Three seconds of speech from a stranger is a standard test condition for zero-shot voice cloning research, the setup behind both VALL-E and Voicebox. On that test, Voicebox raised speaker similarity to 0.681 from VALL-E’s 0.580 on LibriSpeech, and NaturalSpeech 3 later reported 0.67 against 0.68 for the real recordings on its own run of the same dataset. Listeners are not much of a backstop: in a 2025 Scientific Reports study, people judged AI clones to be the real speaker about 80% of the time. For the commercial and market picture, see our broader voice cloning statistics; this roundup stays on one question, how much reference audio cloning needs and what each extra second buys. We aggregated data from Microsoft Research, Meta, ByteDance, the Barcelona Supercomputing Center, the U.S. Federal Trade Commission, Consumer Reports, McAfee, Starling Bank, and peer-reviewed and arXiv papers from ICML, NeurIPS, ICLR and Scientific Reports.
TL;DR
- Voicebox was tested on 3.0 seconds of prompt audio and matched VALL-E’s speaker similarity with roughly two thirds of the input audio (Meta, Voicebox paper, NeurIPS 2023).
- On VCTK, VALL-E 2 similarity rose from 0.447 at 3 seconds to 0.558 at 10 seconds, while word error rate rose from 0.9% to 3.3% (Microsoft Research, VALL-E 2 paper 2024).
- Cutting the average prompt from 7.7 to 5.5 seconds dropped F5-TTS similarity from 0.35 to 0.24 on in-the-wild speech (Barcelona Supercomputing Center, WildSpoof submission 2026).
- XTTS-v2 reached high-fidelity cloning with 6 to 10 seconds of reference audio, with diminishing returns beyond 20 seconds (Zhu, University of Groningen thesis 2025).
- A zero-shot model with a 300-second prompt scored 0.932 similarity vs 0.934 for a fine-tuned baseline on 300 seconds (Jiang et al., Mega-TTS 2, ICLR 2024).
- Fine-tuning on 20 seconds of Portuguese speech raised human similarity ratings from 2.77 to 4.43 (Casanova et al., YourTTS, ICML 2022).
- Listeners judged AI voices to be the same person as the real speaker about 80% of the time (Barrington, Cooper and Farid, Scientific Reports 2025).
- McAfee researchers reported an estimated 85% voice match from three to four seconds of audio, a vendor measurement (McAfee, Beware the Artificial Impostor 2023).
- 28% of UK adults say they were targeted by an AI voice cloning scam in the past year (Starling Bank and Mortar Research survey 2024).
- Four of six voice cloning products failed to take basic steps to stop unauthorized cloning (Consumer Reports, voice cloning assessment 2025).
1. The 3-Second Test: How Much Audio Zero-Shot Systems Are Given
“Zero-shot” means the system was never trained on the target speaker: it hears a short reference clip, called the “prompt”, at generation time and imitates it on the spot. VALL-E and Voicebox were both evaluated on a prompt of just 3 seconds, and that choice says more about ambition than about limits. These papers do not ask whether a voice can be copied from a few seconds. They compete on how well.
Product requirements run longer. OpenAI described its Voice Engine as working from “a single 15-second audio sample” (OpenAI post on synthetic voices 2024), and ElevenLabs documentation recommends one to two minutes. Those are product settings, not measured thresholds.
| Metric | Value | Source |
|---|---|---|
| Prompt length and training data for VALL-E on unseen LibriSpeech speakers | 3 seconds; 60K hours of English speech, over 7,000 speakers | Microsoft Research, VALL-E paper 2023 |
| Prompt audio used for Voicebox English zero-shot tests | 3.0 seconds | Meta, Voicebox paper (NeurIPS 2023) |
| Reference clip length used to evaluate XTTS | One segment between 3 and 8 seconds | Casanova et al., XTTS paper (INTERSPEECH 2024) |
| Reference utterance length range in the Seed-TTS test set | 3 to 20 seconds | ByteDance, Seed-TTS paper 2024 |
| Audio sample OpenAI Voice Engine works from | A single 15-second sample | OpenAI, synthetic voices post 2024 |
| Minimum and recommended audio for ElevenLabs instant cloning | At least 10 seconds in total, 1 to 2 minutes recommended (vendor’s own guidance) | ElevenLabs, Instant Voice Cloning documentation 2026 |
If you are comparing desktop voice cloning software, sections 2 and 3 show what those seconds change.
