v1.8.0 adds offline processing and analysis APIs and changes several processing defaults, timing contracts and packaging details. Review the sections that apply to your product before upgrading, even when existing calls still compile.

Changed processing defaults

The following processors change their default design:

Processor	New default	Previous behaviour
`Oscillator`	Saw and Square edges corrected with a table minBLEP; band-limited edges ring to about 1.45x full scale	`setAntiAliasing(Oscillator<T>::AntiAliasing::PolyBLEP)` keeps every waveform inside [-1, 1]
`Equalizer`	Peak bands use the analog-matched design, so high bells keep their shape near Nyquist	`setMatchedBells(false)` restores the bilinear bell
`SpectralDenoiser`	Decision-directed Wiener gain over a mean-power noise profile	No switch: re-tune the threshold by ear if a preset relied on the hard gate

AlgorithmicReverb was rebuilt: the same parameters now drive a true-stereo feedback delay network with a binaural early field. Presets load unchanged, but the sound and the CPU cost differ, so re-audition mixes that depend on it.

Limiter now holds reduction for 10 ms before release. setHold(0) restores release-only recovery; state blobs that predate this field restore zero hold. This changes recovery after isolated peaks, so re-audition existing settings.

Oversampling and latency

Always query the prepared instance's getLatency() and propagate it to the host and any parallel dry path. Do not hard-code previous latency values. The analog effects retain getLatencySamples() as a compatibility alias.

Processor	Current contract
`TubePreamp`	Default 2x, selectable 1/2/4/8/16. Latency is 0/71/99/113/121 base-rate samples respectively. Stage-count changes preserve this latency and prime the incoming circuit for 5 ms before a 20 ms crossfade. At 1x the point-sampled circuit remains available; it does not have the alias rejection of the 2x path.
`Saturation`	Default changes from 1x to 2x, adding 64 samples. `setOversampling(1)` selects zero added latency.
`TransformerModel`	Default 2x, selectable 1/2/4/8/16. `getLatency()` is now an instance method, not a static constant; it reports 64 samples at 2x and zero at 1x.
`StereoGenerator`	New class; default 1x with 256 samples before an optional delta-only low cut. Higher factors are selectable at preparation. Width zero is delayed identity, not zero latency.

Oversampling is a setup choice; prepare again as the individual class requires. The TubePreamp circuit rendering and TapeMachine bias correction change existing renders. Re-audition analog presets and retain the explicit sample rate/factor when comparing results.

Time, pitch and sample-rate conversion

New TimeStretch and PitchShifter instances use Quality::Studio. At 48 kHz with default frames, the pitch shifter reports 5184 samples; the time stretcher's fixed-rate adaptor reports 5632. Their frame argument defaults to zero to select the engine's frame. Select Quality::Standard explicitly to retain the earlier engine (Quality::High also remains available on PitchShifter). State blobs without the quality field restore Standard. PitchCorrector still selects Standard for its existing retune dynamics.

TimeStretch::setQuality() takes effect at preparation or reset. Crossing between Studio and Standard/High in PitchShifter restarts the stream at the next block and changes latency; Standard/High changes crossfade live. Schedule engine changes where a stream restart and host latency update are acceptable.

Resampler designs filters per rate conversion. Quality tiers specify passband and rejection rather than a fixed tap count, so query getFilterLength() and getLatency(). For example, High at 44.1 to 48 kHz reports 112 output samples instead of 35. Equal-rate streaming is a one-sample delay; offline conversion is time-aligned. processRange() and getReach() expose finite-source support.

Reverb now aligns and scales resampled or stretched impulse responses over the entire conversion kernel. An IR with no initial silence can require added latency to retain its pre-ringing; query the prepared effect after loading it.

Offline processing and interval measurements

The new Offline* dynamics, clipping and stereo adapters require complete-source analysis and run on a worker, never an audio callback. They accept bounded source and transactional sink adapters; convenience run() calls also allocate an owning output. Plans belong to the exact stage input analyzed. Update source revision and analyze again after editing or processing that input. Exclusions are source-relative half-open sample intervals and retain their PCM exactly.

These APIs require exceptions and are omitted from DSPark.h when exceptions are disabled or DSPARK_NO_OFFLINE is defined. Check DSPARK_HAS_OFFLINE. The real-time StereoGenerator core remains separately usable. Offline clipping defaults to 4x local processing, with 1/2/4/8/16 selectable; the offline guide states factor-specific quality and cost.

The new AudioIntervalAnalyzer reports the requested and actually measured interval, numeric validity, and complete loudness-window support. Use IncludeContinuous for additional contextual observations; regional integrated loudness still measures the interval independently. Finalize the true-peak tail without feeding zero padding into loudness windows. See the contract and compiling example.

LoudnessMeter::getLoudnessRange() now uses the EBU percentile ranks and 10 Hz short-term sampling, so previously saved numerical expectations can change.

Parameter changes glide

Updated mix, width, gain, shape and ceiling controls use time-based smoothing instead of stepping once per block. Durations belong to each class's setter contract: for example, TubePreamp limits dry/wet slew to a full-scale change per 20 ms, while StereoGenerator width ramps over 5 ms. Code that compared the first processed block against the new target should allow for the glide.

Plugin layer

Bypass is latency-aligned. The dry path of the soft bypass is delayed by the reported latency, so a bypassed track stays where the host compensates it. Plugins with no latency are unaffected.
Host text parsing refuses garbage. Text that is neither a number, a choice label nor On/Off now makes getParamValueByString return kResultFalse and CLAP text_to_value return false, instead of mapping to the minimum. Hosts keep the current value.
Restart requests reach the host on its UI thread. A latency change detected in the audio callback is delivered by a ~30 Hz UI-thread tick (VST3 and AU) instead of from the callback itself. Hosts that polled getLatencySamples() right after a block see the new value; the notification may arrive up to one tick later.
New parameter helpers. choice(id, name, labels, default) and stepped(...) replace hand-built Param { ... , steps } aggregates. The aggregate form still compiles; a Param gained a trailing labels member that defaults to null.

CMake package and helper

dspark_add_plugin() is defined by add_subdirectory(), FetchContent and find_package(dspark). A project that defines a function of the same name must rename its own.
The installed package now contains the plugin layer and the vendored format SDK headers with their licenses.
dspark_add_plugin() accepts VERSION (bundle and AU component version) and BUNDLE_ID. Set BUNDLE_ID to your own reverse-domain identifier before shipping on macOS; the default is only a placeholder.

Conan include directory

The Conan recipe now installs the headers under include/dspark, like the CMake package, and exposes that directory as the include root. The same #include <DSPark.h> therefore works with every package manager. Code that wrote #include <DSPark/DSPark.h> against the previous recipe must drop the DSPark/ prefix. The recipe also loads dspark_add_plugin() through CMakeDeps.

The repository supplies a local Conan 2 recipe and a vcpkg overlay port. They pin an immutable source commit and authenticated archive checksums; they do not follow main. Their presence here does not mean publication in Conan Center or the central vcpkg registry.

MP3 round trips

Files written by Mp3File now carry an Info frame with a gapless tag, and decoding trims the encoder delay and padding such a tag announces when its CRC checks out. A decoded file therefore has exactly the length of the source instead of carrying the encoder delay and end padding; code that compensated for the codec delay by hand must stop doing so.