- Generated by
1.9.8
|
DSPark 1.8.0
Header-only C++20 DSP for real-time and offline audio
|
All notable user-facing changes to DSPark are documented here.
wdf::ToneStackFMV::analogStateSpace(): the continuous-time state space of the Bassman tone network (capacitor voltages as states), from nodal analysis of the same element values and topology as the R-type adaptor. With an ideal source and load it matches the published analog transfer function to 1e-8 relative. wdf::Resistor::getResistance() and wdf::Capacitor::getCapacitance() report element values.AudioIntervalAnalyzer: bounded block-stream analysis of exact sample intervals, with RMS/peak/finite true peak, scoped loudness, explicit validity, optional continuous observations and no retained PCM. Reuses existing meters and the scaled-energy kernel shared with OfflineEnergyAnalyzer.LoudnessMeter::finalizeTruePeak() completes a finite interpolation tail without advancing loudness windows or gates; getMeasurementInfo() exposes complete-hop coverage and absolute-gate counts to the stream owner. TruePeakDetector::getTailPeak() preserves live channel history.OfflineStereoGenerator: complete-source rendering through StereoGenerator, with explicit mono duplication, exact exclusions, compensated latency and measured output headroom. A host-owned canonical delta cache supports new widths/exclusions while validating source, clock, settings and cached PCM. Worker scratch is bounded independently of duration; output is transactional.OfflineStereoBalance: source-bound side-only leveling from shared mid/side analysis, with an actual-render guard, exact exclusions and transactional output. It preserves the original mid without a master gain trim or limiter.StereoGenerator: twenty moving bands and rational color on a parallel copy, with delta add-back preserving the delayed original mid. Includes width smoothing, optional delta-only low cut, explicit local oversampling, source clock/reset contracts, presets, latency reporting and numerical-error status. Shares Core conversion and clipping kernels; processing allocates nothing.OfflineSoftClipper and OfflineHardClipper: full-source automatic sample-peak reduction using shared Clipper curves, continuous-interval antialiasing and finite-source sinc reconstruction through Core convolution and Cauchy maps. Actual rounded PCM calibrates the ceiling, with explicit local oversampling, original sample alignment, exact exclusions, bounded PCM caches and transactional output. Coarse maps scale with duration and reuse storage across calibration passes. No master gain trim or extra limiter.OfflineBeatCompressor: complete-source pulse peak leveling, upper-median reference, capped reduction, local tempo release and explicit unknown-tempo fallback, reusing shared analysis and the bandlimited gain renderer.OfflineTempoAnalyzer: source-bound beat intervals and ambiguity diagnostics using retained pooled-channel features and the existing BeatTracker engine. Adds per-job allocation accounting and cancellation without another FFT; unsupported pulse intervals return no reliable local tempo.OfflinePunch: source-aligned transient boost with a bounded, stereo-linked gain envelope; shared smoothing, exclusions, verified rendering and source analysis. The attack map also recovers quiet restarts masked by global novelty and rejects end-padding attacks, with sample-clock minimum spacing.OfflinePeakCompressor: automatic complete-source peak reduction, event-local holds, anticipatory smoothing and stereo-linked amplitude recovery. Reuses the transient analyzer, envelope follower and gain transport shared with Leveler. Requested reduction is capped by measured sustained headroom; exclusions and achieved sample-peak reduction are explicit, with no hidden audio clipping.OfflineTransientAnalyzer: separate attack and pulse maps from a shared pooled stereo spectrum, complete-source normalization, refined source-frame intervals, optional reusable novelty features and bounded worker memory. Reuses the same FFT, filterbank, novelty and peak-picking kernels as OnsetDetector. The attack map also recovers upper-register strikes masked by falling low-frequency energy without extending bass holds. Corpus results include the measured precision/recall tradeoff in docs/offline-processing.md.OfflineLeveler: automatic, stereo-linked upward macro RMS leveling with exact source-frame exclusions, reusable immutable plans, cancellation and transactional offline rendering. Reports output sample/true peak without imposing hidden limiting or attenuation. Shared exclusion feathers have three zero edge derivatives and smooth overlapping masks, avoiding gain cusps between neighboring protected regions.OfflineEnergyAnalyzer and shared offline source/sink/job contracts: complete mono/stereo analysis in bounded blocks, 64-bit source positions, actual final-bin duration, content/revision checks and payload-memory budgets. Maps retain energy features rather