stzSound

(no comment in the source: the name reads as a sentence) initno extension

lifecycleQCSXTZZZIBWU

(no comment in the source: the name reads as a sentence) adopt bufferno extension

Hand the engine id out -- for a graph that wants this sound as a source. An ENGINE handle, not a pointer, so it obeys the house law.no extension

(no comment in the source: the name reads as a sentence) releaseno extension

(no comment in the source: the name reads as a sentence) is emptyno extension

(no comment in the source: the name reads as a sentence) last errorno extension

readers: they return DATA and keep their namesno extension

(no comment in the source: the name reads as a sentence) channelsno extension

(no comment in the source: the name reads as a sentence) sample rateno extension

(no comment in the source: the name reads as a sentence) durationno extension

(no comment in the source: the name reads as a sentence) peakno extension

(no comment in the source: the name reads as a sentence) rmsno extension

ONE sample. 1-based frame and channel, like every Haro face.no extension

(no comment in the source: the name reads as a sentence) set sample atQCSXTZZZIBWU

The samples of one channel, as a Haro list. DATA, so To... is right. Be aware this crosses the FFI once per sample -- fine for a look at a few thousand frames, wrong for a whole song.no extension

verbs that change the sound ---------------------------------------- Resample to a new rate. QUALITY is sinc unless you ask otherwise -- SN1 measured sinc at 97x the accuracy of linear for 26x the cost, and still 150x faster than real time, so it is the right default.QCSXTZZZIBWU

(no comment in the source: the name reads as a sentence) resample to fastQCSXTZZZIBWU

(no comment in the source: the name reads as a sentence) change channels toQCSXTZZZIBWU

(no comment in the source: the name reads as a sentence) to monoQCSXTZZZIBWU

(no comment in the source: the name reads as a sentence) to stereoQCSXTZZZIBWU

The symmetric half of loading from memory, and its absence was a real gap: a sound could be decoded FROM bytes but not encoded back TO them, so handing a buffer to another tier meant a temporary file. VC1 went to some trouble to keep a voice out of the filesystem; without this, VC3 would have put it straight back to feed a recogniser.no extension

16-bit by default: it is what everything reads, and SN1 measured the round-trip error at 4.18e-5, one quantum. Ask for 32 when you are handing the file back to another stage rather than to a listener. ADD another sound into this one from pnSeconds in, scaled by pnGain -- notes laid on a timeline. Refused when the sample rates differ: a note mixed at the wrong rate would be a different pitch, silently.no extension

(no comment in the source: the name reads as a sentence) save asQCSXTZZZIBWU

(no comment in the source: the name reads as a sentence) save as floatQCSXTZZZIBWU

hearing it ---------------------------------------------------------- PLAY IT, and wait for it to finish. There is no sample-player in the engine -- the device sink plays a STREAM -- so the face builds the smallest graph that can carry this sound and runs it. That is what a declarative face is for: the caller says "play", not "make a graph, start a producer thread, open a device, wait, then take it all down".no extension

(no comment in the source: the name reads as a sentence) play forQCSXTZZZIBWU

To...() because these RETURN things. A spectrum is a stzSoundGrid: rows by columns of numbers, with x_step and y_step saying what a row and a column MEAN. Drawing one is the graphics plane's job, and keeping that boundary is what makes the analysis assertable -- "the 1 kHz sine is in the 1 kHz bin" is a claim about numbers; "the picture looks right" is a claim about nothing. The magnitude spectrum of one window. Default 4096 points: about 12 Hz resolution at 48 kHz, which resolves musical pitch without smearing time.no extension

(no comment in the source: the name reads as a sentence) to spectrum ofno extension

Frequency against time: one row per window. MULTICORE by default -- every row is an independent FFT over a different slice of the same immutable buffer, which is the one analysis here that parallelises without an argument.no extension

(no comment in the source: the name reads as a sentence) to spectrogram ofno extension

The strongest frequency present, in hertz.no extension

Where notes START, in seconds -- a one-row grid of times.no extension

(no comment in the source: the name reads as a sentence) to onsets ofno extension

Beats per minute, or -1 when there are too few onsets to say -- which is a better answer than a confident number derived from two events.no extension

The standard's K-weighting coefficients are specified at 48 kHz, and the engine REFUSES any other rate rather than applying them anyway. So the face resamples a COPY when it has to -- the caller's sound is not touched, and nobody gets a plausible-looking wrong number.no extension

MomentaryLoudness is the standard's own 400 ms window, ungated, taking the LOUDEST such window rather than a mean -- averaging an earcon against the silence around it answers nothing. BS.1770 arithmetic used as specified, so it is LUFS without qualification.no extension

(no comment in the source: the name reads as a sentence) short term loudnessno extension

The same K-weighting and the same formula over the sound's OWN length. NOT a standard LUFS figure, and LoudnessMetric() says so -- because the alternative is answering the standard's question wrongly instead of answering a different question honestly. A 60 ms earcon at peak 0.5 reads -9.26 here against -9.27 for the same tone at 400 ms: the two agree to a hundredth of a decibel where the standard has an opinion, which is what makes this trustworthy where it does not.no extension

What the number means, so it never travels without its method.no extension