v4.4.3
📦 localaiView on GitHub →
✨ 11 features🐛 17 fixes🔧 11 symbols
Summary
This release introduces significant new TTS capabilities with OmniVoice and expanded Piper voices, alongside performance improvements in request routing and fixes for Darwin builds. It also adds Indonesian and Korean language support.
✨ New Features
- Added Indonesian language support to the react-ui.
- Implemented locate-anything-cpp (open-vocabulary object detection via ggml) in the backend.
- Enabled GGML_CUDA_GRAPHS in the cublas build for parakeet-cpp.
- Bundled espeak-ng and added piper TTS voices to the model gallery via crispasr.
- Added 60 piper TTS voices across 42 languages (Phase 2) to the gallery.
- Added Kokoro TTS + multilingual Piper voices support for sherpa-onnx.
- Added OmniVoice TTS backend (file + streaming, voice cloning + voice design).
- Added Korean (ko) translation.
- Migrated qwen3-tts-cpp to ServeurpersoCom/qwentts.cpp (streaming, speakers, voice design).
- Gated realtime pipeline voice models behind voice recognition.
- Implemented declarative per-model scheduling via environment variables or arguments for distributed setups.
🐛 Bug Fixes
- Implemented production-ready request router + auto-size batching for embedding/rerank requests.
- Fixed publishing of sherpa-onnx and speaker-recognition images for darwin/arm64.
- Fixed crispasr to write piper TTS WAV at the model's native sample rate.
- Fixed routing of vision-language models to the mlx-vlm backend.
- Fixed vibevoice-cpp build linkage and added fail-safe go backend packaging for darwin.
- Fixed agent chat events to emit timestamps in milliseconds.
- Fixed realtime transcription model to remain on a language-only session.update.
- Fixed backend-gating of the top_k=40 sampler default.
- Fixed cogito dependency to resolve MCP image-result panic.
- Fixed neutts by pinning torchaudio to match torch (resolving undefined symbol).
- Fixed gallery to make opus a meta backend for platform auto-selection.
- Fixed meta-backend definitions for platform auto-selection in the gallery.
- Fixed react-ui to stop Talk pipeline overflow and center collapsed-rail avatar.
- Fixed react-ui agent chat timestamps to be format-agnostic.
- Fixed model path listing to use case-insensitive suffix matching and skip .bak files.
- Fixed container-aware total RAM detection (cgroup/lxcfs) for xsysinfo.
- Fixed e2e live-server voice-recognition gate test.