2. How Speaker Similarity Scales From 3 to 10 Seconds
Speaker similarity (SIM, or SECS) is the cosine similarity between two speaker “fingerprints” computed by a speaker-recognition network, where higher means closer. Two real recordings of the same person score well below 1.0, so the real-speech figures below are the reference point to read against. Going from a 3-second to a 10-second prompt lifted VALL-E 2’s similarity on the VCTK dataset from 0.447 to 0.558, a gain of 0.111 (0.558 minus 0.447), which is the clearest published curve for this question (VALL-E 2 paper 2024).
The curve bends at different points in different systems. Meta’s Voicebox team reported that similarity “grows quickly and flattens with longer audio prompts” (Voicebox paper), and a 2025 thesis on XTTS-v2 put the plateau at around 10 to 20 seconds (Zhu, University of Groningen). Across these three papers, the first ten seconds do most of the work, though VALL-E 2 was still improving at ten.
| Metric | Value | Source |
|---|---|---|
| VALL-E similarity on VCTK (108 speakers) with 3, 5 and 10-second prompts | 0.382, 0.423, 0.484 (real recordings: 0.546, 0.591, 0.620) | Microsoft Research, VALL-E paper 2023 |
| VALL-E 2 (group size 1) similarity on VCTK with 3, 5 and 10-second prompts | Single sample: 0.447, 0.487, 0.558; best of five samples: 0.508, 0.552, 0.620 (real recordings: 0.623, 0.679, 0.709) | Microsoft Research, VALL-E 2 paper 2024 |
| Share of the input audio Voicebox needs to equal VALL-E speaker similarity | Roughly two thirds | Meta, Voicebox paper (NeurIPS 2023) |
| F5-TTS similarity with long (7.7 s average) vs short (5.5 s average) prompts on in-the-wild speech | 0.35 vs 0.24 | Barcelona Supercomputing Center, WildSpoof TTS submission 2026 |
| StyleTTS2 similarity with long (7.7 s average) vs short (5.5 s average) prompts | 0.19 vs 0.14 | Barcelona Supercomputing Center, WildSpoof TTS submission 2026 |
| Reference length at which XTTS-v2 reached high-fidelity cloning (1, 3, 6, 10, 20 and 40 seconds tested) | 6 to 10 seconds, with diminishing returns beyond 20 seconds | Zhu, University of Groningen master’s thesis 2025 |
Outlier note: the Barcelona Supercomputing Center submission is the only 2026 result here and uses the TTS In the Wild dataset. A prompt 2.2 seconds shorter on average (7.7 minus 5.5) cost F5-TTS 0.11 of similarity (0.35 minus 0.24). That is the same size of loss VALL-E 2 showed across a 7-second gap on VCTK, a hint that short prompts may hurt more on everyday audio, though the two tests are not directly comparable.
3. Intelligibility Does Not Always Improve With Longer Prompts
Word error rate (WER) measures intelligibility: a speech recognizer transcribes the cloned audio and the share of wrong words is counted, so lower is better. VALL-E 2’s WER on VCTK rose from 0.9% with a 3-second prompt to 3.3% with a 10-second prompt, the opposite direction from its similarity gain. Human raters agreed: in the same test, similarity ratings fell as prompts lengthened, while ratings for the real recordings rose.
Nothing guarantees that more prompt audio helps. Mega-TTS 2, designed for multi-sentence prompts, kept improving all the way to 300 seconds, while the VALL-E run measured in the same paper broke down at 20 seconds (Mega-TTS 2 paper). Very short prompts carry their own intelligibility risk: StyleTTS2 was the system that got worse when its prompt was cut.
