Octa-Initium: A Hierarchical, Physics-Informed Architecture for Full-Song Generation

Model at a glance

Introduction & design principles

A request such as “a dark trap banger with heavy 808s” implies decisions at radically different timescales: global form, harmonic and rhythmic language, timbral palette, lyric content, a sung human-like voice executing that lyric with correct phonetics and expressive pitch, and finally a mix in which all of these coexist at commercial loudness.

Initium's central architectural commitment is factorization. Most properties documented in this card — controllability, lyric fidelity, long-form coherence, auditable safety behavior — follow from it.

2.2Design principles

• Plan, then render. Music-theoretic structure is decided symbolically before any audio is sampled; the acoustic stage is never asked to invent form (§3.3).
• Continuous where sound lives, discrete where structure lives. Structure and semantics are discrete token sequences; audio is generated on a continuous latent manifold by flow matching, avoiding quantization artifacts in the signal path (§3.4).
• The voice is not just another instrument. Sung vocals are produced by an expert branch whose inductive biases come from the physics of human voice production (§4.2).
• Provenance by construction. Every rendered waveform carries an imperceptible watermark; every cover-mode generation is screened against a copyrighted-work fingerprint index before delivery (§8).
• Report honestly. Where Initium underperforms strong commercial systems, the evaluation section says so (§7.4).

2.4Model family & variants

Three checkpoints share one architecture and one tokenizer/codec pair, differing in acoustic-stage depth, sampling schedule and maximum planned duration. They are selected server-side per request class.

2.6Non-goals

Initium is not a general audio model: it does not perform speech-to-text, environmental sound synthesis, or music source separation as user-facing tasks — separation is used internally for data preparation only (§5). It does not attempt to imitate identifiable living artists; requests of that form are abstracted to genre-level style descriptors by the safety layer (§8.2), a deliberate capability subtraction.

System architecture

Initium factorizes song generation into three learned stages around one shared audio representation. The vocal expert lives inside the acoustic stage rather than as a post-hoc synthesizer, so voice and accompaniment share one latent timeline and remain phase- and groove-consistent.

3.1Representations

Three views of the same signal appear throughout this report. One 4:30 song at 43 Hz spans ≈ 11,600 latent frames; the semantic stream spans ≈ 3,375 tokens; the composition graph averages 610 symbolic nodes ^◆.

3.2OctaCodec — the shared latent space

A convolutional–transformer autoencoder maps 48 kHz stereo to latent z through five strided residual stages followed by a bidirectional transformer bottleneck; the decoder mirrors it with a HiFi-GAN-style multi-band head. A KL-regularized bottleneck keeps the manifold smooth for the flow stage. Stereo consistency is enforced by computing mel losses on mid/side channels separately — naive L/R treatment collapses image width, an audible failure mode.

3.3The plan is the interface

The planner emits a typed symbolic object under an attribute grammar. Grammar-constrained decoding masks productions that would violate the schema, so every sampled plan is well-formed by construction: musical quality of plans is learned, their validity is guaranteed. Eight candidate plans are drawn per request and a learned critic scores form balance, harmonic rhythm and lyric-meter fit.

// prompt: "a dark trap banger with heavy 808s" · seed 7141
KEY       F#minor    TEMPO 142 bpm (swing 54%)    METER 4/4
FORM      INTRO[4] > VERSE.A[16] > HOOK[8] > VERSE.B[16] > HOOK[8] > BRIDGE[8] > HOOK.x2[16] > OUTRO[4]
HARMONY   VERSE.A: i - i - VI - VII  |  HOOK: i - VI - III - VII (pedal F#1 under 808)
ROLES     KIT-808{gliss:on, tune:F#1} · HAT{ratchet:1/32} · PLUCK-LEAD{scale:harm.minor}
          PAD{entry:HOOK} · VOX-LEAD{range:A2–E4} · VOX-STACK{HOOK, +5th, wide}
LYRIC.MAP VERSE.A: 8 lines · 11–13 syl · rhyme AABB → anchors L1 ↘ L2 ↗ …
// 612 nodes / 1841 edges total — layers elided for brevity

Because downstream stages treat the graph as ground truth, user-facing controls are graph edits rather than prompt tricks: locking a chord progression, pinning the hook melody while regenerating verses, changing the tempo curve, swapping one section's orchestration. Stage B never sees raw user text — all natural-language interpretation is quarantined in the planner, which makes the acoustic stage auditable (plan-render consistency 0.93 Spearman ^◆).

