nyosegawa/speech.cpp is a C++-based speech synthesis, recognition, and voice activity detection (VAD) project. v0.8.10, released on October 17, 2026, shows model loading and synthesis, transcription, and detection progress in real time. If a flag, subcommand, or model name is entered incorrectly, it infers and suggests the intended value, and the help output is now grouped by item. speech serve gained a banner, one log line per request, and escaped control characters in its output.
The project provides several processing steps needed for voice conversation from a single repository. Synthesis supports Qwen3-TTS and Irodori-TTS, recognition supports Qwen3-ASR, ReazonSpeech, and parakeet, and speech segment detection supports Silero VAD. Targets are Metal, Vulkan, and CPU, and it also includes an OpenAI-compatible server and real-time transcription. The changes in this release are not new models or backends, but focus on the usage flow around progress, command input, and server logs.
The provided release information contains no figures on progress refresh intervals or a backend-by-backend performance comparison. There is therefore no basis for saying this change improves inference speed or processing quality. The verifiable improvement is that work status, guidance for incorrect input, and server request logs are displayed more specifically.