I'm building a voice assistant that needs to generate speech in a custom persona voice with latency close to real time rather than relying on batch rendering. I'm currently using Coqui XTTS v2, which works reasonably well, but I'm wondering whether newer options provide a better speed-versus-quality balance. Ukrainian support is especially important, since many systems perform well in English but struggle with other languages. Has anyone compared XTTS v2 with F5-TTS, StyleTTS2, or other newer libraries for this kind of application?
2 Answers
Near-real-time voice cloning is genuinely possible now. Earlier voice-conversion and cloning projects often required offline rendering and substantial post-processing, but newer streaming approaches and lighter vocoders can bring latency low enough for a consumer GPU. The same capability can support useful applications like accessibility and personalized assistants, though consent and misuse are important concerns.
Be careful with recognizable public figures or anyone else’s voice. Unauthorized celebrity voice clones can create legal problems in addition to the obvious ethical issues, so an original voice or explicit permission is the safer route for a production assistant.
Agreed. The goal here is an original character voice, not an impersonation, and I’d want the same standard applied to any training or reference material.
That’s the direction I’m seeing too. I’m using a custom persona voice rather than copying a real person’s identity, but I’m still trying to find the best practical speed and quality combination.