VOLE-Audio

Procedural sampling and direct audio materialization from deterministic state — a rigorous native-Rust implementation of the VOLE-Audio research architecture, currently specified by v1.1.

de Beer, R. (2026). VOLE-Audio: Procedural Sampling and Direct Audio Materialization from Deterministic State — Entropy-Native Layer and Late-Materialization Architecture (Version v1.1). Zenodo. https://doi.org/10.5281/zenodo.22666746

Original v1.0 broad prior-art disclosure: https://doi.org/10.5281/zenodo.22649073

v1.1 is the current architecture paper — it makes the entropy-native layer explicit (persist the deterministic explanation, entropy-code the exact residual, materialize samples only when an observation requires them), and Phase H.2 of this repository implements the first concrete entropy-native core of that layer. Later phases extend it: Phase O adds learned deterministic prediction, and the speech campaign plus the Phase-6 mechanisms push the exact portfolio past FLAC-8 on the frozen speech corpus. v1.0 remains the earlier broad disclosure.

One Cargo package. PCM is an observation view. SampleObject is authoritative. Literal fallback exists. Everything is measured; nothing is assumed.

The central idea

Persist deterministic state and entropy-coded innovation. Materialize waveform samples only when an observation requires them.

VOLE-Audio's durable representation of an audio object is a deterministic explanation (procedural state — or a literal when no explanation wins) plus a native entropy-coded exact residual for whatever the explanation cannot reproduce, organized as block-addressable pages so any bounded observation decodes only the pages it touches. PCM — the final sample codes a DAC consumes — remains a legitimate observation surface and a literal representation is always the mandatory universal fallback; it is simply not universally privileged as authoritative durable state. The strongest path executed in this repository renders entropy-coded objects on the GPU and writes the exact final sample codes directly into the registered ALSA mmap endpoint region (Phase H/H.2).

Since Phase O the explanation may also be a learned deterministic hypothesis (a quantized predictor proposed by fitting and admitted only after exact residual closure and complete physical-byte accounting), and since the speech campaign the exact portfolio is measured past FLAC-8 on a frozen LibriSpeech-derived corpus. The exact-residual closure contract never changes: the literal fallback remains mandatory, and every candidate is scored by its complete serialized bytes.

What this project is

VOLE-Audio treats pulse-code modulation as a bounded sample-domain materialization requested from persistent deterministic audio state — not as the audio object itself. A SampleObject may be procedural, hybrid, referenced, residual-governed, or literal. Playback is evaluation of a deterministic observation function; the GPU residency and endpoint-mapping experiments (D1/D2) are about pushing the observation boundary as close to the physical DAC as the hardware actually permits.

Non-negotiable reading order (what to know first)

• PCM is an observation view, never the authoritative object.
• SampleObject is authoritative and may be procedural / hybrid / referenced / residual-governed / literal.
• Literal representation is valid and mandatory as the universal fallback. If literal wins the Pareto frontier, that is the result.
• CUDA and ROCm/HIP are the current execution targets; the scalar reference owns semantic authority; the SIMD backend makes the GPU comparison honest.
• D0–D3 name the materialization path: D0 buffered (diagnostic), D1 endpoint-mapped (GPU writes the actual endpoint region), D2 peer-device, D3 endpoint-native. D3 is future conceptual work — it exists in the vocabulary and returns NOT_IMPLEMENTED.
• Directness is measured, never inferred: an mmap exists ≠ D1; a PCIe link exists ≠ D2.
• Negative results are results. Unsupported paths stay visible as UNSUPPORTED_BY_* / INCONCLUSIVE evidence rows. Nothing is silently replaced by an easier buffered path and reported as success.
• Exact is exact: byte/hash equality, not listening. WAV sample-domain reconstruction equality is distinct from WAV container byte equality.
• Every court emits an immutable, self-verifying evidence receipt (vole.audio.evidence.v1) under receipts/.

Repository map

Cargo.toml            one package; std feature gates host-only modules
rust-toolchain.toml   pinned nightly (device/GPU artifact builds only)
src/
  status.rs           the nine verdict classes + NOT_IMPLEMENTED (shared u8 codes)
  limits.rs           hostile-input ceilings (no_std, shared with device)
  hash/               in-repo SHA-256 (FIPS 180-4), no_std
  universe/           vole.audio.u1 exact semantics        (Phase B)
  object/             SampleObject model                    (Phase C)
  sampler/            voices/world/scheduler                (Phase C)
  eval/               scalar + SIMD + shared battery   (Phase C/F)
  entropy/            native rANS + models + pages + residual coding + corpus
                      + EmbeddedStore + DSFB observer       (Phase H.2)
  device/             flat kernel semantics (no_std); nvptx entry live (G);
                      amdgcn entry live (I; code object per gfx target);
                      shared period-scan search primitive (L)
  backend/            flatten (G); cuda/ host runtime live (G) + entropy (H.2)
                      + search (L); rocm/ probe + artifact (I), runtime (J),
                      search (L); entropy_flat (H.2) host flat-job builder
  audio/              ALSA endpoint, directness, topology   (Phase H+)
  format/             canonical `.volea` archive + WAV ingest   (Phase E/N)
  inverse/            bounded inverse-proceduralization + search placement (K/L)
  learned/            learned deterministic prediction: Phase O (exp1), the Exp2
                      addendum, the Exp3 speech/residual mechanisms, and the
                      Phase-6 parsing/representation models (experimental profiles)
  transport/          deterministic framing + receiver + clock  (Phase N)
  evidence/           receipts/counters/timing/environment  (Phase A)
  courts/             executable courts                     (Phase C+; registry in
                      src/courts/mod.rs)
  main.rs             the vole-audio CLI                    (grows by phase)
docs/                 spec + phase charters + evidence + non-claims + CHANGELOG (repo only)
corpus/               frozen flagship corpus, Phase M (repo only)
receipts/             immutable evidence outputs (repo only)
assets/u1/            frozen deterministic tables (resampler, sine)
scripts/              device build + court drivers (repo only)

