Google’s Tacotron 2 scored a mean opinion score (MOS) of 4.53 against 4.58 for professionally recorded speech in December 2017 (Google, Tacotron 2 paper 2017), and since then the gap between synthetic and human speech has closed on almost every standard benchmark. By 2024, zero-shot systems such as VALL-E 2 and MaskGCT were scoring slightly above the ground-truth recordings they were compared with (Microsoft, VALL-E 2 2024; MaskGCT 2024), and in a 2025 listening study 58% of voice clones were mistaken for real human voices (Queen Mary University of London, PLOS ONE 2025). The harder question in 2026 is no longer whether machines can sound human, it is whether the MOS test can still measure the difference. We aggregated data from Google DeepMind, Microsoft Research, ByteDance, the Blizzard Challenge, the VoiceMOS Challenge, ITU-T, PLOS ONE, Artificial Analysis and Hugging Face; for the wider market picture see our text-to-speech statistics roundup.
TL;DR
- Tacotron 2 scored 4.53 MOS vs 4.58 for professional recordings (Google, Tacotron 2 paper 2017).
- WaveNet cut the gap to human speech by over 50%, scoring 4.21 vs 4.55 in US English (DeepMind, WaveNet 2016).
- NaturalSpeech scored -0.01 CMOS vs LJ Speech recordings, with no significant difference (Microsoft Research, NaturalSpeech 2022).
- VALL-E 2 scored +0.033 CMOS and 4.61 SMOS vs 4.13 for ground truth on LibriSpeech (Microsoft, VALL-E 2 2024).
- NaturalSpeech 3 was trained on 200,000 hours with 1 billion parameters (Microsoft, NaturalSpeech 3 2024).
- CosyVoice 3 scaled training data from 10,000 to 1 million hours across 9 languages (Alibaba, CosyVoice 3 2025).
- Human recordings scored anywhere from 4.30 to 4.58 MOS across four landmark papers (Google, DeepMind, Kakao, Microsoft papers 2016-2021).
- Two Blizzard Challenge 2023 systems were not statistically different from natural speech on MOS, a first for the challenge (Blizzard Challenge 2023 summary, Computer Speech and Language 2024).
- 58% of voice clones and 41% of generic AI voices were judged human (Lavan et al., PLOS ONE 2025).
- Listeners identified speech deepfakes correctly 73% of the time (Mai et al., PLOS ONE 2023).
- The VoiceMOS Challenge 2022 covered 187 TTS and voice conversion systems from 22 teams (VoiceMOS Challenge 2022).
- The top Speech Arena model scored 1321 Elo with a plus or minus 189 interval (Artificial Analysis, October 2026).
1. The Human Ceiling Is Not a Fixed Number
Every parity claim depends on a reference score for real speech, and that reference moves. Recorded human speech has scored as high as 4.58 and as low as 4.30 MOS in landmark TTS papers, a 0.28-point spread that is wider than the gap separating many competing systems in the same tables. The LJ Speech dataset alone produced 4.46 in the VITS paper and 4.30 in the FastSpeech 2 paper, because listener pools, instructions and test design differ from lab to lab.
The practical reading: a MOS of 4.4 means nothing in isolation. It only becomes meaningful next to the ground-truth score from the same listening test, which is why modern papers increasingly report CMOS (comparative MOS, a side-by-side score relative to a reference) instead of raw MOS. The VITS paper is a good illustration: it reports 4.46 for ground truth and 4.43 for its own system, a 0.03 gap inside the confidence intervals.
| Metric | Value | Source |
|---|---|---|
| Professionally recorded speech MOS, Tacotron 2 test | 4.58 | Google, Tacotron 2 paper 2017 |
| Human speech MOS, US English, WaveNet test | 4.55 | DeepMind, WaveNet 2016 |
| Human speech MOS, Mandarin, WaveNet test | 4.21 | DeepMind, WaveNet 2016 |
| LJ Speech ground truth MOS, VITS test | 4.46 (plus or minus 0.06) | Kim et al. (Kakao Enterprise), VITS 2021 |
| VCTK ground truth MOS, VITS test | 4.38 (plus or minus 0.07) | Kim et al., VITS 2021 |
| LJ Speech ground truth MOS, FastSpeech 2 test | 4.30 (plus or minus 0.07) | Microsoft and Zhejiang University, FastSpeech 2 2020 |
| Ground truth passed through a vocoder, FastSpeech 2 test | 3.92 | FastSpeech 2 2020 |
Context note: the vocoder row matters. Simply converting real recordings to mel spectrograms and back cost 0.38 MOS points in the FastSpeech 2 test, which shows how much of the historic gap came from waveform generation rather than from text modeling. Most of these papers are older than three years and are shown as most recent available data for their benchmarks.
