Resonance of Tokens

2026-02-16

A sound experiment exploring how an AI, trained on text and images, imagines sound.

No music software or audio samples were used. Each model translated written descriptions of five elements into audio solely by writing Python scripts. They calculated raw frequencies mathematically from scratch, saving the data directly as .wav files.

The exact descriptions provided to the models were:

• Earth: "The deep groan of shifting bedrock."
• Water: "Ocean waves on a shore at night."
• Fire: "A crackling wooden campfire."
• Air: "A sweeping wind through the trees."
• Human: "Calm recitation of poetry."

The cover art is similarly generated: pure SVG code based on the combined descriptions of all five elements.

This project is a comparative study, distributed across two platforms to serve different purposes:

[ YouTube: Selected Iterations ]

A 30-track compilation featuring one selected iteration from each of the six models. Designed for continuous listening and visual comparison of the SVG cover art.

[ SoundCloud: Complete Archives ]

The full scope of the experiment. Six distinct albums (one per model) containing 15 tracks each, demonstrating how a specific model's logic varies across three iterations per element.

• Resonance of Tokens - Anthropic: Claude Opus 4.6
• Resonance of Tokens - DeepSeek: DeepSeek V3.2 [ Coming Soon ]
• Resonance of Tokens - OpenAI: gpt-oss-120b [ Coming Soon ]
• Resonance of Tokens - MoonshotAI: Kimi K2.5 [ Coming Soon ]
• Resonance of Tokens - Z.ai: GLM 4.7 [ Coming Soon ]
• Resonance of Tokens - Google: Gemma 3 27B [ Coming Soon ]

Notes:

• Claude Opus 4.6: With adaptive thinking, this model required more generation time than the other five models combined.
• Exclusions: Models trained on audio data (such as GPT-5.2, Gemini 3, and Grok-4) were excluded.