hum2song（hum2song 音频）

Hum2Song

Turn a hummed melody into a complete song with local audio processing, without uploading sensitive recordings to third-party services.

Overview

This skill converts user humming or singing into complete songs using local AI models. The entire pipeline runs on your machine - no audio data is sent to external services.

Pipeline:

1. 🎤 Audio Input → 2. 🎵 MIDI Extraction → 3. 🎼 Music Generation → 4. 🎧 Complete Song

Triggers

Use this skill when the user:

• Hums or sings a melody and wants to turn it into a full song
• Has an audio recording of humming/singing
• Wants to create music from their own melodic ideas
• Asks to "turn my humming into a song"

Requirements

System Dependencies

# macOS
brew install ffmpeg fluidsynth

# Ubuntu/Debian
sudo apt-get install ffmpeg fluidsynth

# Python packages
pip install basic-pitch pretty_midi librosa soundfile numpy

Optional: ACE-Step for Music Generation (User Choice)

ACE-Step is an optional local AI. Users decide whether to install it.

# User manually installs if they want AI generation
# Otherwise, default SoundFont synthesis works without AI
git clone https://github.com/ace-step/ace-step.git
pip install -r ace-step/requirements.txt

Note: First use downloads ~4GB model weights to local cache. No automatic downloads.

Core Workflow

Step 1: Extract MIDI from Audio

Use Basic Pitch (Spotify's open source tool) to convert humming to MIDI:

from basic_pitch.inference import predict
from basic_pitch import ICASSP_2022_MODEL_PATH

# Convert audio to MIDI
model_output, midi_data, note_events = predict("humming.wav")
midi_data.write("extracted.mid")

Step 2: Enhance MIDI Structure

Clean and enhance the extracted MIDI:

import pretty_midi

# Load extracted MIDI
pm = pretty_midi.PrettyMIDI("extracted.mid")

# Quantize notes to fix timing
for instrument in pm.instruments:
    for note in instrument.notes:
        note.start = round(note.start * 4) / 4  # Quantize to 16th notes
        note.end = round(note.end * 4) / 4

# Save enhanced MIDI
pm.write("enhanced.mid")

Step 3: Generate Full Song

Option A: ACE-Step (Local AI, Optional)

Only if user has manually installed ACE-Step:

from ace_step import MusicGenerator

# Load model (runs locally, downloads weights on first use)
generator = MusicGenerator.from_pretrained("ace-step/base")

# Generate music from MIDI
audio = generator.generate_from_midi(
    midi_path="enhanced.mid",
    style="pop",
    mood="upbeat",
    duration=120
)

# Save result
audio.save("complete_song.mp3")

Option B: MIDI + SoundFont (No AI)

import pretty_midi

# Load MIDI
pm = pretty_midi.PrettyMIDI("enhanced.mid")

# Synthesize with high-quality SoundFont
audio_data = pm.fluidsynth(fs=44100, sf2_path="path/to/good_soundfont.sf2")

# Save as WAV
import soundfile as sf
sf.write("complete_song.wav", audio_data, 44100)

Usage

Quick Start

# Run the complete pipeline
python ~/.openclaw/skills/hum2song/scripts/hum2song.py \
  --input my_humming.wav \
  --style pop \
  --mood upbeat \
  --output my_song.mp3

Parameters

Response Format

When User Provides Audio

🎵 I'll convert your humming into a complete song!

**Processing Pipeline:**
1️⃣ Extracting melody from your audio...
2️⃣ Converting to MIDI notation...
3️⃣ Generating full arrangement...
4️⃣ Rendering final song...

⏱️ Estimated time: 2-5 minutes

**Settings:**
• Style: {style}
• Mood: {mood}
• Duration: {duration}s

Processing now... 🎧

When Complete

✅ Your song is ready!

📁 Output: {output_path}
🎼 Format: {format}
⏱️ Duration: {actual_duration}s

**What was done:**
• Extracted {note_count} notes from your humming
• Generated {instrument_count} instrument tracks
• Applied {style} style with {mood} mood

💡 You can:
• Adjust style/mood and regenerate
• Export as MIDI for editing in DAW
• Try different AI models for variation

Example Interactions

Example 1: User Hums Audio

User: [uploads humming audio]

AI: 🎵 I'll convert your humming into a complete song!

**Processing Pipeline:**
1️⃣ Extracting melody from your audio...
2️⃣ Converting to MIDI notation...
3️⃣ Generating full arrangement...
4️⃣ Rendering final song...

⏱️ Estimated time: 2-5 minutes

What style would you like?
• Pop (default)
• Rock
• Jazz
• Classical
• Electronic

Example 2: User Specifies Style

User: Turn this humming into a jazz song

AI: 🎷 Jazz transformation incoming!

Processing your humming with jazz style:
• Swing rhythm
• Walking bass line
• Piano chords
• Brush drums

[Processing...]

✅ Your jazz song is ready!
📁 Output: ~/Music/hum2song_jazz_20240321.mp3

Example 3: User Wants MIDI Output

User: Can I get the MIDI file to edit myself?

AI: 🎼 Absolutely!

I'll generate:
• Extracted MIDI (raw melody)
• Enhanced MIDI (quantized, cleaned)
• Full arrangement MIDI (all instruments)

All files will be in: ~/Music/hum2song_export/

Technical Details

Audio Processing

Input Formats: WAV, MP3, M4A, FLAC, OGG Sample Rate: Automatically converted to 44.1kHz Channels: Mono/Stereo → Mono for processing

MIDI Extraction

Model: Basic Pitch (Spotify, ICASSP 2022) Pitch Range: C1 to C8 Note Detection: Polyphonic capable Timing Resolution: 10ms

Music Generation

ACE-Step Model:

• Size: 1B parameters (base), 3B (large)
• Training: Licensed music dataset
• Output: 44.1kHz stereo
• Latency: ~1s per second of audio on M1 Mac

SoundFont Synthesis:

• No AI required
• Real-time synthesis
• High-quality instrument sounds
• Deterministic output

Limitations

• Requires local Python environment setup
• ACE-Step needs ~4GB RAM for base model
• Processing time: 2-5 minutes for a 2-minute song
• Quality depends on humming clarity
• Complex harmonies may not be fully captured

Privacy & Security

✅ All processing is local - Your audio never leaves your machine ✅ No cloud services - No API keys or external uploads ✅ Open source tools - Basic Pitch, ACE-Step, Pretty MIDI ✅ No data collection - Nothing is logged or transmitted

References

• basic-pitch.md - Audio to MIDI extraction
• ace-step.md - AI music generation
• pretty_midi.md - MIDI processing
• librosa.md - Audio analysis utilities

Technical Information

Note: This skill requires local setup of Python dependencies. All audio processing happens on your device for maximum privacy.