Audit the file. See the evidence. Understand the verdict.
A cross-platform audio integrity, fidelity, and comparison workstation powered by the Oracle Engine.
Download · Why Audio-V · First audit · Oracle Engine · Validation story · Verdicts · Feature tour · Identity setup · User guide
Audio-V is a desktop application for people who want to inspect music and audio files without becoming an audio engineer first.
Choose one track, a group of tracks, or a whole music folder. Audio-V reads each selected audio stream from beginning to end, measures what it finds, and gives the file a carefully limited verdict:
- Clear — the file passed the checks Audio-V performed.
- Review — the scan finished, but something deserves a closer look.
- Failed — a repeatable integrity check found evidence against the file.
Audio-V is not a music player. It does not judge whether a song is good, and it does not assume that a large bitrate or “Hi-Res” label means high quality. It is an inspection desk: decode the file, measure the signal, show the evidence, and explain the result.
Audio-V has two paths through the same Oracle Engine:
- Quick Inspect is the lightweight path. Choose one file and get its verdict, format, encoded MP3 channel mode, signal, spectrum, integrity, and next step on one screen.
- Deep inspection adds adjustable spectrograms, evidence lanes, comparison, batch processing, reports, history, and collection tools when they are needed.
Quick Inspect does not run a weaker scan or invent a simplified verdict. It summarizes the same complete-stream analysis used by the deeper views. Audio-V is therefore lightweight to use for one file, even though its total scope is larger than a single-purpose spectrogram viewer.
- Music collectors checking a personal library
- Archivists organizing and verifying audio
- DJs and creators checking delivered files
- People comparing an original track with an edited or repaired copy
- Curious listeners learning what technical audio information means
- Developers who need repeatable JSON evidence from a command line
No audio expertise is required. Technical details are available when wanted, but the main workflow is designed to answer three simple questions:
- Could Audio-V read the whole file correctly?
- Did it measure anything that deserves attention?
- What evidence produced the verdict?
Important
Audio-V is an unsigned open-source release still completing its broader release-candidate acceptance matrix. A Clear verdict means the file passed the disclosed checks in this Oracle Engine version. It does not certify an original studio master, prove ownership, or guarantee that the audio was never transformed.
Many audio tools answer one part of the question. A spectrogram viewer shows frequency energy. A metadata utility reports what the container declares. A CD verifier compares a properly assembled rip with a disc database. Audio-V connects those kinds of evidence into one non-destructive library workflow and explains what each result can—and cannot—establish.
| Audio-V strength | What it gives you |
|---|---|
| Complete-stream assessment | The selected audio stream is decoded from beginning to end instead of judging a filename, header, or short visual sample |
| Fast first answer, deep evidence behind it | Quick Inspect puts the verdict, format, encoded MP3 channel mode, signal, spectrum, integrity, and next steps on one screen; every value links back to the same deeper evidence |
| Narrow, explainable verdicts | Clear, Review, and Failed follow disclosed precedence; measurements, heuristics, declarations, and deterministic checks remain visibly different |
| Library-scale operation | Recursive folders, bounded workers and fingerprint finalization, mounted-source staging, pause, prompt cancellation, checkpoints, adaptive recovery, cache reuse, and compact SQLite history |
| Deep visual inspection | Measured waveform and 512–16,384-point spectrogram views with channel isolation, L−R, zoom, pan, regions, scientific palettes, and batch image export |
| Scientific A/B comparison | Independent file loading, synchronized visual modes, alignment, gain and polarity estimation, explicit channel mapping, and full-overlap decoded null evidence |
| Integrity before speculation | Full decoding, SHA-256 identity, FLAC audio-MD5 verification, external checksum manifests, failure stages, and exact evidence |
| Signal and delivery diagnostics | LUFS, LRA, true peak, clipping events, click/pop and stuck-sample candidates, dropout, DC offset, dynamics, bit utilization, ReplayGain, and optional delivery profiles |
| Identity without verdict inflation | Local Chromaprint relationships plus optional AcoustID and MusicBrainz context remain separate from audio-quality verdicts |
| Evidence you can keep | Per-file inspection, human review disposition, resumable history, privacy-safe acceptance records, JSON automation, and PDF/XLSX/DOCX/CSV reports |
| Local-first design | No player, account, advertising, telemetry, automatic upload, or source overwrite |
Audio-V was not built by choosing a few spectral thresholds and calling the result authoritative. Oracle Engine has to earn every stronger claim.