than PCM. See docs/offline-processing.md.PitchCorrector::getState() / setState() (scale, root, retune speed, formant preservation) and SpectralFreeze::getState() / setState() (freeze request, phase mode): they were the two effects with settings that a host could not save with a session.DelayEstimator: the delay between two recordings of the same material (a delivery against its reference), found by GCC-PHAT on a cross-spectrum averaged over the whole program and refined on its phase slope to a fraction of a sample, with a confidence and polarity. Memory is a few frames whatever the program's length; blocks of any size give the whole-signal answer. Measured over 11 s at 48 kHz: within 0.0005 samples on white noise down to 0 dB SNR and 0.003 on coloured noise at 20 dB, for delays from -250.4 to 4000.25 samples.Delay::setPingPong(): the insert-style processBlock(buffer) runs as a ping-pong delay (the first two channels cross their echoes) with the insert's blend, so one call serves both modes. Fully wet it matches the wet-buffer sequence sample for sample; the setting is saved in the state (older blobs restore it off).Resampler::processRange() and getReach(): offline, time-aligned conversion of any span of output samples, including the kernel's ringing before the first input sample and after the last.Sampler: a polyphonic multi-zone sampler. Zones map a recording over key and velocity ranges with root key, tuning, gain, pan and loop (continuous or sustain, with a crossfade baked at load); overlapping zones layer. The interpolator is a 64-tap Kaiser sinc (100 dB) whose kernel follows the playback rate - gathered at or below the recording's rate, scattered in output time above it - so transposition is clean both ways and nothing above the output Nyquist frequency folds: -118 to -130 dB THD+N from -12 to +19 semitones, -110 dB of alias for an 18 kHz tone an octave up. Stolen voices fade over 3 ms, events are sample-accurate at any block size, and zones are replaced while playing through a lock-free handover.TimeStretch::Quality::Studio and PitchShifter::Quality::Studio, a new engine and the default for new instances: phase-gradient heap integration on a reference summed over every channel, an unambiguous instantaneous-frequency estimate, and a time map anchored on strikes found by a look-ahead spectral-flux detector; inside a strike's lock the bins the strike rises into are copied unrotated, so it carries no pre-echo. Scored against an ideal rendering of the same scenes: time stretch spectral distance 4.12 dB (Standard engine 5.01), strike timing 0.59 ms median (4.16); pitch shift spectral distance 1.96 dB at the default frame and 1.47 at 4096 (Standard engine 2.50), strike timing 0.68 ms (13.18).Resampler kernels are designed per conversion from a specification - a passband edge and a stopband attenuation per quality tier, with the stopband starting at the lower Nyquist frequency - and every rational ratio (all common audio rates) uses exact polyphase phases with integer position arithmetic; other ratios read a 512-phase table with cubic interpolation across phases. In double precision, worst over six common conversions, Ultra leaves a -220 dB residual and a -214 dB stopband (before: -105 dB and -53 dB) with a flat passband to 0.915 of Nyquist. getFilterLength() reports the kernel length.Limiter::setHold(): the reduction is held after the last peak that asked for it before the release starts (default 10 ms, 0 to 50 ms). A 50 Hz sine driven 6 dB over the ceiling measures -149.6 dB THD+N instead of -56.8 dB, 40 Hz -149.7 instead of -47.4, and 60 Hz + 7 kHz intermodulation -74.2 instead of -62.4 dB, for 0.21 dB of integrated loudness on a dense mix driven 12 dB over the ceiling. State blobs saved before the field existed restore a hold of 0.TimeStretch::beginOffline() / pushOffline() / finishOffline() with pullOffline(): the offline stretch fed and drained in blocks of any size, bit-identical to process() over the whole signal, holding only unfed input and unpulled output. process() is now that session run over one block.BeatTracker::beginOffline() / pushOffline() / finishOffline() and the same trio on OnsetDetector. Blocks may be any size; the result is bit-identical to analyze() / detectOffline() over the concatenation, and only the onset envelope is kept between calls, not the audio.OnsetDetector::detectOfflineOnsets() and finishOfflineOnsets() return each onset's strength with its position, on the scale getOnsetStrength() reports; getMethod(), getThreshold() and getAdaptiveWhitening() echo the settings in force.FilterEngine::setDriftSeed(): a non-zero seed makes the analog drift reproducible - the same seed renders the same output, and reset() or prepare() restarts the drift from it. Zero keeps the per-instance unique seed.Reverb::loadIR(AudioBufferView<const T>, double): a multi-channel impulse response from memory, for hosts that keep IRs in their own asset store - the same