LocalAI 4.7.0 adds voice cloning, video avatars, and faster streaming TTS

LocalAI 4.7.0, released on July 14 by mudler, introduces significant new capabilities for local AI generation. The headline feature is a managed voice cloning library accessible via the UI: users can record or upload a consented reference audio, save it as a reusable profile, and reference it from any cloning-capable TTS backend using a stable ‘localai://’ URI. This eliminates the need for manual YAML editing or copying audio files into model folders. Additionally, a new longcat-video backend brings text-to-video, image-to-video, and audio-driven talking-avatar generation, integrated into the Studio UI with controls for audio and reference images.

Performance and reliability are also major focuses. The new vibevoice-cpp backend delivers true streaming TTS with a time-to-first-audio of just 2.38 seconds on CPU compared to 39.96 seconds in batch mode — a roughly 17x improvement. The moss-transcribe-cpp backend offers one-pass diarized transcription (multi-speaker recognition with timestamps) and runs 1.6–2.2x faster than reference on CPU. Other improvements include automatic full context window usage via ‘context_size: -1’, GPU device selection for llama.cpp, a model-load cooldown to prevent VRAM leaks from failing models, OIDC authentication fixes, and support for Anthropic-style thinking blocks interleaved with tool calls.