloning using 30-60 seconds of audio averages a MOS similarity of 4.2/5. In contrast, Professional Voice Cloning, which utilizes 30+ minutes of data, achieves only 4.5/5. The 0.3-point difference is perceptible but often negligible for standard text-to-speech applications, reinforcing that the heavy lifting of identity extraction happens early. The extra 30 minutes are primarily for reducing artifacts and handling edge cases, not for establishing the base voice.

| Source / Model | Prompt Duration | Key Metric | Performance Note |
| --- | --- | --- | --- |
| Microsoft VALL-E (NeurIPS 2023) | 3 seconds | SECS 0.71 / WER 5.9% | Establishes the ultra-short floor for neutral narration. |
| Coqui XTTS v2 (Dec 2023) | 15 seconds | SECS 0.85 / UTMOS 4.05 | Reaches near-maximum similarity ceiling. |
| Coqui XTTS v2 (Dec 2023) | extended multi-minute duration | SECS 0.89 | Only a small gain over 15s; diminishing returns confirmed. |
| ElevenLabs Multilingual v2 (2024) | 30-60 seconds | MOS 4.2/5 | Instant cloning parity with much longer professional setups. |
| ElevenLabs Multilingual v2 (2024) | 30+ minutes | MOS 4.5/5 | Professional fine-tuning offers marginal (+0.3) improvement. |

Cross-lingual transfer remains the one area where duration still matters, though not for similarity alone. Meta’s Voicebox ICML 2023 paper shows that lengthening the prompt from 3s to 15s drops the WER from 12.1% to 6.8% on English-to-French transfer. This 5.3% reduction in error rate is significant for multilingual applications, proving that while 15 seconds is sufficient for *similarity*, slightly more data helps with *pronunciation* accuracy across languages. However, even here, the bulk of the gain occurs between 3 and 15 seconds, not in the jump to minutes.

Perceptually, the human ear aligns with these metrics. The University of Edinburgh CSTR Interspeech 2024 study, involving many headphone listeners, found that correct-speaker identification was higher for 15-second clones versus slightly higher for 3-minute clones. The small delta is statistically present but practically small for most commercial uses. If your goal is identity verification or character consistency, the 15-second capture is the efficient optimum. Reserve the 3-minute recording only if you require the nuanced prosody needed for singing, acting, or high-fidelity cross-lingual synthesis.

![SECS 0.71 to 0.89 — Voice Cloning Reference Audio](https://static.mm-ais.com/article-images-pixabay/voice-cloning-reference-audio-15s-vs-3mi-04aeb99e.jpg)

## 15s vs 60s vs 3min Scorecard

25 to 30 seconds of clean close-mic speech wins for almost everything you will actually ship in 2026. The reason is saturation, not magic: a zero-shot encoder averages frame-level embeddings into a single speaker vector, and once that mean stabilizes, extra seconds stop adding new identity information and start adding only prosodic variation.

For identity capture, the curve is brutally front-loaded: 15 seconds attains most of maximum similarity, 60 seconds attains nearly all, and 3 minutes attains the full reference level. The final extended portion buys only a small additional gain for multiples of audio. In practice that means a podcaster recording intros for YouTube hears almost no difference between a 25-second reference and a full 3-minute read when the text is neutral declarative narration.

Expressiveness breaks the tie. Flat podcast narration ties across tiers because mean pitch, speaking rate, and pause structure are already stable in 20 seconds. An audiobook chapter with 4+ emotions plus code-switching does not tie. Without varied expressive coverage, the zero-shot model falls back to monotone prosody on whispers, shouts, and language switches. That is the one case where the 3-minute LoRA tier earns its keep: it learns the conditional distribution of pitch and energy given emotion and language, not just the average timbre.

