Kokoro can now clone any voice with just ~600 MB of extra VRAM!
Kokoro can now clone any voice using approximately 600 MB of additional VRAM, leveraging OpenVoice, an older but highly-ranked TTS model. Users can achieve better results by experimenting with various built-in TTS voices and higher-quality reference audio. For instance, using the same reference sample, am_michael scores 0.4, whereas am_adam can achieve 0.7, demonstrating the impact of voice selection on final quality and speaker similarity.
Time & source
Times shown in UTC
Display time zone: UTC
Local time zone unavailable; showing UTC.
PublishedOffset at this time: UTC+0Oct 29, 2026, 16:59 UTC
IngestedOffset at this time: UTC+0Oct 29, 2026, 20:00 UTC
- Published
- Oct 29, 2026, 16:59
- Ingested
- Oct 29, 2026, 20:00
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Tier is a per-source editorial setting, not a per-item score.
While looking for audio models, I came across OpenVoice. It's getting pretty old, but surprisingly still ranks high on TTS leaderboards.
What caught my attention was its design: essentially TTS + voice conversion. The ToneColorConverter can run independently of the TTS model.
So I added a small recipe any-tts to audio.cpp for Kokoro + OpenVoice ToneColorConverter. Recipes were introduced in v0.9.2 as predefined pipelines that combine multiple audio models.
You can also try pairing any tiny TTS model with ToneColorConverter. The TTS model determines the speech content, pronunciation, and prosody, while the VC model transfers the target speaker's voice characteristics. The built-in voice you choose can also affect the final output quality.
Disclaimer: This experiment is more about extending the capabilities of existing TTS models, not introducing another voice cloning model.
I used synthetic voices as references, and the speaker similarity scores measured by CAM++ range from 0.7 to 0.8. So the generated voices largely match the reference samples I provided. This is one way to create custom voices for arbitrary TTS models.
For better results, you may need to experiment with different built-in TTS voices and higher-quality reference audio. Both can significantly affect the final quality and speaker similarity. For example, with the same reference sample, am_michael scores 0.4, while am_adam can reach 0.7.