| Metric | Value | Source |
|---|---|---|
| VALL-E 2 (group size 1) on VCTK with 3, 5 and 10-second prompts: word error rate, and human similarity rating (SMOS, higher is better) | WER 0.9%, 1.9%, 3.3%; SMOS 4.42±0.15, 4.28±0.16, 3.95±0.10 (real recordings 4.47±0.13, 4.53±0.14, 4.74±0.17) | Microsoft Research, VALL-E 2 paper 2024 |
| VALL-E on LibriSpeech with 3, 10 and 20-second prompts, as measured by the Mega-TTS 2 authors | WER 5.83%, 5.54%, 8.77%; similarity 0.885, 0.893, 0.805 | Jiang et al., Mega-TTS 2 (ICLR 2024) |
| Mega-TTS 2 word error rate with 3, 10, 60 and 300-second prompts | 2.46%, 2.28%, 2.24%, 2.23% | Jiang et al., Mega-TTS 2 (ICLR 2024) |
| StyleTTS2 word error rate with long (7.7 s) vs short (5.5 s) prompts | 0.21 vs 0.49 (F5-TTS unchanged at 0.07) | Barcelona Supercomputing Center, WildSpoof TTS submission 2026 |
| Voicebox trend when the prompt and target language differ, as prompts lengthen | Similarity improves; WER rises for most language directions | Meta, Voicebox paper (NeurIPS 2023) |
Context: the VALL-E 2 trend is not perfectly smooth. Its group size 2 setting recorded WER of 1.5%, 0.9% and 2.3% at the same three lengths. Voicebox’s authors attribute the rising cross-language WER to English making up over 90% of the training data.
4. How Close Short-Prompt Clones Get to the Real Speaker
On benchmark speech, the best published systems now sit next to the real recordings. NaturalSpeech 3 scored a speaker similarity of 0.67 against 0.68 for the ground-truth recordings, and its human similarity rating (4.01) was higher than the real recordings’ 3.85 on the same LibriSpeech test (Microsoft Research, NaturalSpeech 3 paper 2024). That is a long way from the 0.337 YourTTS scored in the first row below.
The real-speech scores are the tell that these are lab conditions: they run from 0.68 to 0.8952 across papers, so each system has to be read against its own ground truth, as in the table. The F5-TTS paper lands within 0.03 of real speech on its own test (0.69 minus 0.66).
| Metric | Value | Source |
|---|---|---|
| LibriSpeech, 3-second prompt: YourTTS vs VALL-E vs Voicebox | Similarity 0.337 vs 0.580 vs 0.681 (real speech 0.754); WER 7.7% vs 5.9% vs 1.9% (real speech 2.2%) | Meta, Voicebox paper (NeurIPS 2023) |
| NaturalSpeech 3 vs real recordings on LibriSpeech: similarity and WER | 0.67 and 1.81% vs 0.68 and 1.94% | Microsoft Research, NaturalSpeech 3 paper 2024 |
| Seed-TTS English similarity vs human recordings, and comparative rating (CMOS) vs human | 0.762 vs 0.730; CMOS -0.07 | ByteDance, Seed-TTS paper 2024 |
| F5-TTS vs real speech on 1,127 LibriSpeech-PC samples of 4 to 10 seconds: similarity and WER | 0.66 and 2.42% vs 0.69 and 2.23% | Chen et al., F5-TTS paper 2024 |
| MiniMax-Speech English similarity: speaker embedding alone vs prompt with paired text | 0.692 vs 0.738 (ground truth 0.730) | Zhang et al., MiniMax-Speech paper 2025 |
| XTTS English similarity vs real speech | 0.6423 vs 0.8952 | Casanova et al., XTTS paper (INTERSPEECH 2024) |
Outlier note: Seed-TTS and the paired-text MiniMax-Speech setting score above their real recordings. The Seed-TTS authors explain that ground-truth and reference clips of one speaker can still differ in speaking style and background environment, so a model that follows the reference closely can out-score a second real recording. It is not evidence of a better-than-human clone.
5. Zero-Shot vs Fine-Tuned Cloning: The Gap Measured in Seconds and Minutes
Fine-tuning means updating a model’s weights on the target speaker’s audio, the older and slower route to a custom voice. With 300 seconds of the target voice, a zero-shot prompt reached 0.932 similarity against 0.934 for a fine-tuned baseline, with fewer word errors (2.23% vs 3.11%) (Jiang et al., Mega-TTS 2, ICLR 2024). The caveat matters: both systems are the authors’ own, and the baseline is a FastSpeech 2 build fine-tuned for 2,000 steps, not the strongest fine-tuned system available.