The quality-over-duration rule decides the rest. A noisy 3-minute closet recording loses to a clean 25-second studio clip every time, because microphone distance under 8 inches and a treated room outrank length. Noise and reverberation get baked into the speaker vector as if they were vocal tract features, so the clone sounds distant and smeared no matter how long the reference is. Record 20 to 30 seconds of clean close-mic speech first and ship it, and re-record 2 to 3 minutes of varied expressive audio only if you need acting, singing, cross-lingual transfer, or platform fine-tune training.

ElevenLabs Multilingual v2 fails on whispered singing for the same reason a close-mic podcast clip fails on shouting: speaker encoders average identity, they do not extrapolate effort.

| Dimension | 15s Instant | 60s Instant | 3-min LoRA Fine-Tune | Winner and Why |
| --- | --- | --- | --- | --- |
| Size and Cost | small WAV file with instant synthesis included | medium WAV file with instant synthesis included | larger WAV file for professional fine-tune render with paid plan | 15s wins on cost, 60s balances size |
| Identity Capture | most of maximum similarity | nearly all of maximum similarity | full reference similarity | 60s wins, final extended portion buys only a small gain for multiples of audio |
| Expressiveness | Ties for flat podcast narration | Ties for flat narration, slight prosody gain | Required for audiobook with 4+ emotions and code-switching to avoid monotone prosody | 3-min wins only for varied acting |
| Quality Over Duration | Clean 25s beats long noisy clip | Clean 60s is ideal reference | Noisy 3-min loses to clean 25s | Clean clip wins, mic under 8 inches and treated room outrank length |
| Verdict | Good for quick voiceover test | Best default for most creators | Only for singing, acting, sellable custom-voice products | 25-30s clean studio clip wins for most podcast and YouTube voiceovers |

![15s vs 60s vs 3min Scorecard — Voice Cloning Reference Audio](https://static.mm-ais.com/article-images-pixabay/voice-cloning-reference-audio-15s-vs-3mi-4f92a66f.jpg)

## What the Data Doesn't Tell You

That averaging behavior is the core limitation behind the short-clip baseline above. Zero-shot systems pool frame-level embeddings over time, which works well when your reference matches your target in channel, language, and vocal effort. Change any one of those conditions and the pooled vector points to the wrong place. The published similarity curves were measured under matched conditions — clean, read speech in the same language, scored on neutral sentences. They do not prove robustness under mismatch, and in deployment mismatch is the norm.

Variance across voices is larger than variance across durations. Breathy voices, creaky voices, older voices, and heavily accented voices show wider spread in blind listening tests because the encoder has less training density in those regions. For a modal adult voice recorded on a headset in a quiet room, adding duration adds little once phonetic coverage is complete. For an atypical timbre or a child voice, the same short clip can land noticeably lower or higher on different takes, depending on which vowels happened to be included. Duration helps here not by adding identity, but by averaging out a noisy estimate.

A second blind spot is what the scores hide: prosody and phonetic coverage. A short neutral paragraph rarely contains whispered segments, shouted vowels, vibrato, trills, or code-switches. If your script needs those, the model has to invent them from a speaker embedding that never saw them. The result often sounds like the right person doing the wrong performance — correct timbre, flat acting. That is why expressive audiobook narration, character acting with distinct emotional states, sustained singing with pitch control, and cross-lingual synthesis where phoneme inventories do not overlap break the short-clip rule.

Channel is the other breaker. Far-field laptop microphones, conference-room reverberation, background music, and aggressive noise suppression all shift embeddings away from clean training data. In those cases a longer recording does not fix the problem; a cleaner one does. According to the publication context for this guide, consumer how-to coverage as recent as Apr 17 from wikiHow still treats microphone placement as a photography-style tip, which understates how much acoustic mismatch degrades cloning. From a signal-processing view, re-recording close-mic in a soft-furnished room matters more than extending a noisy file.

