Best voice cloning models to run locally on a 32GB graphics card
A leaderboard of the best voice cloning models you can run locally on a 32GB graphics card (with 32GB of system RAM). dots.tts SOAR 2B is #1 right now, followed by LongCat-AudioDiT-3.5B and dots.tts MF 2B (MeanFlow, NFE 4). Models are ranked on how clear the cloned voice is and how much it sounds like the original speaker. Each model shows the download that fits and the command to run it.
Scores use reviewed voice cloning evaluations only. Compare scores within this task. Models without task measurements are excluded.
+ more
$ terminal
↗ share
detecting your OS…
Score vs memory to run
● open weights ◆ API only ─┘ best score for the memory · x: GB of RAM / VRAM for the smallest download (log)
| # | model | task score↓ | where↓ | speed↓ | sources |
|---|
| # | model | task score↓ | where↓ | speed↓ | sources |
|---|---|---|---|---|---|
| 2 |  dots.tts SOAR 2B · bf16 (dots.tts package CLI) | 67±4 | local |  | |
| 4 |  LongCat-AudioDiT-3.5B · fp32 weights (official LongCat-AudioDiT repo, transformers>=5.3) | 66±4 | local |  | |
| 6 | dots.tts MF 2B (MeanFlow, NFE 4) · bf16 (dots.tts package CLI) | 62±6† | local | | |
| 7 |  FireRedTTS3 (base) · bf16 (FireRedTTS3 repo) | 62±4 | local |  | |
| 8 |  VoxCPM2 · bf16 (voxcpm) | 61±3 | local | | |
| 9 | LongCat-AudioDiT-1B · fp32 weights (official LongCat-AudioDiT repo, transformers>=5.3) | 60±6† | local | | |
| 10 |  Higgs Audio V3 TTS · bf16 | 59±6*† | local | | |
| 11 |  OmniVoice · fp32 | 59±3 | local | | |
| 13 |  Qwen3-TTS 12Hz 1.7B Base · bf16 (qwen-tts) | 58±3 | local | | |
| 14 |  IndexTTS (v1) · fp32 on CPU (index-tts repo CLI `indextts`) | 57±7 | local | | |
| 15 | VoxCPM 0.5B · bf16 | 57±4 | local | | |
| 16 |  F5-TTS v1 Base · fp32 | 56±6 | local | | |
| 17 |  Fish Audio S2 Pro · bf16 | 56±3 | local | | |
| 18 | Qwen3-TTS 12Hz 0.6B Base · bf16 (qwen-tts) | 56±4 | local | | |
| 19 | IndexTTS-2 · fp32 (index-tts repo) | 54±3 | local | | |
| 20 | ZipVoice · fp32 (ZipVoice repo) | 54±6 | local | | |
| 21 |  DMOSpeech 2 · fp32 (DMOSpeech2 repo) | 53±6† | local | | |
| 22 |  Fun-CosyVoice3-0.5B-2512 · fp32 (CosyVoice repo) | 53±3 | local | | |
| 23 | E2 TTS Base (F5-TTS reproduction) · fp32 | 53±4 | local | | |
| 26 |  MOSS-TTS v1.5 (8B) · fp32 (transformers 5) | 53±3 | local | | |
| 28 | F5-TTS Base (v0) · fp32 | 52±6† | local | | |
| 29 |  GLM-TTS · fp32 (official GLM-TTS repo) | 52±4 | local | | |
| 30 | Higgs Audio V2 · bf16 (transformers >=5.3.0) | 51±6*† | local | | |
| 32 |  MaskGCT · fp32 | 50±3 | local | | |
| 33 | CosyVoice2-0.5B · fp32 (CosyVoice repo) | 48±3 | local | | |
| 35 |  Spark-TTS 0.5B · fp32 (Spark-TTS repo CLI) | 41±5 | local | | |
| 36 | CosyVoice-300M · fp32 (CosyVoice repo) | 40±5 | local | | |
| 37 |  Llasa-8B · bf16 (transformers + xcodec2) | 37±4 | local | | |
| 38 | FireRedTTS (v1) · fp32 (FireRedTTS repo) | 33±5 | local | | |
| 39 |  VoiceCraft 830M TTSEnhanced · fp32 (jasonppy/VoiceCraft repo) | 28±12* | local | | |
| 40 |  XTTS v2 · fp32 (coqui-tts 0.27.5) | 23±6 | local | | |
| 41 | StyleTTS 2 (LibriTTS) · fp32 (styletts2 PyPI package 0.1.6) | 17±7* | local | | |
| 42 | Tortoise TTS v2 · fp16 (neonbjb/tortoise-tts) | 12±7* | local | | |
| 43 | YourTTS · fp32 (coqui-tts) | -1±7*† | local | | |
| — |  Voxtral TTS (4B) · bf16 | not yet calibrated | local | | |
| — |  ZONOS2 8B · bf16 (Zyphra ZONOS2 PyTorch server) | not scored yet | local |
+ 11 more don't fit your machine or budget · show them
47 of 47 catalogue models listed here · sources and licences
* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory