Transcribe Any Music to MIDI Locally

MuScriptor local audio → MIDI pipeline

CPU-capable · optional GPU · no cloud required at runtime

Input

Audio file

WAV / MP3 / stems

Runtime

Pinokio 1-click

Zero-terminal launcher

Native Python / FastAPI

CLI + localhost:8000

MuScriptor model

Small 103M · Medium 307M · Large 1.4B

Output

Per-instrument MIDI

Streaming note events · DAW-ready

Two install paths

Pinokio

Search · Install · Start

or native

HF auth · uvx serve

Transcribe

Drop audio → MIDI

Overview

Model: MuScriptor by Kyutai + Mirelo AI — decoder-only transformer, trained on 170K songs, multi-instrument.
Variants: Small (103M, ~1GB RAM), Medium (307M), Large (1.4B). CPU-capable on all tiers.
Output: per-instrument MIDI with streaming note events, importable into any DAW.
Launcher: 1-click Pinokio app (zero terminal) or native Python CLI with a FastAPI server.

Model Comparison

Variant	Params	RAM	Speed	Best for
Small	103M	~1 GB	Fastest	CPU-only, laptops
Medium	307M	~2 GB	Balanced	Default choice
Large	1.4B	~8 GB	Slow	Max accuracy, GPU

Method 1: Pinokio 1-Click (Easiest)

Install Pinokio from pinokio.co (Mac/Win/Linux).
Search "MuScriptor" in the app store, click Install.
Click Start Small/Medium/Large — downloads the model, starts the server, opens the Web UI.
Drop an audio file → optionally select instruments → download MIDI or WAV.

Pinokio handles the Python env, dependencies, web-client build, and model caching. No terminal needed, and no Hugging Face login required (it uses ungated mirrors).

Method 2: Native Python (Full Control)

Authenticate with Hugging Face

Accept the license at huggingface.co/MuScriptor, then log in.

uvx hf auth login

Launch the Web UI

uvx muscriptor serve   # opens browser at localhost:8000

Transcribe from the CLI

uvx muscriptor transcribe audio.wav --model medium

Automate with the Python API

model.transcribe_to_midi("audio.wav")

HTTP API Endpoints

Endpoint	Method	Description
/health	GET	Server status check
/instruments	GET	List valid instrument names
/transcribe	POST	Upload audio, stream SSE events → MIDI
/auralize	POST	Render MIDI → WAV (FluidSynth)

Transcribe Any Music to MIDI Locally: MuScriptor