Package vs repository: the crates.io tarball ships the library, binary, assets/ tables, licenses, and this README. docs/, corpus/, receipts/, scripts/, and rust-toolchain.toml live in the GitHub repository only (they are excluded from the published package to keep it lean and stable-buildable). See the links under Current status and Building for the repository-only material.

Current status

Phases A–N are complete. Phase O (learned deterministic prediction, profile vole.audio.learned.exp1) is complete; the Exp2 addendum imports every Exp1 candidate and adds the v2 residual codecs and mechanism families; the speech campaign (Seals S0–S8, v0.29.0–v0.37.0) took the exact learned portfolio past FLAC-8 on a frozen LibriSpeech-derived corpus; the fourth-pass entropy/adaptation ladder (E0–F0, v0.42.0–v0.53.0) added the signed/FSM range coders, the natural-gradient predictor and BGMC; Phase 6 is complete (v0.54.0–v0.69.0); the Phase-2 metadata/predictor seals (v0.70.0–v0.73.0) added parameter deltas, SRLA coefficient refinement, the widened natural-gradient ladder and Zstd-style integer metadata codes; Phase 7A (entropy-seed proceduralization, v0.74.0) makes the inverse path discover a deterministic Compound explanation blind from samples and persists it as entropy-coded H state plus an entropy-coded exact residual; Phase 7B (v0.75.0) adds vole.audio.lossy.exp1 — VOLE's first real lossy profile — with a matched-actual-bitrate court against external Opus and Lyra; and Phase 7C (v0.76.0) adds vole.audio.stream.voice.exp1, a voice-call profile whose constitution is latency, packet loss and jitter rather than archival bytes.

Measured position (as of v0.73.0, lossless). On the frozen real-speech corpus (LibriSpeech, CC BY 4.0), the exact learned portfolio is:

split	VOLE portfolio	FLAC-5	FLAC-8	record
effectiveness (dev-clean, 8 clips)	126 264 B	130 331 B	129 713 B	7/8 wins vs both
held-out Mode C (test-clean, 8 clips)	136 783 B	144 145 B	142 714 B	8/8 wins

That is a speech-corpus result on this host — not a general-audio flagship claim, and not a real-time/deadline claim.

Lossy position (v0.75.0). vole.audio.lossy.exp1 is a real lossy encoder/decoder profile — two competing explanation engines (forward-LPC with a pitch predictor; MDCT with a Bark masking model), an MSE-optimal absolute-step rate allocation, and VOLE-native entropy coding — measured against external Opus and Lyra at matched actual bitrate. The measured win/loss split, the defects found while getting there, and the attributed remaining losing cells are in PHASE_7B.md. It is not a competitor wrapper: no competitor code is imported, linked or used as a fallback.

Voice-call position (v0.76.0). vole.audio.stream.voice.exp1 is a real streaming profile: 16 kHz mono, 10/20 ms causal frames, absolute per-packet parameters so a lost packet can never desynchronise a later one, bounded reconstructed state, checksummed state capsules, VAD/DTX with procedural comfort noise, and concealment that repeats the reconstructed excitation ring. Court learned-voice-stream measures it on three axes at once — clean quality at matched actual bitrate against external Opus, EVS and Lyra (EVS also driven through its own G.192 erasure flag), quality under a seeded deterministic loss /burst/jitter/drift engine, and one-way latency with every contribution kept separate. The latency constitution passes with margin (encode p99 1.46 ms against a 5 ms budget, one-way p50 31.1 ms on 20 ms frames against a 40 ms envelope, zero deadline misses); clean quality loses, and the loss is attributed to scalar coefficient transmission setting a ~17.7 kbps floor. The full accounting and the declared remainder are in PHASE_7C.md.

