rmaapk/music

Acoustic-RAG: Ground-Truth Physical Acoustic Codebook

Permanent Research DOI: 10.5281/zenodo.23868153
Direct Zenodo Resolver: https://doi.org/10.5281/zenodo.23868153
Author: RMAA Holdings

---

1. Executive Summary & Thesis

Standard generative neural audio models (e.g., Suno, Udio) attempt to simultaneously predict lyric alignment, musical pitch ($F_0$), vocal tract resonance (formants), and high-frequency phase coherence from coarse Mel-spectrogram latents. Under epistemic uncertainty, the generative model interpolates between candidate states, producing the metallic phase-smeared haze known as the "AI husk".

Acoustic-RAG replaces probabilistic vocoding with a deterministic, physical acoustic codebook:

1. The Base Model: Sequences discrete musical arrangement and timing states without predicting raw waveforms.
2. The Acoustic RAG Codebook: An immutable, memory-mapped spectral library of settled, phase-coherent excitation frames.
3. The Parametric Plug: Reconstructs pristine broadcast audio by grafting $F_0$ pitch contours and target timbre formant vectors directly onto retrieved tokens.

---

2. Mathematical Specifications

• Sample Rate: $40,000\text{ Hz}$ ($40\text{ kHz}$) mono float.
• Nyquist Limit: Exactly $20,000\text{ Hz}$ ($20\text{ kHz}$).
• Linear STFT Grid: $N_{\text{fft}} = 4,000$ points, hop size $H = 800$ samples ($20\text{ ms}$ temporal resolution, $50\text{ frames/sec}$).
• Frequency Bin Width: $\Delta f = \frac{40,000}{4,000} = 10.0\text{ Hz}$ per bin.
• Linear Band Count: Exactly 2,000 linear frequency bins spanning $0\text{ Hz}$ to $20,000\text{ Hz}$.
• Quantization Format: Single int16_t values representing pure decibel energy (dBFS normalized with $+120\text{ dB}$ offset).
• Noise Floor Threshold: Bins below $+10\text{ dB}$ ($<-110\text{ dBFS}$) are clamped to hard zero (0), guaranteeing silent noise floors.
• Frame Geometry: Exactly $2,000\text{ int16}$ values = $4,000\text{ bytes}$ per frame.

---

3. Dataset Structure & Storage Architecture

rmaapk/music/
├── README.md                           # Dataset Card & YAML metadata
├── data/
│   └── train-00000-of-00001.parquet    # Structured parquet view for Hugging Face Viewer
├── databases/
│   └── acoustic_rag_master.db          # Settled SQLite codebook (8.35 GB, 2,017,500 states)
└── tokens/
    ├── cat_0.bin                       # 2,000-band INT16 spectral frames (Cat 0)
    ├── cat_1.bin                       # 2,000-band INT16 spectral frames (Cat 1)
    └── cat_...bin                      # 1,345 categories live (scaling to 8,000)

Table Schema: settled_codebook

In databases/acoustic_rag_master.db:

CREATE TABLE settled_codebook (
    row_id INTEGER PRIMARY KEY AUTOINCREMENT,
    category_id INTEGER NOT NULL,
    state_idx INTEGER NOT NULL,
    band_payload BLOB NOT NULL,             -- Exactly 4,000 bytes (2,000 single int16s)
    total_samples INTEGER DEFAULT 0,        -- Running sample accumulator
    UNIQUE(category_id, state_idx)
);
CREATE INDEX idx_settled ON settled_codebook(category_id, state_idx);

Each category contains 1,500 modulo cyclic temporal states ($30\text{ seconds}$ cyclic resonance window).

---

4. Citation & Reference

@dataset{rmaa_acoustic_rag_2026,
  author       = {RMAA Holdings},
  title        = {Acoustic-RAG: Deterministic Physical Spectral Codebook for AI Music Synthesis},
  year         = {2026},
  publisher    = {Hugging Face},
  doi          = {10.5281/zenodo.23868153},
  url          = {https://huggingface.co/datasets/rmaapk/music}
}