Changelog

Version 0.3.0

Added

• Clean up audio. DeepFilterNet 3 strips background noise out of speech — hiss, hum, traffic, room tone, a fan — and leaves the voice. Strengths are set from the model’s own attenuation limit, so “light” means a measured 12 dB ceiling rather than a guess. Stereo stays stereo: each channel is enhanced separately. Measured on a test mix, at full strength: 20 dB of noise removed with the speech level within 1 dB of the clean reference.
• Command line: --clean-audio light|medium|full.

Changed

• Cleanup runs before silence removal and transcription, and both now read the cleaned track. Noisy audio is precisely the hard case for a speech detector and what makes transcription guess, so doing it in this order makes the other two features better rather than merely coexisting with them.
• The AI models page and the AI options panel gain a fifth entry. The panel already scrolled, which is what that scroll box was for.

Fixed

• Subtitles vanished when audio cleanup and silence removal were both on. Remapping a caption onto the shortened timeline dropped any caption whose end had no mapped position instead of clamping it to the last kept moment — and cleaned audio makes that the common case rather than a rare one, because the transcriber stops splitting clean speech into short segments and returns one long caption spanning the file. Every subtitle was discarded and the row reported “no speech found” for a file plainly full of speech.

Notes

• The published DeepFilterNet export is the neural graph only: it takes normalised spectral features and returns a gain mask and filter coefficients. The STFT, the ERB filter bank, the feature normalisation, the five-tap complex filtering and the overlap-add synthesis are all implemented here against the project’s own reference library. Every one of them fails silently when wrong — the model still returns a well-formed mask and the audio still plays — so the implementation was checked numerically against a reference build of the same contract rather than by ear: correlation 1.000000, residual 71.6 dB below signal, identical SNR.
• Three specifics that are easy to get wrong and quiet when you do are now recorded for future work: the reference library normalises in analysis only (omitting it gives 2 dB of denoising instead of 20); the filter’s taps run t-2 .. t+2, and the other sign also “works” while being 5 dB worse; and 960 is not a power of two, so the transform is Bluestein rather than radix-2.
• This is the first model taken from a third-party re-upload rather than from the project that trained it, because upstream ships weights for its own runtime and not a portable graph. Both licences are recorded in the notices, and the risk that trades away — a repository that could vanish in a way the original would not — is written down rather than left implicit.
• A review of further suggested models is recorded: SCRFD, YOLO-based face detectors and Zero-DCE cannot be used — non-commercial weights, copyleft licensing and a non-commercial Creative Commons licence respectively — all verified at the source rather than assumed.

Version 0.2.0

Added — remove silence

• Finds where people are actually speaking, using Silero VAD, and cuts the gaps between. Three strengths, from trimming only long dead air to removing everything that is not speech, with padding left around each phrase so the cut does not clip the first and last word. The row reports what it removed when it finishes.
• Cutting happens inside the single conversion pass, with the timestamps rewritten — so there is one re-encode, no temporary fragments, and nothing that can drift out of sync at a join.
• Subtitles are moved onto the shortened timeline. Transcription runs against the original audio, because that is where the speech is; every caption is then remapped. Without that step the captions slide further out of sync with every gap removed.

Added — blur faces

• Detects faces on every frame and blurs them. Aimed at footage filmed in public that is going to be published.
• Detection runs on every frame rather than on samples with interpolation between them, and a face lost for a moment stays blurred for a short while afterwards. A one-frame dropout is invisible while scrubbing and perfectly legible when paused, which is the whole problem.
• The app says out loud that no detector catches everything and that the output should be checked. A privacy feature that is 95% right is not 95% useful.
• Face blur runs before upscaling when both are on, so the upscaler is never handed a face to reconstruct in higher fidelity than the original had.

Added — AI upscale

• Real-ESRGAN 4x, applied per frame with overlapping tiles so no seams appear.
• Limited to sources up to about 480p, because 4× of that is 1080p and 4× of anything larger is past 4K.

Added — the frame pipeline and the command line

• A shared frame pipeline behind the per-frame features: the source is decoded to raw frames, the models run, and the frames are piped straight into the conversion that was going to happen anyway. One encode, no intermediate file, and the existing options — resolution, frame rate, subtitles, silence removal — all still apply on top.
• New command-line options: --remove-silence light|medium|aggressive, --blur-faces and --upscale. The headless runner also now prints what the AI steps did, not just where the file went.

Changed

• The AI models page gains the Silero VAD, face-detection and Real-ESRGAN entries. Silero was already in the catalogue and deliberately withheld; it is now switched on.
• Third-party notices rewritten. They named a sibling application throughout — this app’s notices had been copied from it — and covered only FFmpeg, with nothing for the speech runtime or the model weights already shipping. They are now correct, and cover every AI component including the two new licences.
• Every model in the catalogue had been MIT-licensed; Real-ESRGAN is BSD-3-Clause, so the documentation no longer claims otherwise.