Phase 7 is the current program: 7A entropy-seed proceduralization (sealed), 7B vole.audio.lossy.exp1 (sealed), 7C vole.audio.stream.voice.exp1 (this release, with its remainder declared). Phase 7A's evidence is the learned-entropy-seed court: the blind Compound proposer (autocorrelation fundamental estimation, orthogonal matching pursuit over the integer tone atoms the graph can emit with a joint least-squares re-solve and split refinement, harmonic projection, and envelope/onset fitting — no fixture dispatch), entropy-coded H state via the Phase-2B metadata coders, entropy-coded exact residual via the residual-codec family, exact closure, bounded (random-access) materialization, determinism, and the mature learned predictor families competing in the same inverse decision.

Executable evidence today. cargo run -- court <name>. The authoritative court list is the registry in src/courts/mod.rs; the aggregate courts are all (Phase M), h2 (Phase H.2), phase-n (Phase N) and learned (Phase O), and every court emits an immutable receipt under receipts/. The exact seal expectation is checked by vole-audio seal verify --receipts receipts (76 rows at v0.76.0). The full per-phase evidence narrative — the Phase H/H.2 endpoint measurements, the Phase K/L inverse results, and every seal's numbers — lives in CHANGELOG.md.

The exact ledger — completed phases, evidence, blockers, and the next work item — is PROJECT_STATE.md (repository-only). The phase charters (PHASE_H2.md, PHASE_K.md, PHASE_L.md, PHASE_M.md, PHASE_N.md, PHASE_O.md, PHASE_O_EXP2.md) and the normative documents (ENTROPY_NATIVE.md, RANS.md, ENTROPY_ACCOUNTING.md, ENTROPYFS.md, DSFB_SEARCH.md, ADRs 0001–0005) live under docs/. Fixture-level measurements are in PERFORMANCE.md; the spec is U1_SPEC.md, and NON_CLAIMS.md says what this repository does not claim.

Building

cargo build     # host build; needs no CUDA/ROCm/ALSA development files
cargo test      # unit + property + differential tests
cargo clippy --all-targets --all-features -- -D warnings

Toolchain split (documented precisely because it is easy to blur):

• Host builds (the published crate) are stable-Rust: rust-version in Cargo.toml names the minimum stable release this package is verified against. cargo build/cargo test succeed with that stable toolchain and no GPU toolchain installed.
• Device/GPU artifact builds are nightly-only: rust-toolchain.toml in the repository pins the exact nightly used to cross-compile this same package to nvptx64-nvidia-cuda (PTX) and amdgcn-amd-amdhsa (HSA/ELF). The pin is repository-only — it is excluded from the published package so the crates.io tarball never forces nightly on consumers.

The CUDA device artifact (Phase G, live) is produced by cross-compiling this same package via build-cuda-device.sh — one package, no second crate — emitting scripts/out/vole_audio.ptx plus SHA-256 and provenance metadata. court cuda loads that PTX through the CUDA driver API (dlopen'd; nothing links against CUDA at build time). The ROCm artifact is produced by the Phase I script build-rocm-device.sh (repository-only; active in Phase I). GPU support is opt-in and runtime probed; CPU-only machines run every non-GPU court unchanged.

Evidence constitution

Every claim this repository makes is backed by an immutable receipt. Run vole-audio probe to see the environment capture and vole-audio receipt show <file> to verify a receipt's self-hash. Each receipt records its git commit, a source-tree hash, and a dirty-state that excludes receipt-output writes; the seal subject it binds is every tracked source file except receipts/, target/, scripts/out/, docs/ and .git/, so committing evidence or documentation cannot invalidate a seal while any code change does. See the EVIDENCE.md measurement-boundary notes (repository-only).

The court list is not fixed in this README: the authoritative registry is COURT_NAMES in src/courts/mod.rs, and the newest receipt per court is what seal verify checks. Courts arrived with their phases — A–M (oracle, SIMD, CUDA/ROCm, inverse, conventional and runtime courts), N (archive, transport, phase-n), O (learned-* under vole.audio.learned.exp1), and the Exp2/Exp3/Phase-6 courts under the experimental learned profiles. The exact per-seal matrix and every frozen result hash are recorded per receipt and summarized in CHANGELOG.md.

Non-claims

See NON_CLAIMS.md (repository-only). In particular: the speech-corpus result above is exactly that — a frozen-speech-corpus result on this host, not a general-audio flagship claim, and not a real-time/deadline claim. Fixture-level measurements in PERFORMANCE.md are labeled as fixture-level. The entropy phase adds its own non-claims: rANS is never presented as a generator, the DAC does not "consume compressed audio", D1 is not a latency optimization in the Phase-H court, EntropyFS/DSFB are optional and never enter playback, and the GPU never owns semantic authority (scalar does). Learned prediction is an experimental candidate family, never truth: a learned object participates only after canonical quantization, exact residual closure and complete dependency accounting, and the literal fallback stays mandatory. This repository is hostile to self-deception by design.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE / https://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT / https://opensource.org/licenses/MIT)

at your option.

The VOLE-Audio papers — v1.1, the current architecture paper (DOI 10.5281/zenodo.22666746), and v1.0, the original broad prior-art disclosure (DOI 10.5281/zenodo.22649073) — remain separate works with their own disclosure terms.