The project assembled 85 legally acquired, independently tracked reference recordings from five external datasets. From those references, Audio-V created 2,295 controlled and observational cases covering native audio, known lossy transcodes, upsampling, intentional low-pass filtering, defects, speech, narrow-band material, and other difficult cases. Related versions of one recording stay in the same source group so Oracle cannot appear accurate by seeing a close relative of its test material during development.
Candidate rules move through separate development, calibration, frozen test,
external challenge, and edge-case populations. Promotion requires measured
sensitivity and strict protection against false advisories. Experimental
scores cannot be presented as probabilities, cannot produce Failed, and
cannot turn a file into Clear merely because a model likes it.
That policy has already stopped a seemingly promising idea from reaching users. A multivariate candidate performed strongly on controlled material but misidentified too many legitimate external music, speech, and narrow-band recordings. Audio-V published the failed result and left Oracle v12 unchanged. In other words, the validation system did its job: it preferred an honest Inconclusive result over a confident but unreliable story.
Every release also runs the complete production corpus, unit and integration tests, format fixtures, packaged-runtime checks, accessibility checks, and release-artifact verification. Audio-V does not claim that heuristics can prove a studio master or reconstruct an unknowable recording history. It claims something narrower and more useful: every verdict is bounded by disclosed evidence, and stronger rules must survive independent tests before they can affect that verdict.
Read the full, reproducible record in the validation story, Oracle calibration log, and machine-readable promotion policy.
Audio-V does not market any spectrogram as “100% raw” or as proof of an untouched studio master. It decodes the selected audio stream, calculates a short-time Fourier transform (STFT), and discloses the FFT size, window, channel mode, frequency range, display floor, scale, and palette. Those settings affect how every spectrogram looks; exposing them makes Audio-V’s view measurable and reproducible instead of mysterious.
A spectrogram can reveal bandwidth and cutoff patterns. It cannot, by itself, prove the complete history, authenticity, ownership, or artistic intent of a recording. Oracle combines it with integrity and signal evidence and labels uncertain origin findings as Review—not fact.
| If you are… | Start with… | Go deeper with… |
|---|---|---|
| A curious listener | Plain-language verdict and “what to do next” guidance | Spectrogram, loudness, and glossary |
| A music collector | Folder audit, verdict sorting, duplicates, and history | Checksums, origin evidence, cue context, and identity services |
| An audiophile | Effective bandwidth, bit utilization, clipping, true peak, and channel behavior | High-resolution channel spectrograms, null comparison, and delivery profiles |
| An archivist or librarian | Recursive ingest, metadata inventory, hashes, reports, and resumable sessions | External manifests, persistent fingerprints, CLI JSON, and acceptance evidence |
| A producer, DJ, or engineer | Loudness, dynamics, defects, Compare, and safe-copy remediation | Region comparison, per-channel diagnostics, ReplayGain, and profile-specific delivery checks |
This is positioning, not a claim that one tool replaces every other:
| Tool | Its published center of gravity | Audio-V’s different job |
|---|---|---|
| Spek | Fast, approachable single-file spectrogram viewing and image export | Offer an equally direct one-file starting point through Quick Inspect, then extend it to integrity, signal, origin, provenance, delivery, comparison, and reportable evidence |
| Sonic Visualiser | Highly configurable visualization, annotation, layers, playback, and plugin research | Provide a guided audit and decision workflow without requiring plugin selection or becoming an annotation environment |
| AudioAuditor | Broad Windows toolkit combining analysis with playback, EQ, lyrics, metadata tools, and experimental detection | Stay cross-platform and assessment-focused, keeping every verdict reproducible and every uncertain claim explicitly limited |
| CUETools/CTDB and AccurateRip | Specialized verification of properly contextualized audio-CD extractions against shared databases | Assess ordinary files and libraries broadly; identify disc-verification eligibility without pretending eligibility is a database match |
Audio-V is deliberately not the player, tag editor, or mastering suite in this list. Its job is to be the evidence desk you can trust before deciding whether a file is clear, needs review, should be replaced, or deserves a closer comparison.