3.4Acoustic generator — conditional flow matching

Flow matching trains a continuous normalizing flow by regressing a time-dependent velocity field, giving diffusion-class sample quality with an order of magnitude fewer sampling steps — decisive for interactive generation.

Along the optimal-transport displacement path z_t = (1 − (1 − σ_min)t)·z₀ + t·z₁, with logit-normal time sampling and SNR-derived weighting that concentrates capacity in the mid-path region where structure crystallizes. Generation solves the probability-flow ODE from t = 0 to 1.

Velocity regression is decomposed across a multi-scale DCT filterbank and reweighted by the ear's own frequency resolution — bass must be phase-solid while air bands stay stochastic. In ablation this is the single largest contributor to low-end solidity (Δ 808-punch MOS +0.31 ^◆).

3.4.5Serving profile

Vocal synthesis subsystem

Sung voice is the harshest quality bar in generative music: listeners possess dedicated neural circuitry for voice, so errors that would pass in a synth pad — a 20-cent pitch bias, smeared consonants, static vibrato — are immediately salient. Initium's vocal subsystem is built around an explicit model of how humans produce voiced sound, rather than a free-form network asked to rediscover physiology from data.

A glottal source in the Liljencrants–Fant parameterization — open quotient, return-phase constant, asymmetry — shaped by four cascaded formant resonators and lip radiation. The singer's-formant cluster at 2.8–3.4 kHz is exposed as a controllable scalar: the parameter behind “cutting through the mix”.

4.3Pitch as a stochastic process

Human singing pitch is neither the score nor noise around it: it is a structured dynamical process — glides into notes, drift-and-correct around targets, quasi-periodic vibrato whose rate and extent breathe with phrase energy.

Ornstein–Uhlenbeck attraction to the melodic target, an explicit vibrato oscillator with slowly diffusing phase (5–7 Hz, onset-delayed), and a micro-tremor diffusion term. Coefficient trajectories are emitted from phrase context, style tags and voice identity — a drill verse gets tight, near-flat pitch; a power-ballad chorus gets ±60–90 cent vibrato blooming after note onset.

4.4Timbre identity & consent-gated cloning

Voice identity is a point in a timbre space trained contrastively on singing — not speech — augmented against pitch, key, tempo and phonetic content, so identity is disentangled from what is being sung. Enrollment aggregates 2–5 minutes of audio with quality gating on SNR and pitch-range coverage.

4.5Intelligibility is a mix problem

A vocal can be perfectly synthesized and still unintelligible in situ; masking by co-generated instruments is the dominant cause. Because Initium generates voice and accompaniment jointly on one latent timeline, it shapes both sides of the masking equation rather than fixing collisions in post. A differentiable proxy of simultaneous masking is penalized on lyric-bearing frames, with margins largest in the 2–5 kHz consonant-critical region — visible in spectrograms as ≈ 1.5 dB dynamic dips at vocal onsets, i.e. learned side-chain behavior.

Sung-lyric word error rate in-mix, transcribed by humans: 8.9 % EN / 11.2 % TR ^◆, against 15–19 % for strong commercial baselines under identical protocol. Isolated-vocal WER of 5.1 % indicates most residual loss is stylistic — melisma, ad-libs — not masking. Dense screamed styles and deliberately lo-fi aesthetics reduce intelligibility by design; the app exposes a lyric-clarity control mapped to the masking weight and vocal-level prior.

Training data

The audio corpus is assembled exclusively from sources with documented rights chains. We do not train on audio scraped from streaming services or video platforms, and we operate an artist opt-out registry honored retroactively at each training cycle.

Hours-weighted shares after filtering. Absolute scale: 2.3 M raw hours ingested → 612 K hours retained for pre-training; 140 K hours vocal-isolated material; lyric text licensed separately ^◆.

Training procedure

Training ran on a dedicated cluster of H-class accelerators ^◆ with 3D parallelism (tensor, sequence, data) and ZeRO-sharded optimizer state; total pre-training budget on the order of 3 × 10²³ FLOPs ^◆. Numerics: bfloat16 activations with fp32 master weights, and fp32 attention logits in the planner's long-context windows. All stages use AdamW with warmup–cosine schedules and gradient clipping at global-norm 1.0; the flow stage maintains an EMA of weights (decay 0.9999) used for all evaluation and serving.

