---
title: Hum to Song
emoji: 🎤
colorFrom: gray
colorTo: pink
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
short_description: Hum a melody, get a finished song
python_version: "3.12"
startup_duration_timeout: 1h
license: cc-by-nc-4.0
models:
  - Mothersuperior/YuE2-hum-to-song
  - m-a-p/YuE2-3B
  - m-a-p/YuE2-Vae
  - m-a-p/SheetSage2
datasets:
  - amanteur/CHAD_hummings
---

# 🎤 Hum to Song

Hum a melody for 10–30 seconds, add a style line and lyrics, and get back a produced
song that keeps your tune, builds a structure around it, and carries on long after the
hum stops.

This Space runs [`Mothersuperior/YuE2-hum-to-song`](https://huggingface.co/Mothersuperior/YuE2-hum-to-song)
on top of [`m-a-p/YuE2-3B`](https://huggingface.co/m-a-p/YuE2-3B), following the
adapter's reference scripts (`hum_continue.py`, `infer_hum.py`):

1. **Score continuation.** [SheetSage2](https://huggingface.co/m-a-p/SheetSage2)
   transcribes the hum to ABC in YuE2's two-voice score format. The prompt is built as
   `[EOD] instructions+style+lyrics [ABC_START] hum score` and deliberately left **open**
   (no `[ABC_END]`), so the planner treats the hummed bars as the opening of a song and
   keeps writing — new sections, transitions, an ending — before emitting semantic tokens
   for the whole song.
2. **Prosody adapter.** The hum is reduced to a pitch carrier (pYIN f0 → sine wave,
   amplitude = the rectified hum low-passed with a 4th-order 30 Hz and a 2nd-order 80 Hz
   Butterworth), VAE-encoded to 25 Hz latents, and four `Linear(64 → 2048)` projections add
   it into the NAR hidden state at layers 0 / 7 / 14 / 21 while the rank-96 LoRA-adapted
   flow-matching decoder renders the audio. You can hear exactly what the decoder is given
   in the *Pitch carrier* player.

Three melody modes: **Continue from hum** (the above), **Hum only** (close the score right
after the hum, so the song is exactly your melody), and **Ignore hum melody** (the planner
writes its own tune and the adapter only shapes phrasing).

## Notes

- The transcriber needs something voice-like — a real hum works, a synthesised tone does not.
- The adapter is subtle by design; it is best judged over the first hummed section.
- The hum sets the melody and the opening, not the length: the planner decides where the
  song ends, and **Max song length** is only a cap.
- Weights derive from YuE2-3B and are **CC BY-NC 4.0 — non-commercial use only**.
- The example hum comes from the
  [CHAD hummings subset](https://huggingface.co/datasets/amanteur/CHAD_hummings)
  (Amatov et al., *A Semi-Supervised Deep Learning Approach to Dataset Collection for
  Query-by-Humming Task*, ISMIR 2023), CC BY-NC 4.0.