2. From WaveNet to Parity: Single-Speaker TTS, 2016 to 2022
The single-speaker story took about six years. WaveNet reduced the gap between the best existing systems and human speech by over 50% in 2016, scoring 4.21 in US English against 3.86 for concatenative and 3.67 for parametric synthesis (DeepMind WaveNet announcement). Tacotron 2 then brought the gap down to 0.05 points, and by 2021 VITS matched ground truth on the multi-speaker VCTK set at 4.38.
The finish line came with Microsoft’s NaturalSpeech, which scored -0.01 CMOS against LJ Speech recordings at sentence level, with a Wilcoxon signed rank test showing no significant difference. The interesting shift is what came next: once parity was reached on clean read speech, the field stopped competing on single-speaker naturalness and moved to cloning unseen voices from seconds of audio.
| Metric | Value | Source |
|---|---|---|
| WaveNet MOS vs human, US English | 4.21 vs 4.55 | DeepMind, WaveNet 2016 |
| WaveNet MOS vs human, Mandarin | 4.08 vs 4.21 | DeepMind, WaveNet 2016 |
| Concatenative vs parametric MOS, US English (pre-WaveNet) | 3.86 vs 3.67 | DeepMind, WaveNet 2016 |
| Tacotron 2 MOS vs recordings | 4.53 vs 4.58 | Google, Tacotron 2 paper 2017 |
| VITS MOS vs ground truth, LJ Speech | 4.43 vs 4.46 | Kim et al., VITS 2021 |
| VITS MOS vs ground truth, VCTK multi-speaker | 4.38 vs 4.38 | Kim et al., VITS 2021 |
| Tacotron 2 + HiFi-GAN on VCTK, same test | 3.14 | Kim et al., VITS 2021 |
| NaturalSpeech CMOS vs LJ Speech recordings | -0.01 (p >> 0.05) | Microsoft Research, NaturalSpeech 2022 |
Speed improved alongside quality. VITS synthesized at 67.12 times real time, versus 27.48 times for a Glow-TTS plus HiFi-GAN pipeline, and FastSpeech 2 trained in 17.02 hours versus 53.12 for FastSpeech, a 3.12x speedup (FastSpeech 2 paper).
| Metric | Value | Source |
|---|---|---|
| VITS real-time factor | 67.12x | Kim et al., VITS 2021 |
| Glow-TTS + HiFi-GAN real-time factor | 27.48x | Kim et al., VITS 2021 |
| FastSpeech 2 MOS vs Tacotron 2 MOS (20 native raters) | 3.83 vs 3.70 | FastSpeech 2 2020 |
| FastSpeech 2 training time vs FastSpeech | 17.02 h vs 53.12 h (3.12x faster) | FastSpeech 2 2020 |
3. Zero-Shot Voice Cloning Reaches and Passes Ground Truth
Zero-shot TTS, which clones an unseen speaker from a short prompt, compressed the same journey into roughly two years. Microsoft’s VALL-E 2 scored +0.033 CMOS against ground truth on LibriSpeech, and its speaker similarity score (4.61) beat the real recordings (4.13) (VALL-E 2 paper, 20 native American English raters). The original VALL-E, released in January 2023 and trained on 60,000 hours of English, still trailed ground truth by 0.17 CMOS on LibriSpeech.