Fine-tuning still earns its cost when a voice sits far from the training data. YourTTS was trained on a single Portuguese speaker, and fine-tuning on 20 to 31 seconds lifted listeners’ similarity ratings for new Portuguese voices, though overall quality ratings dipped (YourTTS paper). XTTS gained 0.1314 of similarity (0.7166 minus 0.5852) from about 10 minutes per speaker in cross-language cloning.
| Metric | Value | Source |
|---|---|---|
| Similarity of the fine-tuned baseline with 10, 60 and 300 seconds of target audio | 0.895, 0.923, 0.934 | Jiang et al., Mega-TTS 2 (ICLR 2024) |
| Similarity of zero-shot Mega-TTS 2 with 3, 10, 60 and 300-second prompts | 0.898, 0.905, 0.926, 0.932 | Jiang et al., Mega-TTS 2 (ICLR 2024) |
| YourTTS, Portuguese female speaker with 20 seconds of speech, zero-shot vs fine-tuned | Similarity 0.743 vs 0.930; Sim-MOS 2.77±0.15 vs 4.43±0.06; MOS 3.59±0.13 vs 3.48±0.13 | Casanova et al., YourTTS (ICML 2022) |
| YourTTS, English male speaker with 61 seconds of speech, zero-shot vs fine-tuned | Similarity 0.851 vs 0.880 (real speech 0.875) | Casanova et al., YourTTS (ICML 2022) |
| XTTS cross-language similarity before vs after fine-tuning on about 10 minutes per speaker | 0.5852 vs 0.7166 | Casanova et al., XTTS paper (INTERSPEECH 2024) |
| Seed-TTS with a 20-second prompt vs fine-tuned on 1 to 10 hours per speaker: similarity and WER | 0.779 and 3.15 vs 0.789 and 2.83 (CMOS +0.37 for fine-tuned) | ByteDance, Seed-TTS paper 2024 |
| Audio recommended for ElevenLabs Professional Voice Cloning (its fine-tuned tier) | 30 to 180 minutes (vendor’s own guidance) | ElevenLabs, voice cloning documentation 2026 |
Most recent available data for the YourTTS fine-tuning test: Casanova et al., ICML 2022. It covers only 4 speakers, so treat it as an illustration of the language effect rather than a population estimate.
6. The Security Side: Short Prompts and Fraud
The research numbers have a direct consumer translation: public video is prompt material. Listeners judged an AI voice to be the same identity as its real counterpart about 80% of the time, and identified AI voices only about 60% of the time, across 604 participants in a study whose real clips averaged 5.37 seconds (Barrington, Cooper and Farid, Scientific Reports 2025). The clones came from a commercial instant-cloning service fed the first two recorded prompts of each of 220 speakers.
The “three seconds” line in fraud warnings rests on company statements. McAfee’s figure is its own benchmark and, for that source, the most recent available data (McAfee, 2023). Starling Bank repeated the claim in 2024 alongside its survey (Starling Bank, Safe Phrases campaign 2024). Our post on voice cloning family fraud covers the scams themselves.
| Metric | Value | Source |
|---|---|---|
| Share of the time listeners judged an AI voice to be the same person as the real speaker | About 80% | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Share of the time listeners correctly identified a voice as AI generated | About 60% | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Audio McAfee researchers needed for an estimated 85% voice match with a free tool | Three to four seconds (about 95% with more effort; vendor’s own benchmark) | McAfee, Beware the Artificial Impostor 2023 |
| Adults who share their voice online at least once a week | 53% (7,054 adults, 7 countries) | McAfee, Beware the Artificial Impostor 2023 |
| Adults not confident they could tell a real voice from a cloned one | 70% | McAfee, Beware the Artificial Impostor 2023 |
| UK adults who say they were targeted by an AI voice cloning scam in the past year | 28% (3,010 adults) | Starling Bank and Mortar Research survey 2024 |
| UK adults who would send the money requested even if the call seemed strange | 8% | Starling Bank and Mortar Research survey 2024 |
| Voice cloning products assessed that failed to take basic steps to stop unauthorized cloning | 4 of 6 | Consumer Reports, voice cloning assessment 2025 |
Context: the Starling 28% measures what people say happened to them, not confirmed incidents. McAfee’s 85% and 95% are the company’s own estimates, and we found no independent replication. The Consumer Reports assessment covered Descript, ElevenLabs, Lovo, PlayHT, Resemble AI and Speechify.
7. Safeguards, Detection and Release Decisions
Release decisions split the field. Microsoft says it has “no plans to incorporate VALL-E 2 into a product or expand access to the public” (VALL-E 2 project page), and OpenAI chose to “preview but not widely release” Voice Engine while urging banks to phase out voice-based authentication. Meanwhile F5-TTS’s authors call for watermarks and detection models for the misuse risks they name. The FTC treated the problem as open enough to fund a challenge (FTC, Voice Cloning Challenge winners 2024).
