Local Open-Source TTS Compared: Speed, Quality, License (Kokoro, XTTS, F5)
In short: Run open TTS locally and the choice splits along three axes: speed, voice cloning, and license. The fastest and lightest is Kokoro-82M, which at 82M parameters and 2-3GB of VRAM reads faster than real time even on CPU (about 36x on a free Colab T4). The catch is that it cannot clone a voice.
Run open TTS locally and the choice splits along three axes: speed, voice cloning, and license. The fastest and lightest is Kokoro-82M, which at 82M parameters and 2-3GB of VRAM reads faster than real time even on CPU (about 36x on a free Colab T4). The catch is that it cannot clone a voice. XTTS v2 (clones 17 languages from a 6-second sample) and F5-TTS (zero-shot from 3 seconds) sound better, but both ship non-commercial licenses.
In one line: use Kokoro (Apache-2.0) or Piper for commercial products; use XTTS v2 or F5-TTS when you need cloning quality and it is personal use.
How is speed measured, and what is fast?#
TTS speed is measured by RTF (real-time factor). Generate a 10-second clip in 1 second and the RTF is 0.1, that is 10x real time; an RTF below 1 means faster than real time. Kokoro-82M is dominant at about 36x on a T4 GPU and about 5x even on a 32-core CPU. F5-TTS runs about 3-5x on an RTX 4070, and XTTS v2 about 2x. So for plain narration Kokoro is fastest and lightest, while turning on cloning roughly halves the speed.
| Model | Size | Speed (x real-time) | VRAM | Voice cloning | License |
|---|---|---|---|---|---|
| Kokoro-82M | 82M | T4 ~36x, CPU ~5x | ~2-3GB | No (54 fixed voices) | Apache-2.0 (commercial OK) |
| Piper | Small | CPU real-time (edge) | CPU is enough | No | GPL-3.0 (commercial OK, copyleft) |
| XTTS v2 | Medium | ~2x | ~4-6GB | 6s sample, 17 languages | CPML (non-commercial) |
| F5-TTS | Medium | 4070 ~3-5x | ~3-5GB | 3s zero-shot | CC-BY-NC (non-commercial) |
How do quality and cloning differ?#
Quality depends on the job. Kokoro is small at 82M, so it is less natural than larger models and cannot clone voices, but its 54 fixed voices are plenty for fast narration (about a 44% win rate on TTS Arena V2). XTTS v2 clones 17 languages from just a 6-second reference, the broadest multilingual cloning, while F5-TTS uses flow matching (a diffusion-family approach) instead of a GPT backbone to zero-shot clone from a 3-second sample at a competitive MOS of about 4.1. Pick XTTS v2 or F5-TTS when cloning is the point, and Kokoro for fast narration.
Why does the license matter most?#
Because the best-sounding model is not always the one you can use. Kokoro (Apache-2.0) and Piper are safe for commercial products, though Piper's original rhasspy/piper (MIT) was archived in October 2025 and development moved to OHF-Voice/piper1-gpl (GPL-3.0), still commercial but with copyleft to weigh. By contrast, XTTS v2 (CPML) and F5-TTS weights (CC-BY-NC) are non-commercial, and since Coqui shut down in January 2024 there is no one to even sell an XTTS commercial license. No matter how good the demo, check the license first.
How do you measure it yourself?#
Compare fairly on the same sentence.
- Synthesize the same paragraph with each model and compute RTF = generation time / audio duration to get the speed multiple.
- Read VRAM from nvidia-smi and compare audio quality by listening blind (MOS varies by methodology, so never compare it absolutely).
- For a commercial product, check each model card's license before measuring, to filter candidates first.
Note: figures are 2025-2026 community measurements across different hardware and methodologies, so cross-model comparison is not always fair. Licenses and forks (especially Piper) change often, so check the official model card before adopting and measure your own with the method above. Reviewed quarterly.
Responses
No responses yet. Be the first to respond.