Scale is the common thread. NaturalSpeech 3 used 200,000 hours and 1 billion parameters, MaskGCT 100,000 hours of in-the-wild speech, and Alibaba’s CosyVoice 3 one million hours across 9 languages and 18 Chinese dialects. The question for 2026 buyers of voice cloning software is no longer whether a clone can match a reference recording in a lab test, it is how consistently it does so on noisy prompts, accents and expressive speech. Our voice cloning statistics cover adoption and misuse.
| Metric | Value | Source |
|---|---|---|
| VALL-E training data | 60,000 hours of English | Microsoft, VALL-E 2023 |
| VALL-E CMOS vs ground truth, LibriSpeech (ground truth +0.17) | -0.17 | Microsoft, VALL-E 2023 |
| VALL-E 2 CMOS vs ground truth, LibriSpeech | +0.033 | Microsoft, VALL-E 2 2024 |
| VALL-E 2 SMOS vs ground truth, LibriSpeech | 4.61 vs 4.13 | Microsoft, VALL-E 2 2024 |
| NaturalSpeech 3 CMOS (ground truth at +0.08) | 0.00 | Microsoft, NaturalSpeech 3 2024 |
| NaturalSpeech 3 SMOS vs ground truth | 4.01 vs 3.85 | Microsoft, NaturalSpeech 3 2024 |
| MaskGCT CMOS vs ground truth, LibriSpeech | +0.10 | MaskGCT 2024 |
| Seed-TTS CMOS vs human, English / Mandarin | -0.07 / -0.08 | ByteDance, Seed-TTS 2024 |
Objective metrics tell the same story. CosyVoice 3 posted a 1.45% word error rate on the Seed-TTS English test set versus 2.14% for the human recordings (CosyVoice 3 paper), meaning automatic speech recognition transcribes the synthetic audio more accurately than the real speech. That is a sign of very clean articulation, not necessarily of greater naturalness, since real speakers mumble, hesitate and vary.
| Metric | Value | Source |
|---|---|---|
| NaturalSpeech 3 training data and size | 200,000 hours, 1B parameters | Microsoft, NaturalSpeech 3 2024 |
| MaskGCT training data | 100,000 hours (50K English, 50K Chinese) | MaskGCT 2024 |
| CosyVoice 3 training data | 1 million hours, 9 languages, 18 Chinese dialects | Alibaba, CosyVoice 3 2025 |
| CosyVoice 3 WER vs human, Seed-TTS test-en | 1.45% vs 2.14% | Alibaba, CosyVoice 3 2025 |
| Seed-TTS DiT WER vs human, English | 1.733 vs 2.143 | ByteDance, Seed-TTS 2024 |
| F5-TTS CMOS vs ground truth, Seed-TTS test-en / test-zh | +0.31 / +0.21 | F5-TTS 2024 |
| F5-TTS WER vs ground truth, LibriSpeech-PC | 2.42 vs 2.23 | F5-TTS 2024 |
Context note: VALL-E 2’s own VCTK results show how fragile parity is. With 3-second prompts its SMOS matched ground truth (4.47 vs 4.47), but with 10-second prompts the ground truth scored 4.74 and VALL-E 2 between 3.95 and 4.26. CosyVoice 3 also reports that in Chinese all of its models still lag behind human speech, while in English the 1.5B model scores above it.
4. The Blizzard Challenge: Parity Under Independent Testing
Author-run tests have an obvious conflict of interest, which is why the Blizzard Challenge matters: organizers run one shared listening test across all entries. The 2023 edition, the eighteenth, drew 18 teams to its Hub task (51 hours of French audio) and 14 to its Spoke task (2 hours), all using neural approaches (The Blizzard Challenge 2023, ISCA). Microsoft’s MuLanTTS, labeled system F, scored 4.3 MOS against 4.4 for the natural recording on the Hub task and 4.5 against 4.5 on the Spoke task (MuLanTTS paper).