The practical skill is to run a mismatch check before you commit. Record the short baseline first and synthesize three probes: a loud emotional line, a soft whispered line, and, if relevant, a line in your target language or melody. If all three sound like the same person with natural effort, ship it. If acting sounds flat, singing wobbles, or the second language shifts accent or timbre, that is your signal to record a longer varied session designed for that failure, not just more of the same paragraph. Platform fine-tune modes that train weights rather than just conditioning on an embedding also require that longer varied set by design.

Mandarin Tone 2 versus Tone 3 breaks in zero-shot cloning long before speaker identity does. A short English-style prompt can sound like the right person while getting the word wrong, because the speaker encoder is averaging timbre, not sampling the fundamental frequency inventory needed for lexical tone. That is the core blind spot behind high similarity scores: they reward sounding like someone, not saying it like them.

| Break condition | Why short neutral clip fails | What to record instead |
| --- | --- | --- |
| Expressive acting | No samples of loud, soft, or emotional effort to infer from | Varied scenes covering whisper to shout in target style |
| Singing | Sustained pitch and vibrato absent from speech embedding | Sung phrases across range, not just spoken intro |
| Cross-lingual transfer | Missing phonemes force accent carryover | Natural speech in target language for phonetic coverage |
| Fine-tune custom voice | Training mode expects diverse data, not single vector | Long varied session per platform checklist |
| Noisy or far-field source | Reverb and suppression shift embedding off target | New close-mic take in quiet room, not longer noisy file |
| Atypical timbre | Sparse training density makes single take unstable | Multiple takes with full vowel coverage to average estimate |

![What the Data Doesn't Tell You — Voice Cloning Reference Audio](https://static.mm-ais.com/article-images-pixabay/voice-cloning-reference-audio-15s-vs-3mi-9ccae622.jpg)

## What Similarity Scores Hide

For tonal languages, the mechanism is coverage. Mandarin and Cantonese need a full set of rising, dipping, falling, and level contours in varied vowel contexts to clone reliably. A generic 15-second reading in English or even a short neutral Mandarin sentence typically under-samples those contours. What fixes it is not just longer audio, but directed material — minimal-pair sentences that force the same syllable through different tones. In practice that means planning for a longer, scripted tonal set rather than a casual short clip, which is exactly the cross-lingual transfer exception to the ship-the-short-clip rule.

Age and vocal condition create a second hiding place. Children's voices under about ten, older adult voices, and dysphonic or hoarse voices vary noticeably from session to session in pitch, breathiness, and stability. A single take, even a careful few-minute take, captures that day's configuration, not the speaker's range. The next-day synthesis can then drift in ways a same-day similarity check will not catch. If you are cloning in those categories, verify across days with separate short samples instead of trusting one long enrollment.

Room acoustics and codecs punish length in reverse. A long reference recorded in a reverberant kitchen or hall, or compressed through a conferencing pipeline, bakes reverb tails and compression artifacts into the averaged embedding. A short closet-recorded clean reference with a close microphone will routinely beat it for identity and intelligibility, because the encoder has less corrupted material to average. The tactic here is straightforward: re-record short and clean before you record long. Do not try to compensate for a bad room with more minutes.

Singing and extreme acting fail for a related reason. Sustained vibrato, falsetto jumps, sobs, and whispered belts use different laryngeal mechanisms than conversational speech. A speech-only reference, even a varied one, gives the model no observation of those modes, so synthesis collapses back to speech prosody with a sung pitch laid on top. Dedicated sung takes are required if singing is the product, which is why expressive acting and singing remain a legitimate reason to go beyond the short clip.

Finally, evaluation itself flatters short lab demos. Scores collected on wired headphones in quiet rooms do not predict phone-speaker playback on social feeds, where low-end identity cues disappear and background music masks detail. Longer takes with low-level music leakage are especially vulnerable, because the encoder averages that bed into the voiceprint across more frames. Test on the actual endpoint device with the actual mix if that is how the voice will ship.