Every download below points at the latest release, whatever that currently is. The version badge at the top of this page reports it live.
| Your computer | Download | Asset to pick | What you receive |
|---|---|---|---|
| Apple Silicon Mac | Latest release | Audio-V-<version>-mac-arm64.dmg |
Native build for M-series Macs |
| Any supported Mac | Latest release | Audio-V-<version>-mac-universal.dmg |
Apple Silicon and Intel in one package |
| Windows | Latest release | Audio-V-<version>-win-x64.exe |
Guided installation |
| Windows portable | Latest release | Audio-V-Portable-<version>-x64.exe |
Runs without installation |
Installer filenames carry their version, so this table names the asset to pick rather than linking each file directly. A direct link would break on the next release and quietly serve a 404 to anyone arriving from an older page.
These builds are not signed with paid Apple or Microsoft certificates. Verify
the download with
SHA256SUMS.txt,
then follow the
unsigned installation guide. The
UNSIGNED_RELEASE_MANIFEST.json
records the signing policy and artifact identities. Those two filenames carry no
version, so both links always resolve to the current release.
Open Audio-V and select:
- Choose files for one or more tracks; or
- Choose folder for a folder and its subfolders.
Audio-V can work with local folders, user-selected removable drives, mounted macOS volumes, Windows mapped drives, and UNC network paths.
- Full Oracle audit decodes and assesses the audio. Use this when you want a verdict.
- Metadata inventory quickly catalogs declared properties without decoding the signal. Its result is Not analyzed, because no audio verdict was attempted.
In Settings, you may also select an optional delivery profile for future full audits: EBU R 128 programme QC, ATSC A/85 delivery, or AES internet-music track normalization. Audio-V stores the exact profile and limits with the session. A miss becomes Review, never Failed; with no profile selected, loudness and true peak remain measured evidence only.
Completed files appear while the audit continues. Progress remains below 100% until file analysis, fingerprint linking, and durable history finalization are finished.
You can pause the scheduling of new files or cancel the audit. Completed results are checkpointed, and interrupted scans can be resumed from History.
Select a file and Quick Inspect opens first. It places the one-file verdict, plain-language explanation, format, signal, spectrum, integrity, and next-step summary on one screen. Its buttons open the full Spectrogram, Loudness, and Evidence views; it summarizes the existing Oracle result and never issues a different verdict.
The deeper views include:
- the plain-language Oracle explanation;
- the evidence lane that raised a concern;
- hashes and integrity results;
- signal, loudness, clipping, and defect measurements;
- Origin Assessment and its limitations;
- provenance and acoustic-identity clues;
- measured waveform and spectrogram views.
For MP3 files, MP3 frame mode is read directly from MPEG audio frame headers and reported as Stereo, Joint Stereo, Dual Channel, or Mono. This is kept separate from the decoded output channel layout.
- For Clear, save a report if you need a record.
- For Review, inspect the named evidence and compare with a trusted copy when possible.
- For Failed, preserve the source and obtain or compare another copy before deleting anything.
- For Analysis error, read the failure stage and retry; Audio-V has not called the file damaged.
The Oracle Engine is Audio-V’s versioned inspection and verdict system. Think of it as a line of specialists. Each specialist checks one part of the file, then Oracle brings their evidence together without pretending that every clue is proof.
flowchart LR
A[Discover] --> B[Identify]
B --> C[Decode all audio]
C --> D[Verify integrity]
D --> E[Measure signal]
E --> F[Inspect spectrum]
F --> G[Inspect provenance]
G --> H[Five evidence lanes]
H --> I[Explain verdict]
- Discover — find supported audio and normalize its format.
- Inventory — read declared properties and metadata.
- Identify — calculate exact SHA-256 file identity.
- Decode — read the selected audio stream from beginning to end.
- Verify — check decode completion, FLAC audio MD5, duration, and supplied checksum manifests.
- Measure — calculate loudness, peaks, clipping, continuity, dynamics, channel behavior, waveform, and spectrum.
- Assess origin clues — look for conservative multi-feature patterns compatible with possible transcoding or upsampling.
- Inspect provenance and identity — examine C2PA statements, generator indicators, MQA-related declarations, CD-database eligibility, and Chromaprint relationships.
- Organize evidence — keep integrity, signal, origin, provenance, and delivery findings in separate lanes.
- Issue and explain the verdict — apply narrow rules and show exactly which evidence mattered.
- Finalize — link bounded library relationships and commit resumable history before showing 100%.