The organizers’ follow-up summary drew the sharper conclusion: two systems were rated not statistically different from natural speech on MOS, a first in the challenge’s history, yet in a finer MUSHRA test every system was judged significantly different from natural speech (Refining the evaluation of speech synthesis, Computer Speech and Language 2024). In other words, the 5-point MOS scale saturated before the voices actually became identical. The same summary found non-native listeners and non-experts gave higher scores than native listeners and speech experts.
| Metric | Value | Source |
|---|---|---|
| Blizzard 2023 Hub / Spoke teams | 18 / 14 | Perrotin et al., Blizzard Challenge 2023 |
| Hub / Spoke training data (French) | 51 hours / 2 hours | Perrotin et al., Blizzard Challenge 2023 |
| MuLanTTS MOS vs natural recording, Hub | 4.3 vs 4.4 | Microsoft, MuLanTTS 2023 |
| MuLanTTS MOS vs natural recording, Spoke | 4.5 vs 4.5 | Microsoft, MuLanTTS 2023 |
| Systems not statistically different from natural speech on MOS | 2 (first time in Blizzard) | Blizzard Challenge 2023 summary, CSL 2024 |
| Systems judged different from natural speech in MUSHRA | All | Blizzard Challenge 2023 summary, CSL 2024 |
| DelightfulTTS MOS / SMOS, Blizzard 2021 task SH1 | 4.17 / 4.35 | Microsoft, DelightfulTTS 2021 |
The 2025 edition moved the hard problem to low-resource languages. It asked teams to build voices for Bildts, a Dutch language variety with about 6,000 native speakers, from a 7-hour dataset, plus a zero-shot task with 10 seconds of reference speech (The Blizzard Challenge 2025, University of Groningen). Seven teams entered the main task and five the zero-shot task.
| Metric | Value | Source |
|---|---|---|
| Blizzard 2025 target language native speakers | About 6,000 (Bildts) | Do et al., Blizzard Challenge 2025 |
| Main task training data | 7 hours, single speaker | Do et al., Blizzard Challenge 2025 |
| Zero-shot task reference audio | 10 seconds per speaker | Do et al., Blizzard Challenge 2025 |
| Main task / zero-shot teams | 7 / 5 | Do et al., Blizzard Challenge 2025 |
5. Why MOS Numbers Disagree: Methods, Standards and Automatic Prediction
The MOS method was built for telephone lines, not neural voices. ITU-T P.800, approved in August 1996, still governs subjective transmission quality testing; P.800.1 (July 2016) standardizes MOS terminology; and P.808 (June 2018, revised June 2021) covers crowdsourced speech quality tests (ITU-T P.808). A review of Interspeech and SSW papers from 2021-2022 found that most authors do not report scale labels, increments or instructions to listeners (Kirkland et al., Stuck in the MOS pit, SSW 2023), and that quality MOS averages higher than naturalness MOS for the same natural speech.
The same team ran controlled listening tests with identical stimuli and found that changing only the scale increment or the rating instruction moved MOS for some systems. Benchmark choices also diverge: a 2025 position paper noted that VALL-E evaluates on 1,234 LibriSpeech utterances, NaturalSpeech 3 and MaskGCT on 40, and F5-TTS on 1,127, and that MaskGCT’s WER is 2.63% on the 40-utterance set but 4.22% on the 1,234-utterance set (Towards Responsible Evaluation for Text-to-Speech, 2025).
| Metric | Value | Source |
|---|---|---|
| ITU-T P.800 approval | August 1996 (in force) | ITU-T |
| ITU-T P.800.1 MOS terminology, latest version | July 2016 | ITU-T |
| ITU-T P.808 crowdsourced testing | June 2018, revised June 2021 | ITU-T |
| Papers reporting scale labels, increments or instructions | Minority (2021-2022 Interspeech and SSW) | Kirkland et al., SSW 2023 |
| LibriSpeech test-clean utterances used: VALL-E / NaturalSpeech 3 / F5-TTS | 1,234 / 40 / 1,127 | Yang et al., 2025 |
| MaskGCT WER on 40 vs 1,234 utterances | 2.63% vs 4.22% | Yang et al., 2025 |
| Speaker similarity with vs without prompt included | 0.905 vs 0.754 | Yang et al., 2025 |
| CosyVoice 3 MOS panel | 200 sentences, 10 native raters, 0.5-point steps | Alibaba, CosyVoice 3 2025 |
Because human tests are slow and inconsistent, the field built automatic MOS predictors. The first VoiceMOS Challenge in 2022 drew 22 teams to predict human ratings for 187 TTS and voice conversion systems spanning over a decade of research (VoiceMOS Challenge 2022), and fine-tuned self-supervised speech models proved most effective. Participation has since thinned to 10 teams in 2023 and 8 in 2024, whose first track focused on zoomed-in, high-quality synthesis where predictors struggle most.
