speech.cpp v0.11.0 lands Haru-4B, a full-duplex voice chat model running locally in C++
alexcovo_eth · x · 2026-11-15
Developer gyakuse released speech.cpp v0.11.0, adding support for higurashi-ai's Haru-4B, a full-duplex conversational model that can speak while listening, enabling voice chats on Mac, Windows and Linux.
The project is a ggml-based C++ speech stack combining TTS (Qwen3-TTS, Irodori-TTS), ASR (Qwen3-ASR, ReazonSpeech, parakeet) and Silero VAD, running on Metal, Vulkan and CPU — no Python or CUDA needed — with an OpenAI-compatible server and live transcription. A demo video shows OpenAI's gpt-live-1 talking with Haru-4B. Requirements: 11GB memory on Mac (F16) or 7GB (Q80); 8GB VRAM on Windows/Linux for Q80 chats up to 10 minutes. Speed is close to but slightly slower than the official MLX version.
More from Infra
- ggml tensor library trends on GitHub with 15.8k stars, up 154 in a day — ggml-org · 2026-11-15
- Open-source h3metal runs image gen on M5/M6 matrix units, 2.1-2.3x faster than Draw Things — audenvenn · 2026-11-15
- BiRefNet-lite runs fully on a phone NPU at 1024px in 23.7ms on Snapdragon 8 Elite Gen 5, open weights — subhashg547 · 2026-11-15
- Azure outage draws silence: big consultancies avoid vendor concentration risk talk — DavidLinthicum · 2026-11-15
- 2 Starship launches equal 20 Falcon 9 missions as 52 V3 sats match 500+ V2 capacity — XFreeze · 2026-11-15
- HPE Beats Estimates but Stock Slides as Analysts Doubt AI Server Margins — DavidLinthicum · 2026-11-15