What is openmoss?

OpenMOSS is a C++ implementation of a text-to-speech (TTS) system that converts written text into spoken audio, including a voice cloning feature that can match a specific person's voice from a short audio sample. It is a port of MOSS-TTS-Delay, an open-source speech synthesis model from the OpenMOSS Team, compiled into a self-contained binary using GGML, the same low-level machine learning library that powers llama.cpp. The system has two components. The first is a language model backbone (Qwen3-8B, an 8-billion-parameter text AI) that generates sequences of audio codes from your input text. The second is an audio codec, a 1.6-billion-parameter neural network, that converts those codes into actual waveform audio at 24 kHz (CD-quality mono). Voice cloning works by encoding a reference audio clip into the same code format and prepending it to the generation, so the model continues in the same voice. It runs as either a one-shot command-line tool (you provide text, it writes a WAV file and exits) or as an HTTP server that keeps the model loaded in memory and accepts repeated requests via API. The server also includes a small browser-based interface for generating and playing back audio. On a 16 GB GPU (an RTX 5060 Ti), the quantized model produces about 10 seconds of speech in 4 seconds of wall-clock time. You would use this to run high-quality AI speech synthesis locally without sending audio or text to a cloud service. Building from source requires llama.cpp, CMake, and a C++17 compiler. The full README is longer than what was provided.