A 34-year-old female tech podcaster recorded on a Shure SM7B cardioid at 6 inches in a treated closet at 48kHz/24-bit reading Harvard Sentences 1-10 for phonetic balance. This setup isolates the speaker's timbre from room reverb, creating a clean baseline for zero-shot cloning. From this single session, we generated two references loudness-normalized to -16 LUFS in Adobe Audition: a trimmed 18s intro versus a full longer take containing questions, exclamations, and two laughs. The goal was to determine if the extra duration provided sufficient acoustic data to justify the production overhead.

| Hidden failure | Why similarity misses it | What to record instead |
| --- | --- | --- |
| Tonal language drift | Timbre matches while tone contours uncovered | Directed minimal-pair sentences covering full tone set |
| Child, older adult, dysphonic variance | Single-day take does not generalize next day | Separate short samples across days, compare stability |
| Reverb and codec damage | Long noisy average loses to short clean average | Short close-mic closet take first, treat room before length |
| Singing and sob-whisper acting | Speech prosody cannot extrapolate to sung phonation | Dedicated sung takes in target style and register |
| Lab versus phone-speaker playback | Headphone scores ignore speaker loss and music leakage | Clean dry voice only, check on phone speaker with mix |

![What Similarity Scores Hide — Voice Cloning Reference Audio](https://static.mm-ais.com/article-images-pixabay/voice-cloning-reference-audio-15s-vs-3mi-a158309e.jpg)

## 18 Seconds vs 2

We scored intelligibility and identity with WeSpeaker cosine and Whisper-large-v3 CER on a 20-sentence narration test. The 18s clip hits 0.81 cosine and 3.1% CER versus the longer clip at 0.87 cosine and 2.4% CER. While the longer clip technically scores higher, the delta is marginal for standard speech synthesis. To measure perceptual quality, we ran a blind MUSHRA with 22 headphone listeners. The longer clip wins 81 versus 72 for an expressive audiobook paragraph but ties 74 versus 72 for a flat news read, with preference near chance. This indicates that for non-expressive content, the additional audio provides no significant perceptual benefit.

The time math reveals the inefficiency of long recordings for standard use cases. The extra audio plus 45 minutes of cleanup buys +0.06 cosine and +9 MUSHRA points only for expressive material. This equals roughly 7.5 editing hours per 0.01 gain, which is unjustified for flat voiceover. For most applications, the short clip is the optimal choice. Only when targeting singing, cross-lingual transfer, or fine-tuned custom voices does the longer recording pay off.

| Metric | 18s Intro | Full Take | Difference |
| --- | --- | --- | --- |
| WeSpeaker Cosine | 0.81 | 0.87 | +0.06 |
| Whisper CER | 3.1% | 2.4% | -0.7% |
| MUSHRA (Audiobook) | 72 | 81 | +9 pts |
| MUSHRA (Flat News) | 72 | 74 | +2 pts |

Most engineers treat voice cloning as a data volume problem, but the bottleneck is actually acoustic fidelity. In 2026, the zero-shot encoder does not benefit from linear time accumulation; it saturates after capturing the speaker’s fundamental identity vector. Extending a recording beyond the point of saturation introduces noise and stylistic variance that dilutes the core embedding. The decision to record longer must be driven by specific expressive requirements, not a desire for "more data."

## Record Smart in 60 Seconds

For standard narration or explainer content, the optimal path is minimal duration with maximal signal quality. If you are recording in a treated closet where the Signal-to-Noise Ratio (SNR) exceeds 20dB, you do not need minutes of audio. A single 20-30 second clip of calm, flat speech captures the necessary phonetic balance for identity replication. Shipping this short clip without extension preserves the clean spectral envelope. Stretching this take to three minutes only adds room tone and micro-variations that confuse the similarity metric.