Fixed

• The Start button could end up off the bottom of the window. Four AI bands added roughly 320px to the output-options card, which is an auto-sized row — so it grew and pushed the action bar off screen. The AI options are now one bounded, scrollable group, which caps that growth for good, and the window opens maximized.
• The Queue page said “Skipped” and nothing else. It never bound the job’s message, so the reason a file was rejected — the only thing you need at that moment — was visible on the Convert page and nowhere else.

Notes

• Per-frame models use your graphics card through DirectML when a DirectX 12 GPU is available, and fall back to the processor when it is not. The difference for upscaling is minutes per minute of video versus hours, so the job row says which one it got rather than leaving you to guess why it is slow.

Version 0.1.0

Added — the converter

• Conversion engine with 20 output formats — MP4 (H.264 / HEVC / AV1), WebM (VP9 / AV1), MKV, MOV (ProRes / H.264), AVI, animated GIF, seven audio formats, and container remuxing. Every generated command was run against the bundled binary before shipping, not assumed to work.
• Remux — “change container, no re-encode”. Copying the streams into MP4, MKV or MOV: seconds instead of minutes, with no quality loss at all. Offered whenever the source codec is legal in the target container.
• Batch queue. Drop a file, a selection, or a whole folder. One set of options applies to the whole list — that is the entire promise of the app, and there is deliberately no per-file override. A single file is just a list of one.
• An encoder ladder that is tested rather than trusted. H.264 and HEVC output comes from a hardware encoder — NVENC, Quick Sync or AMF — or from the encoder built into Windows. Each candidate is test-encoded before it is chosen, because the encoder list advertises NVENC even on a machine with no NVIDIA card, so the list alone cannot be trusted.
• Live progress and real ETAs, per file and overall, mirrored onto the taskbar button. The ETA is withheld for the first two seconds because encoders spend that moment initialising, and an estimate that swings wildly reads as a bug.
• History, kept across restarts, showing what each conversion did to the file size.
• Settings — encoder override, how many files to convert at once, what to do when a filename is taken, thumbnails on or off.
• “When done” — do nothing, open the folder, close the app, sleep, or shut down. Sleep and shutdown go through a 60-second delay so an unattended batch that finishes while you are back at the keyboard can still be cancelled.
• Headless mode, so everything the window does is scriptable.

Added — subtitles and the AI model system

• Automatic subtitles. Speech is transcribed on your own machine and written as a separate .srt, muxed as a real subtitle track, or burned into the picture. Can translate to English while transcribing. Nothing is uploaded anywhere.
• Optional AI models, downloaded on demand, verified by exact size and SHA-256, and written through a temporary file so a failed download can never be mistaken for a working model. Model files you already have can be imported from disk, for machines with no internet.
• Model download URLs are pinned to immutable revisions. A moving URL would fetch different bytes if a model were ever re-uploaded, fail the fingerprint check, and break downloads for everybody at once with nothing in the app to explain why.

Fixed before release

• Subtitles were unfindable. The control was correct and enabled, but it sat as the sixth dropdown in a wrapping row of six, and the AI models page named what a download added only in small grey text with no route to it. Subtitles now have their own band in the options card, and an installed model gets a “Use it” button that goes straight there.
• “Remove silence” was advertised without existing. Its model was downloadable and the page reported “Ready”, then pointed at a control that had never been built. Capabilities are now hidden, along with their models, until the feature actually ships.
• The best installed speech model is now the one used. Selection took the first model in list order, which is sorted by download size — so downloading a better model while keeping the old one silently carried on using the old one.
• A sidecar .srt alongside an audio-only output was refused, with a message describing the very mode being asked for. Transcribing a podcast to MP3 plus subtitles now works.

Decisions worth recording

• Converting a file into its own folder in the same format always gets a suffix, whatever the overwrite setting says. Reading and writing one file at once would corrupt it.
• Impossible combinations are rejected before the batch starts — extracting audio from a silent file, remuxing HEVC into AVI, burning subtitles into a stream being copied untouched. The row explains why in a sentence instead of showing raw encoder output.
• Conversions run one at a time by default. Parallel encodes compete for the same GPU encoder session and usually finish slower; 2–4 at once is available in Settings for the formats that encode on the processor.
• GIF gets a two-pass palette. A naive one-pass GIF is 8-bit mud.
• Thumbnails are taken 10% into each file, not from the first frame, which is usually black or a title card and made every row look identical.