Post-training runs three passes: supervised fine-tuning on 92 k curated request → render exemplars authored by producers and lyricists in five languages, teaching prompt-register — what users mean by “darker”, “more space”, “radio-ready”; preference optimization on ≈ 1.1 M pairwise judgments ^◆ drawn from production A/B listening; and safety fine-tuning against the policy taxonomy of §8.

Evaluation

All systems are evaluated on OctaBench-600: 600 prompts stratified over 20 genres × 5 languages × 3 structural demands, authored blind to any system's strengths and frozen before model training completed. Comparators are three strong commercial full-song systems, anonymized here per their terms of service. All outputs are loudness-normalized to −14 LUFS. Human studies use 210 paid, music-literate raters balanced across the five languages, each item rated by at least nine raters, randomized and blinded; we report 95 % bootstrap confidence intervals.

7.4Where comparators win

Blind studies also localize current deficits. Publishing them is card policy: unreported deficits become user-discovered deficits.

Safety & provenance

Every render traverses five layers; layer verdicts are logged to the audit trail referenced by the render manifest.

Watermark detection is a correlation test with per-band whitening over an inaudible spread-spectrum mark in the mid-band mel domain, carrying a 96-bit payload (render-manifest digest plus policy flags). AUC 0.997 clean and 0.981 after re-encoding with ±2 % pitch shift. Robustness holds under MP3/AAC re-encode, resampling and moderate EQ; gradient-based removal attacks degrade audible quality before detection fails.

Quality parity is tracked across genre × language cells; the max-min MOS gap is 0.38 ^◆, worst cell Hindi-language dense-arrangement pop, which drives the commissioned-data program of §5. Rater panels are language-balanced to avoid Anglophone quality bias in the headline numbers.

Limitations & intended use

Initium is deployed as a consumer creative tool: original-song creation, self-voice covers, lyric-driven composition and social sharing. Appropriate professional uses include sketching, demo production and content soundtracks under the app's licence terms. The model is not offered for impersonation of real artists, voice cloning of anyone but the consenting account holder, unlicensed commercial replication of protected works, or as a raw API without the deployment stack of §8 — it is not distributed separately from that stack.

Release notes & citation

10.1Citation

@techreport{octaverum2026initium,
  title       = {Octa-Initium: A Hierarchical, Physics-Informed
                 Architecture for Full-Song Generation},
  author      = {{Octaverum Research}},
  year        = {2026},
  number      = {OCT-MC-2026-001},
  institution = {Octaverum},
  note        = {Model card and technical report, rev. 1.0}
}

10.2References

1. Vaswani, A. et al. Attention Is All You Need. NeurIPS, 2017.
2. Ho, J., Jain, A., Abbeel, P. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.
3. Lipman, Y. et al. Flow Matching for Generative Modeling. ICLR, 2023.
4. Peebles, W., Xie, S. Scalable Diffusion Models with Transformers. ICCV, 2023.
5. Zeghidour, N. et al. SoundStream: An End-to-End Neural Audio Codec. IEEE TASLP, 2021.
6. Défossez, A. et al. High Fidelity Neural Audio Compression. TMLR, 2023.
7. Kong, J., Kim, J., Bae, J. HiFi-GAN. NeurIPS, 2020.
8. Fant, G. Acoustic Theory of Speech Production. Mouton, 1960.
9. Fant, G., Liljencrants, J., Lin, Q. A four-parameter model of glottal flow. STL-QPSR, 1985.
10. Sundberg, J. The acoustics of the singing voice. Scientific American, 1977.
11. Moore, B.C.J., Glasberg, B.R. Auditory-filter bandwidths. JASA, 1983.
12. ITU-R BS.1770-4. Loudness and true-peak level. 2015.
13. Kilgour, K. et al. Fréchet Audio Distance. INTERSPEECH, 2019.
14. Wu, Y. et al. Large-scale Contrastive Language-Audio Pretraining. ICASSP, 2023.
15. Loshchilov, I., Hutter, F. Decoupled Weight Decay Regularization. ICLR, 2019.
16. Rafailov, R. et al. Direct Preference Optimization. NeurIPS, 2023.
17. Mitchell, M. et al. Model Cards for Model Reporting. FAT*, 2019.
18. Gebru, T. et al. Datasheets for Datasets. CACM, 2021.