Skip to content

README.md

Latest commit

 

History

History
68 lines (53 loc) · 2.38 KB

README.md

File metadata and controls

68 lines (53 loc) · 2.38 KB
Raw

MOSS-Music

MOSS-Music support for mlx-audio lives in the STT bucket because lyrics ASR is one of the model's primary tasks. The model is prompt-driven audio-text-to-text, so it can also answer music captioning, key/tempo, chord, structure, and general music QA prompts.

Supported Models

Repo Precision Size
OpenMOSS-Team/MOSS-Music-8B-Thinking bf16 ~16.9 GiB
OpenMOSS-Team/MOSS-Music-8B-Instruct bf16 ~16.9 GiB
mlx-community/MOSS-Music-8B-Thinking-8bit 8-bit ~9.5 GiB
mlx-community/MOSS-Music-8B-Thinking-6bit 6-bit ~7.6 GiB
mlx-community/MOSS-Music-8B-Thinking-4bit 4-bit ~5.6 GiB

Python Usage

from mlx_audio.stt import load

model = load("mlx-community/MOSS-Music-8B-Thinking-8bit")

result = model.generate("song.wav")
print(result.text)

The default prompt is:

Please give a detailed musical description of this clip.

Use prompt for transcription or music QA:

result = model.generate(
    "song.wav",
    prompt="Transcribe the lyrics of this song.",
)
print(result.text)

CLI Usage

mlx_audio.stt.generate \
  --model mlx-community/MOSS-Music-8B-Thinking-8bit \
  --audio song.wav \
  --output-path output \
  --prompt "Please give a detailed musical description of this clip."

Notes

  • Time markers are enabled by default to match the upstream serving path. In Python, pass enable_time_marker=False to generate() to disable them for a single call.
  • result.text contains the decoded model response. result.segments is structured when the model emits timestamp markers such as [00:12] lyric or 00:12: lyric; otherwise it falls back to a single full-text segment spanning the input audio.
  • For timestamped JSON/SRT/VTT output, prompt the model to emit timestamped lines and use --format json, --format srt, or --format vtt in the CLI.
  • The Thinking checkpoint may emit <think>...</think> reasoning. mlx-audio strips those blocks by default; pass strip_thinking=False in Python to keep raw output.