whatmodel.sh / tts / MaskGCT

MaskGCT

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

MaskGCT is an open-weights text-to-speech model from Amphion (CUHK-Shenzhen), released October 2024. Its published size is 1.43B parameters. Licence: cc-by-nc-4.0.

Catalogue identity notes

amphion/MaskGCT (semantic and acoustic codecs, T2S model, 1-layer and full S2A models).

Model information sources

· Parameter count ↗

OUR SUM of the safetensors headers of the three generative checkpoints the official README downloads (pinned revision): T2S 746,399,744 + S2A full 353,439,744 + S2A 1-layer 330,348,544 = 1,430,188,032. The semantic codec (44.3M) and acoustic codec encoder/decoder (42.5M + 128.2M) are not included. The paper (arXiv 2409.00750v3, Table 7) lists T2S-Large 695M and S2A 353M. Observed 2026-10-07.

58.8±1.2#44 of 118 text-to-speech modelsfrom 35 benchmark results across 10 sources

Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.

$ how it compares

The text-to-speech models ranked either side of MaskGCT, out of 148 with a score, then more from Amphion (CUHK-Shenzhen).

#modelscorewhereRAMreleased
51MiniMax Speech-02-Turbo MiniMax60±3api $60.00/M chars2025-04
52TTS-1 OpenAI59±4api $15.00/M chars2023-11
53MaskGCT Amphion (CUHK-Shenzhen)59±1local~6.6GB2024-10
54MGM-Omni-TTS 2B (0927) MGM-Omni (CUHK)59±4comparison only2025-09
55CosyVoice v3 Plus Alibaba Cloud58±4api $26.00/M chars2025-09
more from Amphion (CUHK-Shenzhen)
95Amphion VALL-E (LibriLight 6k) Amphion (CUHK-Shenzhen)47±14local~1.6GB2024-01
97Vevo (Amphion) Amphion (CUHK-Shenzhen)46±3local~6.8GB2024-12
121Amphion VALL-E v2 recipe (MLS) Amphion (CUHK-Shenzhen)36±3local~2.9GB2024-06

