Reviewed 2026-08-20
These exact filenames are attached to the tagged release. Checksums let you verify a completed download independently.
Scroll horizontally to review every column.
| Asset | Platform | Size | SHA-256 |
|---|---|---|---|
| DepoAudio_1.0.2_universal.dmg | macOS universal | 137.0 MB | 9ab0e1f348816b850a19d15226cda8140e5ffd9be4e5e3ca158a9ac3e2f788a2 |
| DepoAudio_1.0.2_universal.app.tar.gz | macOS universal | 137.1 MB | 3c754dc69ba43f68f5c82e5fb671c85e8dce2e7c15e8907cb783f7f46869347f |
| DepoAudio_1.0.2_x64-setup.exe | Windows x64 | 308.0 MB | 08b31d66870a4dcb40529b35fd0e5a74db6746f435567483f391750361638856 |
| DepoAudio_1.0.2_x64_en-US.msi | Windows x64 | 338.2 MB | 14cf21148c3e880d0e4256de41e64c3a3b88f671a0a943970eee56b2b84dcabe |
macOS app bundle
The exact v1.0.2 release job used the ad-hoc identity "-" because Developer ID credentials were absent, and it skipped notarization. Deep codesign verification of the published bundle was unavailable on the Windows review host. Gatekeeper behavior was not clean-tested for the published v1.0.2 app.
Read macOS install steps →Windows installers
Direct byte- and SHA-256-matched inspection found no Authenticode signature, signer, or timestamp on either the published EXE or MSI. SmartScreen and organization-policy warning behavior was not clean-tested for the published v1.0.2 installers. Verify the filename and checksum before continuing.
Read Windows install steps →Application updater
The published release was built without an updater endpoint or signing key and has no updater manifest or updater signature assets. Use this page or the tagged GitHub release to check for a newer build when automatic updates are unavailable.
Local processing
Core conversion does not upload recordings. Selecting Install for an optional model contacts GitHub without sending the recording. The published release was built without an updater endpoint or signing key and has no updater manifest or updater signature assets.
Read the privacy details →“Released” means supported by the tagged v1.0.2 source and published installer lineage. “Experimental” requires output verification. Hidden, unsupported, or planned work is not marketed as available.
Open supported court recordings and export WAV, MP3, FLAC, Opus, or M4A.
Mix to stereo, keep the source channel layout, or export one file per channel. Channel labels are user-editable roles; they do not identify a person from their voice.
Reduce steady background noise locally. Noise reduction cannot restore speech that is missing from the recording.
Balance quiet and loud channels using bounded per-channel loudness analysis. Auto-level is unavailable in Keep Original mode. This is level analysis, not speaker identity or a quality score.
Apply the released clipping-repair processing option.
Review source and converted audio with speed, loop, track-label, and bookmark controls.
Organize converted audio by case and session, then search, play, archive, or re-export it.
Find recordings from recognized court-reporting software locations for import.
Attempt conversion of FTR/TRM recording segments. Important decoder and segment-ordering fixes are present only in inspected source. Preserve originals and verify the complete output.
Estimate speech and silence from a bounded local excerpt. This detects voice activity, not who is speaking.
Estimate likely turn boundaries during local analysis. This estimates turn completion; it does not identify a speaker. No model-specific packaged inference smoke was recorded.
Estimate active segmentation-model slots from the first 60 seconds. This is not a distinct-person count or voice identification; embedding and clustering do not run. No model-specific packaged inference smoke was recorded.
Apply per-channel bandwidth extension to narrowband audio. It cannot restore missing words or guarantee clearer speech. No model-specific packaged inference smoke was recorded.
Estimate speech, background, and overall quality from a short excerpt after an optional model install. The result is an estimate, not a measured improvement or intelligibility guarantee. No model-specific packaged inference smoke was recorded.
Import SRT, VTT, or TXT, proof or build text against playback, stamp lines to the playhead, and export SRT or TXT. VTT is import-only. This is a local proofing editor, not automatic transcription or named-speaker identification. Text autosaves per track path in local app storage; export a copy before moving audio or clearing app data.
Use Fast (RNNoise) for released noise reduction. DeepFilterNet model files are reserved and unused because the production spectral pipeline is not implemented. The backend rejects Best-quality denoise.
Not part of the supported current-release workflow. A downloadable asset exists, but no released inference, clustering, or identification path uses it.
Not available in the current release. A source path exists, but its DCCRN+ model is absent from the installers, app catalog, and models-v1 release.
Not part of the supported current-release workflow. Release readiness remains tracked in app issue #90.
Release methodology
Public claims are checked against the published tag, exact release commit, asset list, build provenance, and installed behavior where available. Source-only fixes stay labeled unreleased. Merge remains independently gated under app issue #90. The synced transcript editor is in the exact published source and asset lineage, with import, export, persistence, error, and accessibility paths covered by tagged tests. A final manual packaged-UI smoke remains documented separately.
For v1.0.2, the release workflow started the exact macOS bundle executable and the application extracted from the exact Windows MSI; both remained alive through a 12-second startup window. The NSIS setup wrapper was not independently installed by that gate, and this evidence does not replace a clean-install or complete workflow test.