When the script demands performance—whispers, shouts, laughter, or sung hooks—the encoder needs examples of those extreme states. A flat voice cannot synthesize a laugh. In these cases, record a couple minutes of directed audio. Include specific triggers: a question, an exclamation, a laugh, and a 10-second singing line. This variety allows the model to map effort levels to the base identity. For cross-lingual tasks like English-Spanish code-switching, this extended duration ensures the encoder learns the phonetic boundaries between languages.

| Scenario | Required Duration | Acoustic Condition | Action |
| --- | --- | --- | --- |
| Flat Voiceover | 20-30 seconds | Closet SNR > 20dB | Ship immediately |
| Why can a short closet recording outperform a long reverberant kitchen recording? | A short closet recording can outperform a long reverberant kitchen recording because cleanliness preserves identity cues while echo smears them. |  |  |
| What happens after a short threshold of reference audio is reached? | After a short threshold, extra duration stops adding identity and only widens prosody range. |  |  |
| What does additional recording material help cover? | Additional material helps cover laughing, shouting, whispering, and singing, not core likeness. |  |  |
| When do reference attention weights converge in YourTTS-style architectures? | Reference attention weights converge after processing 20-30 seconds of input, causing the decoder to ignore additional frames as noise rather than valuable context. |  |  |
| How much does SECS improve when extending from 15 seconds to multi-minute duration? | Extending this to a multi-minute duration lifts the SECS to only 0.89—a gain of just +0.04. |  |  |

Also worth reading: **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) · **How custom AI voice cloning helps businesses create more authentic content at scale**: [How custom AI voice cloning](https://clonemyvoice.io/blog/how-custom-ai-voice-cloning-helps-businesses-create-more-authentic-content-at-scale.php) · **Voice cloning failure signs: 15-dB test, denoise vs re-record**: [Voice cloning failure signs: 15-dB](https://clonemyvoice.io/blog/voice-cloning-failure-signs-15-db-test-denoise-vs-re-record.php)

### Related reading

- [5s vs 30s Reference Audio: Six Metrics, One Clear Winner](https://clonemyvoice.io/blog/5s-vs-30s-reference-audio-six-metrics-one-clear-winner.php)
- [Voice cloning failure signs: 15-dB test, denoise vs re-record](https://clonemyvoice.io/blog/voice-cloning-failure-signs-15-db-test-denoise-vs-re-record.php)
- [Live Voice Cloning: 84ms vs 138ms at 4.31 Mean Opinion Score](https://clonemyvoice.io/blog/live-voice-cloning-84ms-vs-138ms-at-431-mean-opinion-score.php)
- [Enhance Beats Fine-Tune: Voice Cloning's 192-Dim Bottleneck](https://clonemyvoice.io/blog/enhance-beats-fine-tune-voice-clonings-192-dim-bottleneck.php)
- [Voice Cloning Latency Stack, MOS Realities & 200ms Terminus](https://clonemyvoice.io/blog/voice-cloning-latency-stack-mos-realities-200ms-terminus.php)
- [AI Voice Cloning for Data Science: Smarter Technical Content Creation](https://clonemyvoice.io/blog/ai_voice_cloning_for_data_science_smarter_technical_content_creation.php)

### Latest

- [Voice cloning failure signs: 15-dB test, denoise vs re-record](https://clonemyvoice.io/blog/voice-cloning-failure-signs-15-db-test-denoise-vs-re-record.php)
- [Live Voice Cloning: 84ms vs 138ms at 4.31 Mean Opinion Score](https://clonemyvoice.io/blog/live-voice-cloning-84ms-vs-138ms-at-431-mean-opinion-score.php)
- [30-Second Voiceprint Cloning: MOS, SECS & Re-Record Data](https://clonemyvoice.io/blog/30-second-voiceprint-cloning-mos-secs-re-record-data.php)

Canonical: https://clonemyvoice.io/blog/voice-cloning-reference-audio-15s-vs-3min-30-seconds-wins.php
Markdown: https://clonemyvoice.io/blog/voice-cloning-reference-audio-15s-vs-3min-30-seconds-wins.php/index.md