| Metric | Value | Source |
|---|---|---|
| VoiceMOS 2022 teams / systems in main track | 22 / 187 | Huang et al., VoiceMOS Challenge 2022 |
| VoiceMOS 2023 teams | 10 from 7 countries | VoiceMOS Challenge 2023 (ASRU 2023) |
| VoiceMOS 2024 teams | 8 | Huang et al., VoiceMOS Challenge 2024 |
| MOS-Bench training / test sets | 8 / 17 | MOS-Bench 2024 (revised 2026) |
| NaturalSpeech 3 UTMOS vs ground truth | 4.30 vs 4.14 | Microsoft, NaturalSpeech 3 2024 |
Context note: automatic scores such as UTMOS now routinely rate synthetic speech above real recordings, as NaturalSpeech 3’s table shows, which is a reason to treat them as screening tools rather than as proof of naturalness.
6. Can Listeners Still Tell? Perception Studies
Parity on a rating scale is one thing; fooling a listener who is asked directly is another. In a 2025 Queen Mary University of London study, 58% of voice clones and 41% of generic AI voices were judged to be human, while real human voices were recognized as human 62% to 81% of the time depending on the experiment (Lavan et al., PLOS ONE 2025). The clones were built from about 4 minutes of recordings per speaker with commercially available tools. The authors titled the paper ‘realistic but not (yet) hyperrealistic’: clones did not beat real voices overall, but they were close enough that any single listener’s judgment is unreliable.
The social impressions are the less discussed finding. Generic AI voices were rated more dominant than human voices (58.1 vs 40.4 on a 0-100 scale) and slightly more trustworthy (57.9 vs 53.8). For anyone producing synthetic narration or using an AI voice generator, that means listeners respond to these voices as people, which raises the stakes of disclosure.
| Metric | Value | Source |
|---|---|---|
| Voice clones judged human | 58% | Lavan et al., PLOS ONE 2025 |
| Generic AI voices judged human | 41% | Lavan et al., PLOS ONE 2025 |
| Real human voices judged human | 62% to 81% (by experiment) | Lavan et al., PLOS ONE 2025 |
| Recording time used per voice clone | About 4 minutes | Lavan et al., PLOS ONE 2025 |
| Participants per experiment / voices per experiment | 50 / 80 | Lavan et al., PLOS ONE 2025 |
| Dominance rating, generic AI vs human voices | 58.1 vs 40.4 | Lavan et al., PLOS ONE 2025 |
| Trustworthiness rating, generic AI vs human voices | 57.9 vs 53.8 | Lavan et al., PLOS ONE 2025 |
Earlier detection data points the same way. A 2023 study of 529 participants found listeners identified speech deepfakes correctly only 73% of the time, with English and Mandarin speakers performing about the same, and showing them five deepfake examples beforehand raised accuracy by just 3.84% (Mai et al., PLOS ONE 2023). Our audio deepfake detection accuracy statistics track how machine detectors compare.
| Metric | Value | Source |
|---|---|---|
| Participants | 529 | Mai et al., PLOS ONE 2023 |
| Deepfakes correctly identified, single clips | 73% | Mai et al., PLOS ONE 2023 |
| Genuine clips correctly identified, single clips | 67.78% | Mai et al., PLOS ONE 2023 |
| Deepfakes recognized when paired with real clip | 85.59% | Mai et al., PLOS ONE 2023 |
| Accuracy gain from 5 training examples | 3.84% | Mai et al., PLOS ONE 2023 |
7. Preference Arenas and the Market Behind the Scores
With lab MOS saturating, rankings have shifted to blind pairwise preference. Hugging Face launched TTS Arena on February 27, 2024 with six models, asking users to pick the more natural of two anonymous samples and ranking results with an Elo-style system (Hugging Face, TTS Arena launch). On the Artificial Analysis Speech Arena leaderboard read on October 3, 2026, ElevenLabs’ Eleven v4 led at 1321 Elo, but with a plus or minus 189 interval that overlaps every other top 10 model (Artificial Analysis TTS leaderboard). Read honestly, the top tier is a statistical tie between at least six providers.