Read The Oracle Engine, explained simply for a complete walkthrough of every stage, rule, threshold, limitation, and sensible next step.
| Evidence | Meaning | Example |
|---|---|---|
| Deterministic | A repeatable yes/no check | A FLAC audio MD5 matches or does not |
| Measured | A number or behavior observed in decoded audio | Loudness is −14 LUFS |
| Heuristic | Several measurements match a disclosed pattern | Possible prior lossy encoding |
| Declared | Metadata or a signed statement says something | A tag names a generator |
This labeling is important. A heuristic clue is never silently presented as deterministic proof.
| Lane | The question it answers |
|---|---|
| File integrity | Could the whole file be read and verified? |
| Signal defects | Did the decoded sound contain material warning patterns? |
| Spectral origin | Does its frequency shape resemble a prior transformation? |
| Provenance | Are there attached claims, generator clues, or identity relationships? |
| Delivery compliance | Does it fit a selected delivery target? |
The lanes prevent unlike issues from being mixed together. Missing provenance does not mean broken audio. A loud master does not mean a corrupt file. A possible transcode pattern does not mean the decoder failed.
Clear means:
- the complete selected audio stream decoded;
- deterministic integrity checks did not fail; and
- no current finding reached Review or Critical severity.
Clear files may still contain advisory observations. Clear does not mean “original master,” “perfect sound,” or “never edited.”
Review means the audit completed, but at least one material finding deserves a person’s attention. Examples include:
- strict decode failed but a tolerant confirmation decode succeeded;
- material clipping or possible scaled-clipping patterns;
- repeated or large click/pop candidates;
- repeated or long stuck-sample candidates;
- multiple internal digital-silence dropout candidates;
- large DC offset;
- a material duration mismatch in a sample-exact format;
- an external checksum-manifest mismatch; or
- a strong, multi-feature possible-transcode or possible-upsample pattern.
Review is not the same as damaged. Music production, synthesis, old recordings, hard edits, analog transfers, and intentional mastering can create similar measurements.
Failed is intentionally rare. It currently requires deterministic evidence such as:
- strict and tolerant complete decodes both failing with recognized corruption evidence; or
- decoded native FLAC audio not matching its stored STREAMINFO audio MD5.
These are workflow states, not verdicts:
- Analysis error means a tool, resource, source connection, worker, or internal stage prevented completion. The file is not labeled damaged.
- Not analyzed means Metadata Inventory ran without decoding the signal.
Human review is recorded beside the Oracle verdict, never on top of it. A reviewer can mark a finding as acknowledged, accepted as intentional, confirmed, a possible false positive, remediated, needing replacement, or needing follow-up, and can add a short note. This disposition does not change the file, erase evidence, or turn Review or Failed into Clear.
Audit shows the active queue, sortable verdict results, library distribution, selected-file summary, Oracle explanation, measured evidence, and progress.
Audio-V accepts 39 FFmpeg-backed audio and container extensions. Examples include FLAC, WAV, AIFF, ALAC, MP3, AAC/M4A, OGG, Opus, WavPack, APE, DSF, Matroska/WebM, WMA, AC-3, and E-AC-3.
Audio-V reports container and codec, sample rate, bit depth, channel count and layout, duration, and average bitrate. For MP3 it additionally parses the frame headers directly for MPEG version, layer, and the encoded channel mode, so a joint-stereo file is not reported as plain stereo.
Bitrate mode is measured rather than read from a declared field. Audio-V probes every audio packet in the stream, builds the per-packet bitrate distribution, and classifies from its spread:
- packet count, and how much of the stream's duration those packets cover;
- minimum, maximum, and mean per-packet bitrate;
- the 5th and 95th percentiles, and the standard deviation;
- CBR when the P05–P95 spread and the coefficient of variation are both within 3% of the mean, otherwise VBR.
Two consequences worth knowing. First, this works for any codec FFmpeg can demux, not only MP3, so AAC, Opus, and Vorbis get a real bitrate-mode answer instead of a shrug. Second, a file whose header or tags advertise one mode while the packets say otherwise is reported as what the packets actually are. The underlying distribution is exported alongside the verdict, so the classification can be checked rather than taken on trust.
Bitrate mode appears next to the bitrate in Quick Inspect and on the technical detail panel, and is included in exported reports. When packet probing is unavailable for a stream it reads "Not probed" rather than guessing.
Choose 512, 2,048, 4,096, or 16,384-point inspection; combined, left, right, or L−R channel views; scientific color maps; zoom, pan, region readout; and single or batch PNG export.
The spectrogram is evidence, not decoration—but one picture alone does not prove source history.
