Developers Evaluating Coqui XTTS-v2

[SansoftIMS]

After extensive testing and development time, I’d like to share an objective observation regarding Coqui’s XTTS-v2 model for text-to-speech and voice cloning.

While the repository is presented as open-source, in practice XTTS-v2 behaves much more like a closed API-driven system than an accessible research model. Developers expecting deep customization or reproducible cloning results may find the current release limiting.

Key Technical Limitations

1. Loss of 256-D / 512-D Embedding Support
   Earlier Coqui and VITS versions allowed direct embedding access, enabling genuine speaker cloning and reproducibility. XTTS-v2 abstracts this away entirely, preventing control over or reuse of learned voice profiles.

2. Hidden Voice-Cloning Pipeline
   The clone_voice() functionality is now internally managed. Parameters once available to developers (speaker embeddings, model weights, reproducible seeds) are no longer user-accessible.

3. Limited Reproducibility
   Because of the internalized pipeline, cloned voices cannot be recreated consistently across sessions — a serious drawback for research and production workflows.

4. Multilingual Claims Are Overstated
   Although the model advertises multilingual support (including Hindi), the resulting voices sound generic and do not preserve speaker identity across languages.

5. Open-Source in Name, Not in Function
   The visible code only exposes an inference wrapper; the actual voice-cloning mechanism resides behind Coqui’s proprietary API infrastructure. This creates a mismatch between the repository’s “open-source” label and what developers can realistically use or study.

Summary

XTTS-v2 should be treated as an API-bound inference interface rather than a true open-source voice-cloning framework. Developers seeking transparency, reproducibility, or fine-grained control over embeddings will not find it here.

This note is shared to help other developers set realistic expectations before investing time in XTTS-v2 for local or research-grade cloning tasks.

Please correct me if what I mentioned is wrong.

[eginhard]

This looks like AI hallucinations. Speaker embeddings can be accessed, seeds can be fixed for reproducibility and there is no proprietary API, everything runs locally.