Moss Tts

Moss Tts is a text-to-speech model distributed in GGUF format and designed for use with the CrispASR moss-tts backend. The model is a GGUF conversion of OpenMOSS-Team/MOSS-TTS-v1.5 and incorporates a Qwen3-8B backbone that emits 32 RVQ audio codebooks under a delay pattern, along with the MOSS-Audio-Tokenizer 1.6B transformer codec. It is intended for generating speech from text inputs and is validated to produce intelligible, accurate speech end-to-end, as tested on CUDA (P100) hardware through a round-trip process of synthesis and automatic speech recognition. The model files are available in different configurations, including a Q4_K backbone version and an F16 backbone version, with the latter requiring more than 20 GB of VRAM or CPU resources. There is also an F16 transformer codec companion file. Usage instructions are provided for integrating the model with CrispASR, specifying how to synthesize speech from text and output audio files. The model is released under the Apache-2.0 license. Moss Tts is positioned as a resource for those seeking to implement or experiment with text-to-speech capabilities, particularly in conjunction with the CrispASR backend. The evidence does not specify particular user roles or application domains beyond this context.