Hum

name: hum description: > Persona humming pipeline. Downloads songs, separates vocals, converts to persona voice via RVC, and caches for playback during idle conversation. Orchestrates: /ingest-youtube, /create-stems, /learn-artist, /create-music. triggers:

• hum
• add hum
• teach humming
• hum a song
• humming pipeline
• convert vocals
• persona singing allowed-tools:
• Bash
• Python
• Read
• Write metadata: short-description: "Persona humming via vocal stem conversion" author: graham version: "0.1.0"

provides: - hum composes: [, task-monitor]

/hum

Convert any song into a persona's humming voice. Downloads audio, stems vocals, converts to the persona's RVC voice model, and caches with Federated Taxonomy metadata for mood-driven playback during idle conversation.

Quick Start

cd .pi/skills/hum

# Train a persona's RVC voice model (one-time per persona, ~2hr)
./run.sh train --persona embry
./run.sh train --persona brandon

# Add a song to any persona's humming library
./run.sh add "https://youtu.be/Dordpe3KX_I" \
  --persona embry \
  --mood playful,curious \
  --bridges Loyalty,Resilience

# Different personas, different songs
./run.sh add "https://youtu.be/xyz" \
  --persona brandon \
  --mood melancholic \
  --bridges Fragility,Resilience

# List cached hums for a persona
./run.sh list --persona embry

# Play a cached hum
./run.sh play hawaiian_war_chant --persona embry

# Sanity check all dependencies and all personas
./run.sh sanity

Multi-Persona

Any persona with TTS samples at /mnt/storage12tb/media/personas/<name>/tts_output/ can be trained and used. Each persona gets its own RVC voice model and hum cache.

/mnt/storage12tb/media/personas/
  embry/
    tts_output/     ← voice samples (any persona needs these)
    hum-cache/      ← generated humming audio + manifest
  brandon/
    tts_output/
    hum-cache/

Pipeline

YouTube URL
    |
    v
[ingest-youtube]  yt-dlp --extract-audio
    |
    v
full_mix.wav
    |
    v
[create-stems]    Demucs htdemucs_6s --two-stems vocals
    |
    v
vocals.wav
    |
    v
[create-music]    RVC inference with persona model
    |
    v
persona_vocals.wav
    |
    v
[hum-cache]       /mnt/storage12tb/media/personas/<name>/hum-cache/
                   + manifest.json with taxonomy metadata

Commands

Options

Storage Layout

/mnt/storage12tb/media/personas/<persona>/hum-cache/
  manifest.json              # Index of all cached hums
  hawaiian_war_chant.wav     # Converted audio
  hawaiian_war_chant.json    # Per-track metadata

Track Metadata Schema

{
  "id": "hawaiian_war_chant",
  "title": "Hawaiian War Chant",
  "artist": "Lennon Sisters",
  "source_url": "https://youtu.be/Dordpe3KX_I",
  "source_video_id": "Dordpe3KX_I",
  "bridge_attributes": ["Loyalty", "Resilience"],
  "mood": ["playful", "curious"],
  "persona_connection": "Linguistics degree, Hawaiian cultural ties",
  "duration_s": 120,
  "pitch_shift": 0,
  "f0_method": "rmvpe",
  "created": "2026-02-11T12:00:00",
  "forbidden": false
}

Manifest Schema

{
  "persona": "embry",
  "tracks": [
    { "id": "hawaiian_war_chant", "file": "hawaiian_war_chant.wav", ... }
  ],
  "updated": "2026-02-11T12:00:00"
}

Integration

With /converse

The converse idler reads the hum-cache manifest to select tracks matching the current emotional state. Playback goes through the AudioMixer "humming" channel at 60% volume, ducking to 0% when speech starts.

# In converse/idler.py
from hum.src.cache import HumCache

cache = HumCache(persona="embry")
track = cache.select(mood="playful", bridges=["Loyalty"])
mixer.play("humming", track.audio_path, volume=0.6)

With /learn-artist

Training uses the learn-artist skill's Docker-based RVC pipeline:

# learn-artist handles Docker container lifecycle
cd .pi/skills/learn-artist
./run.sh train "embry" \
  --source-dir /mnt/storage12tb/media/personas/embry/tts_output \
  --category voice --epochs 200

With /create-stems

Vocal separation uses the create-stems skill:

cd .pi/skills/create-stems
./run.sh separate --mix song.wav --out /tmp/stems --instrument vocals

With /create-music

RVC inference uses the create-music skill:

cd .pi/skills/create-music
./run.sh rvc-infer \
  --model-name embry \
  --input vocals.wav \
  --output humming.wav \
  --f0method rmvpe

Dependencies

Safety

• The Kamakawiwoole guard remains active. Hawaiian music tagged as sentimental or grief-triggering is marked forbidden: true in track metadata.
• The converse idler checks forbidden before playback.
• Bridge attribute Fragility above 0.7 triggers extra review before caching.