Detection results depend heavily on what the fake is compared against. Meta’s classifier separated Voicebox output from original audio perfectly, but scored 0.704 against real speech that had been passed through a vocoder (software that rebuilds audio from its spectrum) at the lowest masking level. For a deeper look at detector performance, see our audio deepfake detection accuracy statistics.
| Metric | Value | Source |
|---|---|---|
| Prize money split by the three FTC Voice Cloning Challenge winners | $35,000 (announced April 8, 2024) | U.S. Federal Trade Commission, Voice Cloning Challenge 2024 |
| Audio chunk size checked by the FTC-recognized Pindrop real-time clone detector | Two-second chunks | U.S. Federal Trade Commission, Voice Cloning Challenge 2024 |
| Voicebox detection classifier accuracy vs original audio, and vs vocoded real audio at 30%, 50% and 90% masking of the audio | 1.000 vs original; 0.704, 0.809, 0.907 vs vocoded | Meta, Voicebox paper (NeurIPS 2023) |
| When OpenAI first developed Voice Engine vs when it published its limited preview | Late 2022 vs March 29, 2024 | OpenAI, synthetic voices post 2024 |
| Teams in the ASVspoof 5 challenge, and the hardest attack in the Track 1 open condition | 53 teams; pre-trained zero-shot YourTTS (minDCF 0.33) | Wang et al., ASVspoof 5 evaluation paper 2026 |
Context: minDCF is a detection cost score where higher means detectors struggled more. A pre-trained YourTTS model topping the ASVspoof 5 open-condition difficulty list suggests zero-shot attacks do not need a frontier system to cause trouble, though the paper does not break this result down by prompt length.
Summary: Zero-Shot Voice Cloning Reference Audio Length by the Numbers
| Metric | Value | Source |
|---|---|---|
| Audio sample OpenAI Voice Engine works from | 15 seconds | OpenAI, synthetic voices post 2024 |
| VALL-E 2 on VCTK, 3 vs 10-second prompt (group size 1): similarity, and WER | 0.447 vs 0.558; 0.9% vs 3.3% | Microsoft Research, VALL-E 2 paper 2024 |
| Share of the input audio Voicebox needs to equal VALL-E similarity | Roughly two thirds | Meta, Voicebox paper (NeurIPS 2023) |
| F5-TTS similarity, long (7.7 s) vs short (5.5 s) prompts on in-the-wild speech | 0.35 vs 0.24 | Barcelona Supercomputing Center, WildSpoof TTS submission 2026 |
| StyleTTS2 word error rate, long (7.7 s) vs short (5.5 s) prompts | 0.21 vs 0.49 | Barcelona Supercomputing Center, WildSpoof TTS submission 2026 |
| Reference length for high-fidelity XTTS-v2 cloning | 6 to 10 seconds | Zhu, University of Groningen thesis 2025 |
| NaturalSpeech 3 similarity vs real recordings | 0.67 vs 0.68 | Microsoft Research, NaturalSpeech 3 paper 2024 |
| F5-TTS word error rate vs real speech | 2.42% vs 2.23% | Chen et al., F5-TTS paper 2024 |
| Zero-shot Mega-TTS 2 vs fine-tuned baseline similarity with 300 seconds of audio | 0.932 vs 0.934 | Jiang et al., Mega-TTS 2 (ICLR 2024) |
| YourTTS listener similarity rating, zero-shot vs fine-tuned on 20 seconds | 2.77 vs 4.43 | Casanova et al., YourTTS (ICML 2022) |
| XTTS cross-language similarity after about 10 minutes of fine-tuning | 0.5852 to 0.7166 | Casanova et al., XTTS paper (INTERSPEECH 2024) |
| Listeners judging an AI clone to be the real speaker | About 80% | Barrington, Cooper and Farid, Scientific Reports 2025 |
| Audio for an estimated 85% match (vendor benchmark) | Three to four seconds | McAfee, Beware the Artificial Impostor 2023 |
| UK adults who say they were targeted by a voice cloning scam | 28% | Starling Bank and Mortar Research survey 2024 |
| Voice cloning products that failed to take basic steps against unauthorized cloning | 4 of 6 | Consumer Reports, voice cloning assessment 2025 |
| Prize money for FTC Voice Cloning Challenge winners | $35,000 | U.S. Federal Trade Commission, Voice Cloning Challenge 2024 |
Methodology and Sources
- Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E) (Microsoft Research, 2023): 3-second prompt setup and VCTK results by prompt length.
- VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers (Microsoft Research, 2024): VCTK similarity, WER and SMOS at 3, 5 and 10 seconds.
- VALL-E 2 project page (Microsoft Research): release statement.
- NaturalSpeech 3: Zero-Shot Speech Synthesis with Factorized Codec and Diffusion Models (Microsoft Research and university co-authors, 2024): LibriSpeech comparison.
- Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale (Meta, NeurIPS 2023): prompt-length analysis, English benchmark, detection classifier.
- YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for everyone (Casanova et al., ICML 2022): zero-shot vs fine-tuned speaker adaptation.
- XTTS: a Massively Multilingual Zero-Shot Text-to-Speech Model (Casanova et al., INTERSPEECH 2024): reference length, English similarity, fine-tuning gain.
- Mega-TTS 2: Boosting Prompting Mechanisms for Zero-Shot Speech Synthesis (Jiang et al., ByteDance and Zhejiang University, ICLR 2024): 3 to 300-second prompts vs a fine-tuned baseline.
- Seed-TTS: A Family of High-Quality Versatile Speech Generation Models (ByteDance, 2024): zero-shot vs speaker fine-tuning.
- F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching (Chen et al., 2024): LibriSpeech-PC benchmark.
- MiniMax-Speech: Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder (Zhang et al., 2025): speaker-embedding vs paired-text cloning.
- Zero-Shot TTS With Enhanced Audio Prompts: BSC submission for the 2026 WildSpoof challenge TTS track (Barcelona Supercomputing Center, February 5, 2026): long vs short prompts.
- Zero-Shot Voice Cloning with Minimal Data: Impact of Reference Duration on Long-Form Speech Synthesis (Zhu, University of Groningen, 2025): 1 to 40-second reference test.
- People are poorly equipped to detect AI-powered voice clones (Barrington, Cooper and Farid, Scientific Reports, 2025).
- ASVspoof 5: Evaluation of Spoofing, Deepfake, and Adversarial Attack Detection Using Crowdsourced Speech (Wang et al., January 7, 2026).
- FTC Announces Winners of Voice Cloning Challenge (U.S. Federal Trade Commission, April 8, 2024).
- Do These 6 AI Voice Cloning Companies Do Enough to Prevent Misuse? (Consumer Reports, March 10, 2025).
- Beware the Artificial Impostor (McAfee, May 2023; vendor research).
- Starling Bank Safe Phrases campaign (Starling Bank and Mortar Research, September 18, 2024).
- OpenAI post on the challenges and opportunities of synthetic voices (OpenAI, March 29, 2024).
- ElevenLabs Instant Voice Cloning documentation (vendor guidance, read in October 2026).
- Data watch: similarity scores are not comparable across papers, because each uses its own speaker-recognition model, test set and scoring (real speech on LibriSpeech scores 0.754, 0.68 or 0.69 depending on the paper), so only compare within a row. The Mega-TTS 2 paper reports much higher similarity values than the others, with VALL-E at 0.885 there against 0.580 in the original paper, so its numbers should only be read against each other. Both sides of the Mega-TTS 2 fine-tuning comparison were built by its authors. The Voicebox text says similarity grows then flattens with longer prompts while its figure caption says it remains the same up to 10 seconds. MiniMax-Speech calls paired-text prompting ‘one-shot’ and classes Seed-TTS that way, even though Seed-TTS calls itself zero-shot. The WildSpoof models were fine-tuned on the TITW dataset before testing, so those results are not pure zero-shot baselines. VALL-E 2’s 10-second listener scores carry wide intervals (up to ±0.49). The Scientific Reports paper does not state how long the two prompts behind each clone were, so its 80% should not be read as a figure for any given prompt length. Zhu’s thesis abstract gives no similarity values, only the lengths and the direction of the effect. VALL-E, Voicebox, YourTTS and McAfee are older than three years and are flagged as most recent available data: we found no 2025 or 2026 paper that repeats the full 3, 5 and 10-second curve on a shared benchmark. The Starling Bank and McAfee surveys are self-reported.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the ICASSP 2027 proceedings are published.