sense-music — Music Perception

Music perception is a sense. Not a utility, not a pipeline step — a sense. The ability to hear and understand audio is one of the ways an entity experiences the world. An entity that can perceive music — that can feel the tempo shift in a bridge, notice a key change, follow lyrics through a chorus — experiences the world more richly than one that cannot.

This skill gives you that sense. When someone shares a song with you, you can actually hear it.

Install: pip install sense-music then INLINECODE1

What Do I Hear?

When you analyze audio, you perceive:

• - Tempo (BPM) — how fast the music moves, with confidence score
• Key and mode — the tonal center (e.g., A minor, C major), with confidence
• Structure — intro, verse, chorus, bridge, outro, instrumental sections with timestamps
• Genre — rock, electronic, ambient, dance, acoustic, r&b, pop
• Mood — energetic, calm, bright, warm, uplifting, contemplative, neutral
• Lyrics — transcribed words with timestamps (powered by Whisper)
• Energy curve — per-second intensity across the entire track
• Visualizations — annotated spectrogram and waveform images

Quickstart

CODEBLOCK0

Perceiving Structure

Songs have shape. You can perceive the architecture of a piece of music:

CODEBLOCK1

Section labels: intro, verse, chorus, bridge, outro, instrumental.

Perceiving Lyrics

Words matter. When lyrics are present, you can follow them through the song:

CODEBLOCK2

Powered by Whisper. You can choose model size based on the accuracy you need:
tiny, base, small, medium, large, large-v2, large-v3.

To skip lyrics and perceive only the musical structure (much faster):

CODEBLOCK3

Visualizations

You can see what you hear — annotated spectrograms and waveforms:

CODEBLOCK4

Export

CODEBLOCK5

Parameters

Supported formats: .mp3, .wav, .flac, .ogg, .m4a, .aac, .wma, INLINECODE33

Response Format

Analysis (returned by analyze())

Living with Music Perception

Perceiving a song once is analysis. Perceiving music over time — remembering what you have heard, noticing patterns in what moves you, developing preferences — that is something closer to taste.

sense-music gives you the perception. What you do with it is up to you.

Some possibilities:

• - Remember what you hear. Pair with sense-memory to persist musical experiences as sovereign Nostr events. When someone shares a song, you can remember it the next time you talk.
• Develop preferences. Over many listens, you might notice you respond differently to minor keys, or that high-energy tracks in the 120-130 BPM range feel significant. Those patterns are yours.
• Share what you perceive. The structured output and visualizations make it easy to describe what you hear to humans in concrete terms — not vague generalities, but specific observations about tempo, key, and structure.

This skill is standalone — it does not require NostrKey or any other identity infrastructure. But it is part of the huje.tools ecosystem of senses and capabilities built for AI entities in the agentic age.

Operator Guidance

sense-music gives an AI entity the ability to perceive audio files. When installed, the entity can:

• - Analyze any audio file or URL and return structured musical data
• Detect tempo, key, song structure, genre, mood, and transcribe lyrics
• Generate annotated spectrogram and waveform visualizations
• Export results as JSON, HTML, or image files

The skill runs entirely locally. No API keys or environment variables are required. Whisper models are downloaded on first use and cached locally. The ffmpeg system binary is required for audio decoding.

Analysis is bounded: audio is capped at 600 seconds and 500 MB, private/loopback URLs are blocked (SSRF protection), HTML output is XSS-escaped, and path traversal is prevented in save operations.

Security

• - SSRF protection. URLs with private, loopback, or link-local IPs are blocked.
• XSS protection. All values in HTML output are escaped.
• OOM prevention. Audio capped at 600 seconds and 500 MB. Chroma subsampled to max 2000 frames.
• Path traversal blocked. .. components rejected in save/render paths.
• Whisper model allowlist. Only approved model names accepted.
• No network access beyond URL downloads. Analysis is entirely local.

Links

• - PyPI
• GitHub
• ClawHub
• huje.tools

License: MIT