resampling, per-channel convolvers and atomic publication as loading from a file.choice() (named positions whose labels hosts list, display and parse, including AU value strings and VST3 list / CLAP enum flags) and stepped() (evenly spaced discrete positions shown as whole numbers). Unparsable host text is refused instead of guessed.find_package(dspark)) now ships the plugin layer and defines dspark_add_plugin(), as add_subdirectory() and FetchContent do; the helper gains VERSION and BUNDLE_ID.HilbertIIR, a zero-latency analytic pair in quadrature from 20 Hz; SincInterpolator (32-tap windowed sinc) and StretchedSincReader; PitchShifter Quality::High; NoiseGate and Expander lookahead with reported latency; first-order ADAA in WaveshapeTable; fractional SampleAndHold periods; FilterEngine::setShelfSlope().Delay insert API (prepare(spec), setMix(), in-place dry/wet processBlock), Crossfade equal-power sine law with curve glides and gainsFor().TubePreamp tabulates the implicit Koren load line instead of iterating it per sample (the independent current reference stays within 5e-10 A over five internal rates), and at factors 2/4/8/16 solves the circuit in continuous time inside each internal sample interval: a Farrow polynomial reconstructs the grid input, intervals are split at the triode knees (closed form where a stage is saturated, Gauss-Legendre nodes elsewhere), the tone circuit is propagated exactly in its analog modal form and the plate voltage is band-limited once, at the output, with an exact compensation of the projection droop. At the 2x default the worst alias component below 20 kHz of a 23-tone sweep up to +36 dB drive is -80.6 dBc; the September 26 source (d8a98a6) measured -9.8 dBc for two stages at +36 dB on the same measurement. Against independent dense solutions of the same circuit equations the 2x waveform error is at most -73.5 dB. Reported latency is 71/99/113/121 samples at 2/4/8/16x; 1x keeps the point circuit and zero latency. The shared per-instance table occupies 156672 bytes. Stereo processing at 2x costs 0.5-1.7 times that September 26 source at the same factor (less for clean and moderate settings, more for top-octave tones at +36 dB), with the measurement conditions in the header. The default remains 2x. Stage changes prime a prepared second circuit for 5 ms and then crossfade for 20 ms, preserving latency and avoiding callback allocations.processBlock() and getLatency() are the canonical in-place and audio delay names. Saturation's process() and the analog effects' getLatencySamples() remain compatible aliases. The cookbook documents separate-input utilities, analysis delay, branch alignment and setup-only configuration. The channel-strip example uses ProcessorChain to report the complete chain latency.BeatTracker::analyze() settles the metrical level with a model fitted to three public tempo-annotated collections: the proposed reading, its half and its double are scored on how each one's pulse sits in the full, the register-balanced and the four register envelopes, and at what rate it would be tapped. The header records the measured accuracy and distinguishes reuse of fitting material from the additional evaluation excerpts. The model is not consulted where the proposal's alternate beats cannot be told apart, so clicks, swing and quiet subdivisions keep their level; when it moves the level, the reading it moved from is secondaryTempoBpm. Rational period alternatives at 1/3, 2/3, 3/2 and 3 of the proposal now participate when range and pulse support allow, with phase-aware scoring and the original binary coefficients retained. These alternatives do not identify a time signature; the header documents grouped validation and training overlap.TransformerModel oversamples its core 2x by default, with setOversampling() (1, 2, 4, 8, 16), getOversamplingFactor() and the factor saved in the state (older blobs restore 2x). At 1x the loop's harmonics of high tones folded back into the band: a 10.1 kHz tone at -6 dBFS put its third at 17.7 kHz 61 dB down at the default drive and 37.5 dB down at +24 dB; at 2x the worst fold below 20 kHz over 1-15 kHz tones is 83.4 and 59.4 dB down. getLatency() is no longer a static constant: it reports the oversampler's delay (64 samples at 2x, 0 at 1x), getLatencySamples() returns the same value as on TapeMachine and TubePreamp, the dry path of the mix is delayed to match, and blocks longer than the prepared maximum are processed in pieces. DSParkLab's Saturation and Transformer slots start at the library's 2x.Saturation oversamples 2x by default (it was 1x, the one saturator in the framework that did not): at 1x a 10.1 kHz tone at -6 dBFS through the default SoftClip left an alias at 17.7 kHz only 34 dB down, at 2x it measures 124 dB down. The oversampler adds 64 samples of latency at 48 kHz, reported by getLatency(); setOversampling(1) restores zero latency.Compressor::DetectorType::Hilbert detects on the zero-latency allpass pair (HilbertIIR), backed by the rectified input at onsets, instead of the 191-tap FIR. The FIR was blind below about 200 Hz, so low tones were modulated at their own frequency exactly as with the peak detector; now a 30 Hz tone 14 dB into 4:1 at 5/100 ms measures -96 dB THD+N (before -35 dB) and 20 Hz -107 dB, the static curve holds within 0.02 dB, and the detector adds no latency (it reported 95 samples). A step escapes as it does with the peak detector.Resampler quality tiers are now specifications (Draft 0.80 of Nyquist and 60 dB, Normal 0.90 and 100 dB, High 0.91 and 140 dB, Ultra 0.915 and 210 dB), so kernels are longer than the former fixed 8/32/64/128 taps and latencies grow accordingly (High 44.1 -> 48 kHz: 112 output samples, was 35). A ratio of exactly 1 is a one-sample delay. Draft's passband now ends at 0.80 of Nyquist.TimeStretch and PitchShifter instances use Studio. Latencies at the default frames and 48 kHz: PitchShifter 5184 samples (was 4096), TimeStretch's fixed-rate adaptor 5632 (was 2048). prepare()'s frame argument now defaults to 0, each engine's own frame. setQuality() selects the earlier renderings (Standard, and High for the pitch shifter), bit-exact as before; state blobs without the quality field restore Standard. PitchCorrector keeps the Standard engine, whose retune dynamics it is tuned to. Crossing between Studio and Standard/High in PitchShifter restarts the stream at the next block.AlgorithmicReverb mixes its tail through a time-varying feedback matrix: after the Hadamard mix, line pairs turn through slow Givens rotations (lossless, so every T60 stays exact), which spread modes alike at every frequency and let the delay lines wander less. A steady 3 kHz tone keeps 11-14 dB more of its energy within +/-3 Hz in Hall and Cathedral (Hall -44 dB, Cathedral -45 dB outside), the modal ringing of Room and Chamber drops by about half, the late field of a lateral source decorrelates better (500 Hz coherence 0.17 instead of 0.20) and the CPU cost is unchanged. The Hall preset's modulation depth is 0.10 (was 0.13) and the Cathedral's 0.12 (was 0.16).AlgorithmicReverb::setDecay() is now the ISO 3382-1 mid-frequency reverberation time T_mid (mean of the 500 Hz and 1 kHz octave T60s). The loop's DC anchor is solved so that T_mid lands on the setting; before, the first-order absorption shelves left it 3-12% short (Cathedral at 12 s measured 10.6 s, now 11.8 s; Spring at 1.5 s measured 1.37 s, now 1.46 s). Tails are correspondingly longer at the same setting.BeatTracker::analyze() on dense mixes: the metrical level is decided on a register-balanced onset envelope, correlation noise bumps are no longer tempo candidates, and the grid is chosen among a moving-period, a steady and a tighter steady reading by how much of that envelope each explains. tempoBpm is fitted to the grid with skipped or extra beats counted as such. On 46 synthetic pop masters (92-150 BPM) the level is right on all 46 (35 before) and tempoBpm agrees with the grid's median interval within 0.55% (up to 17.7% before). OnsetDetector::OdfFrame gains registers, the flux per register group.Compressor AutoMakeupMode::Static keeps its textbook, program- independent offset, which also lifts passages that never reach the threshold; the behaviour is now pinned by a test and the documentation points to Adaptive for makeup only where gain is reduced.AlgorithmicReverb rebuilt as a true-stereo 32-line FDN (16 in Eco) with exact per-band decay, a velvet-noise early field joined to the late field on one physical decay, a binaural stereo image with directional early reflections, and a dispersive spring model - at about half the CPU.Oversampling runs a true polyphase decimator on the shared SimdOps layer (about 2x faster).Oscillator waveforms are minBLEP band-limited by default; Equalizer bells default to the analog-matched design; SpectralDenoiser uses a decision-directed Wiener gain instead of a hard gate.Limiter gain computer turns peaks down before the hard-clip backstop; AutoGain matches integrated K-weighted loudness; decibel conversions run on exp/log at half the cost.restartComponent and AU Latency listeners are called on the host's UI/main thread only: a latency change detected in the audio callback raises an atomic flag that a ~30 Hz UI-thread tick hands to the host, with no host call or allocation on the audio thread.include/dspark like the CMake package, so #include <DSPark.h> works with every package manager, and ships the plugin layer with dspark_add_plugin(). Both recipes pin the 1.8.0 source.OfflineBeatCompressor compiles after the Windows SDK, whose near macro previously erased a local helper name. Include-order coverage uses the real SDK header; numerical processing is unchanged.LoudnessMeter: the -100 LUFS floor is monotone across its former power-domain discontinuity, including positive sub-floor powers in float and double.AlgorithmicReverb: compute feedback rotation sine/cosine in double before rounding to