see all text-to-speech models · find the ones that fit your machine

$ what it's good at

voice cloning▁·▁▁▁▁▁▁▇█▃▃ #27 of 37 · 33% of its score · 4 tests

Beats CosyVoice2-0.5B here, though they rank higher overall. Trails Higgs Audio V2, F5-TTS Base (v0) here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Seed-TTS test-en · speaker similarity (WavLM-large SV) +271.3% ✎▄····▄·█▄█▄▄ beats 5/8medium · avg 53%▇▇▇▇▇▇ 16%at its level · several runs
Seed-TTS test-zh · speaker similarity (WavLM-large SV) +177.2% ✎▁·▂··▂·█▄·▃▄ beats 16/25medium · avg 62%▇▇▇ 8.1%at its level · several runs
Seed-TTS test-en · Whisper large-v3 WER +32.6% ✎▄···█▄▄·█▄·▄ beats 2/8avg 3.6%▇▇▇ 6.7%blunt · several runs
Seed-TTS test-zh · Paraformer CER +22.4% ✎▁·▂▁▃▁▂▃▅█·▂ beats 4/30avg 2.1%▇ 2.8%blunt · several runs
other tests23 not in a task · 66% of its score
testresultvs the others in its rundifficultycountswhy
CV3-Eval EN (zero-shot English) · WER7.7% ✎▄··▄··▄▄·█·▄ beats 0/6avg 16%▇▇▇ 8.3%typical
TTSDS2 listening test · speaker-similarity MOS (4 English domains)4.39 / 5 ✎▅·█▅▅·█▃▃▅·▅ beats 17/17easy · avg 4.03 / 5▇▇▇ 6.3%typical
CV3-Eval hard-en · speaker similarity (ERes2Net SV)62.4% ✎▄··█··▄·█··▄ beats 0/6medium · avg 49%▇▇ 6.1%at its level
CV3-Eval hard-en · DNSMOS (predicted quality)3.48 / 5 ✎▄·····█·▄▄·█ beats 0/6medium · avg 2.98 / 5▇▇ 5.5%at its level
Raon-OpenTTS-Eval · overall speaker similarity (WavLM)63.5% ✎▄·█·█▄·····▄ beats 3/6medium · avg 44%▇▇ 5.0%at its level
Raon-OpenTTS-Eval · human speaker-similarity MOS (overall)3.58 / 5 ✎█··██······▄ beats 3/6medium · avg 2.84 / 5▇▇ 4.6%at its level
VoxCPM listening test · Chinese speaker-similarity MOS (S-MOS)3.77 / 5 ✎█·······██·█ beats 0/3medium · avg 3.29 / 5▇▇ 4.5%at its level
VoxCPM listening test · English speaker-similarity MOS (S-MOS)4.00 / 5 ✎██········██ beats 1/3medium · avg 3.43 / 5▇▇ 4.3%at its level
15 more tests
testresultvs the others in its rundifficultycountswhy
CV3-Eval hard-en · WER41.1% ✎▄·······█·██ beats 0/6avg 26%▇▇ 4.1%sharp
VoxCPM listening test · English naturalness MOS (N-MOS)3.84 / 5 ✎█······█··██ beats 0/3medium · avg 3.36 / 5▇▇ 4.1%at its level
VoxCPM listening test · Chinese naturalness MOS (N-MOS)3.20 / 5 ✎██·······█·█ beats 0/3medium · avg 2.85 / 5▇ 3.5%sharp · at its level
Raon-OpenTTS-Eval · overall WER (4 acoustic conditions)8.6% ✎▄··▄█····▄·█ beats 3/6avg 20%▇ 2.3%noisy
RVCBench LibriTTS clean voice cloning · ECAPA speaker similarity57.0% ✎▃·····▂▂▂▃█▆ beats 11/15medium · avg 46%▇ 2.1%blunt · at its level
SAP voice reconstruction · human BWS reconstruction preference1024.5 ✎█▄█····▄··▄▄ beats 5/7avg 995▇ 1.7%typical
TTSDS2 listening test · naturalness MOS (4 English domains)3.28 / 5 ✎▅··▃▅··▅█▅██ beats 14/17medium · avg 3.30 / 5▇ 0.9%blunt · at its level
TTSDS2 multilingual benchmark · Chinese overall (YouTube, 2025 round)89.1% ✎▄▄·····▄··▄█ beats 3/5very easy · avg 88%▇ 0.8%blunt
LibriSpeech test-clean zero-shot TTS · HuBERT-large WER lab2.6% ✎█·█········█ beats 2/2avg 2.8%▇ 0.7%blunt · small run
TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round)90.3% ✎▂▂█·▂·█▆▂·▄▃ beats 17/17very easy · avg 89%▇ 0.5%blunt
CodecMOS-Accent VCTK voice clone · human naturalness MOS (S-NAT)3.76 / 5 ✎█··▄▄██▄·▄▄▄ beats 4/11easy · avg 3.99 / 5▇ 0.3%blunt
CodecMOS-Accent VCTK voice clone · human speaker-similarity MOS (S-SPK-SIM)4.59 / 5 ✎▄▄█·▄▄▄▄▄·▄█ beats 11/11easy · avg 4.20 / 5▇ 0.3%blunt
SAP voice reconstruction · TTSDS dual-reference distributional score85.1% ✎▄····▄▄▄▄▄·█ beats 5/7very easy · avg 86%▇ 0.3%blunt
SAP voice reconstruction · WER of reconstructed speech (Whisper turbo)17.5% ✎▄▄▄··█▄··▄·▄ beats 3/7avg 17%▇ 0.0%blunt · noisy
RVCBench LibriTTS clean voice cloning · Whisper-medium WER9.0% ✎▂▂▂··▂·▃▃▃█▂ beats 4/15avg 8.6%▇ 0.0%blunt · noisy

$ where to run it

Other ways to run it

  1. Run it as published · needs ~6.6 GB, NVIDIA only · from github.com/open-mmlab
    git clone https://github.com/open-mmlab/Amphion.git && cd Amphion && pip install -r models/tts/maskgct/requirements.txt && python -m models.tts.maskgct.maskgct_inference

Official weights and documentation: huggingface.co/amphion ↗

$ how the score was worked out · Raon-OpenTTS: Open Models and… 45%, VoxCPM: Tokenizer-Free TTS fo… 16%, OmniVoice: Towards Omnilingua… 11% and 7 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech (arXiv:2605.20830v2)9▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 45%CC BY 4.0
VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning (arXiv:2509.24650v1)4▇▇▇▇▇ 16%CC BY 4.0
OmniVoice: Towards Omnilingual Zero-Shot Text-to-Speech with Diffusion Language Models (arXiv:2604.00688v3)4▇▇▇ 11%CC BY 4.0
DMOSpeech 2: Reinforcement Learning for Duration Prediction in Metric-Optimized Speech Synthesis (arXiv:2507.14988v1)4▇▇▇ 11%CC BY-SA 4.0
TTSDS2 listening test results (ttsds/listening_test)2▇▇ 7%MIT
Amphion MaskGCT README · zero-shot TTS results table · lab-reported3▇ 2%MIT
RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443v2)2▇ 2%CC BY 4.0
An Evaluation Framework for TTS Voice Reconstruction (SAP dysarthric speakers; arXiv:2606.21343v1)3▇ 2%CC BY 4.0
TTSDS2 multilingual benchmark results (ttsds/leaderboard_data)2▇ 1%MIT
CodecMOS-Accent: listening-test MOS of zero-shot TTS across English accents (arXiv:2603.14328v1)2▇ 1%CC BY 4.0
prior · pull toward the middle▇ 1%
weighted average= 58.8

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json