Hugging Face’s own answer to slow vote collection, the Open TTS Leaderboard launched September 30, 2026, swaps human votes for objective metrics (WER, speaker similarity, speed) and says evaluating a model drops from a couple of weeks to a couple of hours, across a Hub that now hosts more than 8,000 TTS models (Hugging Face, Open TTS Leaderboard).
| Metric | Value | Source |
|---|---|---|
| TTS Arena models at launch (Feb 2024) | 6 | Hugging Face, Feb 2024 |
| #1 Speech Arena model: Eleven v4 (ElevenLabs) | 1321 Elo (plus or minus 189), 1,930 samples | Artificial Analysis, Oct 2026 |
| #2: Qwen-Audio-3.1-TTS-Plus (Alibaba) | 1292 Elo (plus or minus 189) | Artificial Analysis, Oct 2026 |
| #3 / #4: Sonic 3.6 (Cartesia) / Gemini 3.8 Flash TTS (Google) | 1278 / 1275 Elo | Artificial Analysis, Oct 2026 |
| #10: Breeze TTS 2 (BreezeBlue) | 1216 Elo | Artificial Analysis, Oct 2026 |
| TTS models hosted on Hugging Face Hub | More than 8,000 | Hugging Face, Sept 2026 |
The commercial stakes explain the race. Neural and AI voices already hold a 67.18% revenue share of the text-to-speech market, which Mordor Intelligence sizes at USD 4.36 billion in 2026 rising to USD 7.92 billion by 2031 (Mordor Intelligence, Text to Speech Market). Research and Markets puts the 2023 base at USD 3.8 billion with a USD 9.3 billion 2030 forecast (GlobeNewswire release, Dec 2024).
| Metric | Value | Source |
|---|---|---|
| TTS market size, 2025 / 2026 | USD 3.87B / USD 4.36B | Mordor Intelligence, 2026 |
| TTS market forecast, 2031 | USD 7.92B (12.66% CAGR) | Mordor Intelligence, 2026 |
| Neural / AI voice revenue share | 67.18% (15.08% CAGR) | Mordor Intelligence, 2026 |
| North America share, 2025 | 36.78% | Mordor Intelligence, 2026 |
| TTS market, 2023 to 2030 | USD 3.8B to USD 9.3B (13.4% CAGR) | Research and Markets, Dec 2024 |
| China TTS market forecast, 2030 | USD 2.3B (17.5% CAGR) | Research and Markets, Dec 2024 |
Summary: Synthetic Speech MOS by the Numbers
| Metric | Value | Source |
|---|---|---|
| Tacotron 2 MOS vs professional recordings | 4.53 vs 4.58 | Google, 2017 |
| WaveNet gap reduction to human speech | Over 50% | DeepMind, 2016 |
| Human speech MOS range across landmark papers | 4.30 to 4.58 | Google, DeepMind, Kakao, Microsoft, 2016-2021 |
| VITS MOS vs ground truth, VCTK | 4.38 vs 4.38 | Kakao Enterprise, 2021 |
| NaturalSpeech CMOS vs recordings | -0.01 (not significant) | Microsoft Research, 2022 |
| VALL-E 2 CMOS vs ground truth, LibriSpeech | +0.033 | Microsoft, 2024 |
| VALL-E 2 SMOS vs ground truth | 4.61 vs 4.13 | Microsoft, 2024 |
| MaskGCT CMOS vs ground truth | +0.10 | MaskGCT, 2024 |
| Seed-TTS CMOS vs human, English | -0.07 | ByteDance, 2024 |
| CosyVoice 3 WER vs human, English | 1.45% vs 2.14% | Alibaba, 2025 |
| CosyVoice 3 training data | 1 million hours | Alibaba, 2025 |
| Blizzard 2023 systems equal to natural speech on MOS | 2 (first time) | Blizzard Challenge 2023 summary, 2024 |
| MuLanTTS vs natural recording, Blizzard 2023 Hub | 4.3 vs 4.4 | Microsoft, 2023 |
| Papers with full MOS method reporting | Minority | Kirkland et al., SSW 2023 |
| VoiceMOS 2022 systems / teams | 187 / 22 | VoiceMOS Challenge 2022 |
| Voice clones judged human | 58% | Lavan et al., PLOS ONE 2025 |
| Deepfakes correctly identified by listeners | 73% | Mai et al., PLOS ONE 2023 |
| Top Speech Arena Elo | 1321 (plus or minus 189) | Artificial Analysis, Oct 2026 |
| Neural / AI voice share of TTS revenue | 67.18% | Mordor Intelligence, 2026 |
| TTS market, 2026 | USD 4.36B | Mordor Intelligence, 2026 |
Methodology and Sources
Every MOS, CMOS and SMOS figure above comes from the paper or challenge report that ran the listening test, read from the arXiv abstract or full-text tables. Where a paper only showed results in a chart (for example CosyVoice 3’s MOS bars), we used the text conclusion and dropped the approximate values. Self-reported parity claims from model developers are labeled with the developer’s name so readers can weigh them against independent tests such as the Blizzard Challenge.
