Speech synthesis

CTranslate2 integrates experimental Drama 3 text-to-speech inference. It generates mono speech from text using a converted local checkpoint.

Tip

See the Drama 3 guide for conversion and synthesis examples.

The implementation supports offline CPU synthesis with optional CUDA/HIP acceleration for backbone and tower projections. The remaining graph and codec execute on CPU. Voice cloning from reference audio, streaming, automatic long-text splitting, and batching are not implemented. English and Spanish have been exercised, but broad language coverage and speech quality have not been established. Evaluation has found repeated codes, prolonged silence, and generation reaching the frame limit; release acceptance remains incomplete.