the sample type, avoiding amplified float-libm discrepancies. Full-signal x64 MSVC/GCC/Clang fixtures now gate numerical agreement across three sample rates and static, preset and maximum modulation.AudioBufferView conversions now retain small-capacity channel views, allowing Oversampling<T, 2> and other compact Core buffers to use the common processing API. Const sample access remains const; implicit capacity narrowing stays prohibited.FIRDesign propagates allocation failures from its coefficient factories, allowing callers to recover instead of terminating the process. Successful designs produce the same coefficients.BeatTracker no longer counts an isolated half-beat phase transition as a full beat when stable intervals surround it. The worst tempo error on six 85/100/140 BPM phase-change trains falls from 4.01% to below 0.003%. With the level model and beat positions unchanged, tempo accuracy on the same 454 ballroom excerpts rises from 75.8% to 76.0%, Salsa stays at 68.9%, and the expanded 3223-loop evaluation rises from 49.7% to 50.5%. On 242 additional ballroom excerpts kept apart from the original fit and the slope development, accuracy rises from 73.6% to 74.8%.ADSREnvelope, the declared type. Windows comment-path checks normalize Git paths with forward slashes; the VST3 smoke host avoids a shadowed local rejected by strict compiler warnings.TapeMachine no longer mirrors the programme about a quarter of the sample rate. Its AC-bias carrier sat at 0.375 of the internal rate, where the carrier's third harmonic folds onto the base-rate Nyquist frequency: every tone came out with an image at 24 kHz minus its frequency (9.9 dB below a 10.1 kHz tone at -30 dBFS, 22 dB below a 5 kHz one at 48 kHz). The carrier now runs at a quarter of the internal rate, where all its harmonics fold onto 0, itself or the internal Nyquist frequency: the image measures 80 dB down, and the reference-level response, the odd-dominant saturation and the drive law are unchanged within their tests.Reverb (convolution) converts an impulse response held at another rate, or stretched, without changing what it does. The IR was streamed through the resampler: its gain moved with the rate ratio (+6.02 dB for a 48 kHz IR at 96 kHz, -6.02 dB the other way), the resampler's latency stayed inside the response (the direct sound 71 samples late from 44.1 to 48 kHz) and was never flushed, so a short IR came out as a fragment (-18.2 dB). It is now converted offline and time aligned over the kernel's whole reach, and scaled by the rate ratio: a unit impulse keeps a DC gain of 1.000000 with its peak exactly on getLatency(), for every rate pair, IR length and stretch. The kernel's ringing ahead of an IR that starts at full level is kept and declared as latency (112 samples for 44.1 into 48 kHz), with the dry path delayed to match; an IR with pre-delay holds it in its own silence and adds none.LoudnessMeter::getLoudnessRange() percentiles follow EBU Tech 3342: 0-based rank round((n - 1) * p), never the relative-gate threshold, and short-term values sampled at 10 Hz (the minimum Tech 3342 has required since V3). A steady tone below 12 s read LRA 20 LU and now reads 0; on a music fragment the reading matches the Tech 3342 reference code within 0.06 LU.wav_process example compensates its chain latency, so the written file stays aligned with the source and keeps its tail.Changed defaults, plugin timing and packaging are covered in the v1.8.0 migration guide.
OnsetDetector, BeatTracker and LoopFinder for transient analysis, tempo/phase tracking and bounded crossfade-ready loop discovery.LoudnessNormalizer for offline LUFS normalization under a true-peak ceiling.TimeStretch, PitchCorrector and SpectralFreeze for phase-vocoder time, pitch and spectral processing. Spectral freeze retains captured magnitudes while phases advance, are reconstructed or are decorrelated according to the selected mode.FlacFile for dependency-free native FLAC decoding, MidiFile for Standard MIDI File reading and writing, and KeyDetector for major/minor key estimation.Biquad::setCoeffsNow() for coefficients computed by the stream owner and for single-threaded or offline processing.PitchDetector::getWindowSize().BiquadCoeffs is a non-template, double-precision coefficient set, and the Biquad recursion remains in double precision for both float and double buffers.AudioBufferView converting and pointer-array constructors express const-correctness as constraints, so type traits and requires expressions now report the legal conversion direction.AudioProcessor now describes a genuinely in-place processor. Read-only analysers no longer satisfy the concept.prepare() receives non-finite or non-positive sample rates, including PitchCorrector.AudioBuffer safely empty after an allocation failure instead of retaining dangling channel views.Source and latency changes are covered in the v1.7.0 migration guide.
AlgorithmicReverb Eco quality mode.ConvolutionReverb. 1.9.8