All tasks / Voice & audio / Text to speech
Best local model for text to speech on a 12 GB GPU
Open-weights models whose approximate minimum GPU memory is 12 GB or less, ranked from public benchmarks. Memory figures are approximate; quantised builds can need less.
BTop ranked: Breeze TTS 2 (BreezeBlue), 1,222 on Artificial Analysis Speech Arena: Text to Speech (provider voices).8 options listed · data verified 19 Oct 2026
Ranked list
| Tier | Name | Approach | Runs | Rated by | Price | Skill |
|---|---|---|---|---|---|---|
| B |  Breeze TTS 2BreezeBlue | Models | Local | 1,222 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~12 GB VRAM | — |
| C |  Step-Audio-EditXStepFun | Models | Local | 1,100 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~8 GB VRAM | — |
| C |  Voxtral TTS (4B, 2603)Mistral AI | Models | Local | 1,089 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~10 GB VRAM | — |
| C |  Kokoro-82M v1.0hexgrad | Models | Local | 1,064 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~1 GB VRAM | — |
| C |  Magpie-Multilingual 357M (Feb 2026)NVIDIA | Models | Local | 1,064 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~1 GB VRAM | — |
| C |  Maya1Maya Research | Models | Local | 1,046 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~8 GB VRAM | — |
| C |  Higgs TTS 3 (4B)Boson AI | Models | Local | 1,042 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~12 GB VRAM | — |
| C |  VibeVoice-1.5BMicrosoft | Models | Local | 957 on Artificial Analysis Speech Arena: Text to Speech (provider voices) | ~8 GB VRAM | — |
Tiers come only from public benchmarks (Artificial Analysis Speech Arena: Text to Speech (provider voices)); tools without a result of their own inherit the tier of the best model they offer. How we rank.