MOSS-Music
A local AI music analysis model transcribes lyrics with timestamps, describes songs, and answers questions about full tracks through SGLang or Transformers.
MOSS-Music is a local AI model for people who need to describe, transcribe, or analyze recorded music. It works with full tracks and combines sung lyrics with musical details in its responses.
- Lyrics transcription: Extract sung words over backing music with timestamps for individual words and sentences.
- Song descriptions: Generate captions and tags covering genre, mood, production style, instruments, and vocal characteristics.
- Harmony and rhythm: Identify key, tempo, beats, and downbeats, and transcribe chord progressions with timestamps.
- Song structure: Locate sections such as verses, choruses, and bridges, and analyze repetition and contrast across a track.
- Questions about music: Ask open-ended questions about a recording and get analysis grounded in the full song.
MOSS-Music-8B-Instruct focuses on direct responses to musical prompts. MOSS-Music-8B-Thinking is tuned for reasoning through musical analysis. Both use Qwen3-8B and MOSS-Audio-Encoder.
You can run the models on your own hardware or serve them through SGLang, with a Gradio app available as a browser interface. The Python runtime supports NVIDIA CUDA GPUs. SGLang is the recommended backend for generation quality and throughput; the local Transformers path is primarily for verification and debugging.
Logotyper
