audio.cpp v0.9.4 supports Drama3. Under 7 GB VRAM with CPU offloading! No Python required.
#12
by audio-cpp - opened
audio.cpp v0.9.4 supports Fish Audio Drama3, including voice cloning, streaming generation and long-form narration.
On an RTX 5090, CPU expert-weight offloading reduced peak VRAM from 22.9 GiB to 6.5 GiB, with 0.4199 RTF—still faster than real time!
These results use Q4_K_S experts, Q8_0 tower projections and eight pinned expert-cache slots. Offloading trades speed and additional host RAM for lower VRAM; memory requirements vary by input.
Try it through the CLI or WebUI: audio.cpp
If audio.cpp helps with your projects, a star on GitHub or a like here would mean a lot. Thanks for the support!