speech.cpp is a C++ speech toolkit on ggml that bundles text-to-speech, speech recognition and voice activity detection behind one command, speech, and one shared library, libspeech, with a C API. It runs on CPU, Metal and Vulkan, needs no Python or CUDA, and fetches each model by name as a single GGUF file with its codec. The synthesis models are Qwen3-TTS (0.6B and 1.7B, ten languages, nine named speakers, streaming audio) and Irodori-TTS for Japanese, including voices cloned from a reference recording or described in words. Recognition covers Qwen3-ASR (30 languages), ReazonSpeech v2 and parakeet-tdt models for Japanese and 25 European languages, with Silero VAD to split speech into utterances. Subcommands include speech tts, asr, vad, voice, quantize, serve and worker; speech asr --live transcribes the microphone as someone speaks. speech serve exposes OpenAI-style /v1/audio/speech and /v1/audio/transcriptions endpoints plus a browser page for trying models, listening on 127.0.0.1 with no authentication, and speech worker offers a JSON Lines protocol for programs that keep models loaded. The README reports measurements from 2026-10-14 on an Apple M5 (Metal) and an RTX 2080 (Vulkan): Qwen3-TTS 0.6B at Q8_0 sends its first audio 0.035 s and 0.024 s after the text, and parakeet-tdt_ctc-0.6b-ja reaches 7.9% CER on the Common Voice 8.0 Japanese test set at about 0.045 s median latency on the M5. The README states that each stage of every port is compared with tensors from the official implementation. The code is MIT; model weights keep their authors' licenses, including Apache-2.0 for Qwen3 models, CC-BY-4.0 for parakeet, and ethical restrictions on Irodori-TTS voices.