Sound

Release notes

What we're building, unfiltered.

Neither instrument is for sale yet. This is not the pitch. It is the workbench. Every issue from both, newest first: what we built, what it measures, and what it still cannot do. The numbers stay honest, which is the whole point.

All Sibilance notes →All Thunder Clarity notes →
Thunder Clarity
Issue 03 July 18, 2026 Sydney, AUS v0.4In the build

Studio Restore learns to rebuild a voice

The big one. Cleanup can only remove what's wrong with a take; it can't put back what a phone, a codec, or a cheap mic never captured. So we taught Thunder Clarity to re-synthesise the voice itself.

New
  • Studio Restore, in the plugin and standalone: capture a take and a generative model rebuilds it at full studio bandwidth. The top octaves a phone call or a Zoom recording threw away come back. It runs offline, faster than real time.
  • Restore a file, not just a live take: open any WAV and it loads as the take; Save writes the result. The same path works at your desk in the standalone or on a track in your DAW.
  • A dial between natural and fully restored: generative models can sound… processed. So restoration is a blend, not a switch: 0% is your untouched take, 100% is fully rebuilt, and everything between is yours.
Refined
  • It changes tone, not volume: the model used to hand back a voice about four times quieter than it went in. Now the restored take is matched to your original's loudness, so the blend is a clean morph, not a level dip.
  • 24-bit masters: restored files export at 24-bit, the depth a studio actually expects.
  • Long takes stay whole: a full-song vocal or a podcast segment no longer gets quietly clipped at five minutes; and if a take ever does hit the ceiling, we say so instead of trimming in silence.
Little fixes
  • 48 kHz sessions used to restore fine, then come back as gibberish. They don't anymore.
  • Hit Restore with nothing captured and it used to think about it. Now it just waits for you.
  • If the model file is missing, Restore steps aside instead of taking the audio with it.
Notes from the roadmap
  • Making it faster still (quantising the model), and a cleaner resampler for 48 kHz sessions.
  • And the part no test can sign off: hours of listening, on real spoken and sung takes.
Thunder Clarity
Issue 02 July 18, 2026 Sydney, AUS v0.3In the build

Meters that don't flatter you

A clarity plugin is only as trustworthy as its numbers. This pass was all about honest measurement, and making the cleaned voice sit at the level it came in at.

New
  • Auto-balance, loudness-preserving: de-noising and de-reverb remove energy, which used to leave the output quieter. Now it's ramped back to the input's loudness automatically (cleanup changes tone, never level), and it's peak-safe, so it can't push you into clipping.
  • The delivery trio: LUFS, true-peak, integrated: ITU-R BS.1770 short-term loudness, 4×-oversampled true-peak (dBTP), and gated integrated LUFS with a reset: the three numbers streaming and broadcast actually check you against.
  • Studio polish, on a dial: a gentle, self-tuning tonal move toward a studio-balanced voice: it brightens a boomy source and warms a thin one, whatever it was recorded on.
  • A/B, loudness-matched: compare processed against dry at matched loudness, so you judge the tone and not just 'louder sounds better'.
Refined
  • Music is the default: a fresh insert opens ready for a sung or spoken take: gentle room reduction, natural reverb kept.
Little fixes
  • The studio polish could place an EQ band above Nyquist at 8 kHz and take the whole channel down. It can't reach that high anymore.
Thunder Clarity
Issue 01 July 17, 2026 Sydney, AUS v0.2In the build

A neural core that tells the truth

Under the three faders is a real neural separator (DeepFilterNet). This week was wiring it in honestly, and refusing to ship anything that only looks like it's working.

New
  • The honest SNR readout: the model reports its own estimate of the input's speech-to-noise ratio, and we surface it as-is: no invented 'clarity score'.
  • A 48 kHz notice: the neural core is happiest at 48 kHz; if your session isn't, the plugin says so plainly instead of pretending.
  • User presets: save, recall and delete your own fader states, across projects.
  • Zipper-free faders: per-sample gain smoothing, so moving a fader (or toggling bypass) never clicks.
Little fixes
  • Removed a 'Low-latency' switch that did precisely nothing. A control that lies is worse than no control.
Sibilance
Issue 03 July 16, 2026 Sydney, AUS v2.1Shipping

It knows a hiss from a cymbal

Most de-essers hear energy in the 5–8 kHz band and duck it, which is why they also duck cymbals, breath, and the top of a bright vowel. Sibilance now identifies the actual phoneme first, and only acts when it's really an ess.

New
  • Phoneme-aware detection, real-validated: a small classifier tells s from sh from f/th from t before anything moves. Measured on real speech (s 97%, sh 93%, f/th 76%, t 57%), and we publish the honest numbers, not the synthetic-only ones.
  • A temporal model, not a single frame: it reads a short sequence rather than one instant, so a burst like a 't' is caught by its shape in time. Real 't' detection climbed from 17% to 57% over the work.
Refined
  • Surgical resonance bells: the Smooth engine's cuts got narrower (Q 6.5 to 8.0): more precise, less collateral on the surrounding tone.
  • De-Pop and Breath, on by default: the two things every close-mic'd vocal needs, working out of the box.
  • A preset browser that's actually organised: categorised, source-specific presets, now 21, so you start from something near your voice.
Notes from the roadmap
  • Lifting plosive and fricative detection further. t and f/th have the most headroom left.
  • More source-specific starting points. As always, validated on real voices before it ships.
Sibilance
Issue 02 July 16, 2026 Sydney, AUS v2.1Shipping

Aerospace-inspired safety

The detection path is now aerospace-inspired: it fails safe, and tells you when it has. Before it acts, two independent paths must agree that a sound is sibilant, the command-and-monitor redundancy that flight-control computers rely on.

New
  • Command/monitor redundancy: two independent paths agree on whether a sound is sibilant before the plugin acts, the same command/monitor pattern flight computers use. If they disagree, it defaults to caution.
  • A fail-safe input guard: pathological input (NaNs, denormals, dead silence, clipping) can't push the engine into a bad state; it detects and rides through.
  • A signal-integrity annunciator: a little cockpit-style lamp on the panel: a lamp-test on load, and a caution light if a lane ever goes unhealthy. You can see the plugin watching itself.
Refined
  • New character options: a Warmth flavour, plus a Tracking Assistant for setting up while you record.
Sibilance
Issue 01 June 18, 2026 Sydney, AUS v2.0Shipping

A new face, machined

A full visual rebuild, and the installer to match.

New
  • A physical, machined interface: canvas-drawn knobs with real material, a brushed-metal surface, and Human Engine's own typeface. The plugin looks like one object, milled rather than drawn.
  • A spectral painting field: the old spectrum readout became something you can actually read the voice in.
  • One installer, every format: v2.0.0 ships VST3, AU and CLAP in a single macOS package.

The list

Be there for the first release.

When the installers are signed and the first download goes live, the list hears first and gets the intro price. That is the only thing this form does.

Release news only, unsubscribe any time. Privacy.

FormatPlatform
VST3macOS · Windows
Audio Unit (AU)macOS
CLAPmacOS · Windows
StandalonemacOS · Windows

Apple Silicon & Intel · JUCE 8 · C++20 · No AAX · See what ships →