Audio-V measures:
- EBU R128 integrated loudness and loudness range;
- true peak, sample peak, RMS, and crest factor;
- clipping percentage, per-channel counts, grouped events, and timeline;
- click/pop, stuck-sample, dropout, and steep-transition candidates;
- DC offset, stereo correlation, dual mono, and near mono;
- DR descriptors and integer bit utilization;
- track ReplayGain and eligible album ReplayGain.
Chromaprint relationships work inside the current audit and in a persistent historical Identity library. Browse, rebuild, prune missing paths, clear the index, or run Identify current audit to recognize already measured files without decoding them again. Optional AcoustID lookup sends a fingerprint and rounded duration—not the audio—to the external service. Requests are serialized below the service limit, repeated fingerprints are cached, and temporary server failures receive bounded retries. Official builds can include Audio-V's registered client identity; source builds and forks can supply their own application key as an operating-system-protected per-user override. A separate optional MusicBrainz enrichment step can use an embedded recording ID or the strongest AcoustID candidate to add credited artists, ISRCs, first-release dates, and release-group context. Audio-V then produces a parallel identity result: Metadata corroborated, Identity matched, Metadata conflict, or Inconclusive. The result shows which declared fields agree with the external identity, but neither service changes the Oracle audio-quality verdict.
Want to use these optional services? Follow the beginner-friendly AcoustID and MusicBrainz setup guide. It explains which key to create, what each switch does, what data leaves the computer, and how to interpret or troubleshoot a match.
Load File A and File B directly from disk or from the current audit. Audio-V compares hashes, formats, loudness, waveforms, spectra, offset, gain, polarity, and a complete overlapping decoded null test when channel mapping permits. Large audits keep their result list compact; when a comparison file is chosen, Audio-V restores its full waveform and spectrum evidence on demand and shows a clear loading or error state instead of an unexplained empty panel. Stacked, overlay, movable wipe, and automatic blink views share one zoom and horizontal position. The aligned residual has its own waveform and spectrogram, and a chosen time region can be re-measured for local correlation, gain, residual level, and peak evidence.
Compare has no playback component. It is built to answer “how are these files related?” with measurements.
Repair is a triage workspace. It explains whether replacement, comparison, manual verification, acknowledgement, or optional mitigation makes sense.
The current automated true-peak action:
- creates a separate FLAC working copy;
- never overwrites the source;
- preserves a known 16-bit or 24-bit source depth by default;
- supports explicit 16-bit or 24-bit output;
- preserves supported metadata and artwork; and
- audits the new copy.
It reduces level to avoid new positive true peaks. It does not reconstruct samples already lost to clipping.
Open a complete per-file evidence report or export a full audit as PDF, XLSX, DOCX, CSV, or lossless JSON.
Reports can contain filenames, full paths, hashes, metadata, and measurements. Review them before sharing.
Audio-V checkpoints completed records in local SQLite history. After an unexpected interruption, adaptive resume:
- reuses completed records that still match;
- safely validates files that were active during the interruption;
- restores the saved worker profile after that isolated check; and
- quarantines a repeatedly unstable file as Analysis error rather than crashing the remaining audit or calling the file damaged.
History can also be cleared after confirmation when no audit is running. The list clears immediately, while large saved sessions are reclaimed in small, restart-safe background batches so the app stays responsive. Your audio files, reusable Oracle cache, and separate Identity fingerprint index are left alone. Settings shows live and reclaimable database space. Legacy databases can run one isolated compaction; new databases return free pages incrementally.
The command line uses the same Oracle worker, verdict rules, and JSON evidence model as the desktop app:
npm run cli -- /music/archive \
--output audit.json \
--concurrency 2 \
--memory-mb 256 \
--ffmpeg-threads 2 \
--native-memory-mb 1024 \
--fail-on failedUse --metadata-only for inventory, --fail-on review for a stricter policy,
or --fail-on never when evidence should never return policy exit code 2.
Desktop packages include a CLI launcher and their bundled Oracle tools.
Audio-V does not:
- play music or manage a listening queue;
- prove that a file is an original studio master;
- rate musical taste or mastering style;
- treat high sample rate, bit depth, or bitrate as automatic quality;
- recover information removed by lossy encoding or clipping;
- identify every possible defect;
- issue a universal “AI generated: Yes/No” badge;
- upload selected audio or collect telemetry.