- DeepMind: WaveNet announcement (Sept 2016) and WaveNet paper
- Google: Tacotron 2 paper (Dec 2017)
- Microsoft and Zhejiang University: FastSpeech 2 (2020)
- Kakao Enterprise: VITS (June 2021)
- Microsoft Research: NaturalSpeech (May 2022), NaturalSpeech 3 (March 2024), VALL-E (Jan 2023), VALL-E 2 (June 2024), MuLanTTS (Sept 2023), DelightfulTTS (Oct 2021)
- ByteDance: Seed-TTS (June 2024)
- MaskGCT (Sept 2024) and F5-TTS (Oct 2024)
- Alibaba: CosyVoice 3 (May 2025)
- Blizzard Challenge: The Blizzard Challenge 2023 (ISCA), Refining the evaluation of speech synthesis, Computer Speech and Language (2024), The Blizzard Challenge 2025
- ITU-T: P.800, P.800.1, P.808
- Kirkland et al.: Stuck in the MOS pit, SSW 2023
- Yang et al.: Towards Responsible Evaluation for Text-to-Speech (Oct 2025)
- VoiceMOS Challenge: 2022, 2023, 2024; MOS-Bench
- Lavan et al.: Voice clones sound realistic but not (yet) hyperrealistic, PLOS ONE (Sept 2025)
- Mai et al.: Warning: Humans cannot reliably detect speech deepfakes, PLOS ONE (Aug 2023)
- Hugging Face: TTS Arena launch (Feb 2024), Open TTS Leaderboard (Sept 2026)
- Artificial Analysis: Text to Speech leaderboard (read Oct 3, 2026)
- Mordor Intelligence: Text to Speech Market; Research and Markets via GlobeNewswire (Dec 2024)
- Data watch: MOS, CMOS and SMOS values are not comparable across papers because listener pools, scales and instructions differ; CMOS scales also differ (F5-TTS uses -3 to +3). Most parity claims in sections 2 and 3 come from developers’ own tests with small panels (for example 12 raters for VALL-E CMOS and 20 for VALL-E 2), and the WaveNet, Tacotron 2, FastSpeech 2 and VITS results are older than three years, shown as most recent available data for those benchmarks. The Blizzard 2023 findings on statistical parity and MUSHRA come from the organizers’ published summary abstract. Artificial Analysis Elo scores change daily as votes arrive and their confidence intervals overlap across the top 10. Market sizing conflicts: Mordor Intelligence puts 2025 at USD 3.87 billion with 12.66% CAGR to 2031, while Research and Markets projects USD 9.3 billion by 2030 at 13.4% CAGR from a 2023 base of USD 3.8 billion; the gap reflects different scope definitions. Full Blizzard 2025 listening test results and the next VoiceMOS Challenge are the main newer data expected.
Last updated: October 3, 2026. We update this roundup quarterly, and the next refresh is expected when the next VoiceMOS Challenge results and the full Blizzard Challenge 2025 listening test analysis are published.