The project has deterministic synthetic regression fixtures and a complete external-corpus framework. The first acquisition plan covers Slakh2100, MUSAN, MAESTRO, EBU SQAM, and MUSDB18-HQ without requiring private studio masters or maintainer-created recordings. Dataset audio remains local and license-governed; the repository stores official source records, source-level partitions, controlled transformation recipes, checksums, and aggregate scorecards. Audio-V does not publish real-world false-positive or false-negative rates until enough independent external references and a held-out challenge set satisfy the documented thresholds.
The measured corpus now contains 85 independent references across MUSAN, Slakh2100 Redux, EBU SQAM, MAESTRO, and MUSDB18-HQ, producing 2,295 controlled cases. The public population reaches the documented 50-reference target, while MAESTRO, EBU SQAM, and MUSDB18-HQ remain local, nonredistributable challenge sets. Results show zero false lossy-origin advisories across 975 eligible negative cases and zero false upsample advisories across 1,495 eligible negative cases. They also show that Oracle’s current positive-origin recall remains unacceptably low, so Audio-V publishes that limitation rather than presenting the corpus as proof of universal provenance accuracy.
A source-separated multivariate experiment subsequently evaluated 1,870
origin-candidate records across all 85 source groups, including untouched
Slakh test groups, EBU SQAM, MAESTRO, MUSDB18-HQ, and an all-negative MUSAN
abstention population. Although controlled sensitivity improved, the candidate
failed the frozen external false-advisory gates and was rejected. Oracle v12
therefore remains unchanged; the complete negative result and promotion policy
are published under validation/real-world/ so a promising laboratory score
cannot silently become a user-facing claim.
- Audio analysis runs locally.
- Selected audio is not uploaded.
- There is no account, advertising, or telemetry.
- AcoustID lookup is optional and off by default.
- Direct MusicBrainz enrichment is independently optional, uses no API key, and is rate-limited to one request per second.
- Source files are never overwritten by Repair.
- Privacy-safe diagnostic exports omit filenames, paths, hashes, tags, and report evidence.
- Privacy-safe acceptance exports record workload, throughput, sampled memory, storage growth, cancellation, source class, and recovery without exposing filenames, paths, hashes, tags, or audio evidence.
- Local rotating logs may contain filenames and paths so a failed file can be identified; they stay on the user’s computer unless deliberately shared.
Read the complete privacy and security model.
| Start here when… | Document |
|---|---|
| You want normal step-by-step operation | User guide |
| You want to enable AcoustID or MusicBrainz | External identity services setup |
| You want to understand Oracle and every verdict | Oracle Engine guide |
| A technical word is unfamiliar | Plain-language glossary |
| You want formulas, standards, and limitations | Methodology |
| You want to understand validation claims | Validation corpus contract |
| You want to understand how corpus evidence tunes Oracle | Oracle calibration guide |
| You want the current GA score and remaining gates | GA readiness audit |
| You want the release-blocking defect rules | Defect severity policy |
| You need to install an unsigned build | Unsigned installation |
| You want the implemented capability contract | Product audit |
| You are preparing a release | Release workflow and RC checklist |
| You want to contribute | Contributor guide |
Audio-V requires Node.js 22 or newer.
git clone https://github.com/DRAZY/Audio-V.git
cd Audio-V
npm ci
npm run provision:engines
npm run devRun the complete source gate:
npm run verify
npm run validate:platform
npm run validate:formats
npm run verify:release-candidateCreate local development packages:
npm run dist:mac
npm run dist:winMaintainers build packages on controlled local systems and upload the verified artifacts directly to a versioned GitHub Release. Pushes and tags do not start hosted builds. The optional GitHub verification workflow is manual-only.
Audio-V uses Electron, Vue 3, TypeScript, Vite, FFmpeg/ffprobe 8.1.2 LGPL builds, SQLite-backed storage, Vitest, C2PA Tool, Chromaprint, and Electron Builder. The renderer is context-isolated and receives a narrow preload API.
Audio-V source is licensed under GNU AGPL version 3 only. Dependencies and bundled tools keep their own licenses.
Contributors accept the Audio-V Contributor License Agreement, which preserves contributor ownership while granting the project explicit relicensing rights. The Audio-V name, logo, icon, Oracle Engine identity, and official-release designation are covered by the trademark policy.
Audio submitted for validation uses the separate Validation Corpus Contribution Agreement.
Evidence before assumption. Measurement before verdict.