/**
 * The psychoacoustic model: how much quantisation noise each scalefactor band
 * can hide.
 *
 * Everything else in the encoder is arithmetic with a right answer. This is the
 * part that decides what a listener will not notice, and it is the difference
 * between an MP3 that is merely valid and one that sounds acceptable at 128 kbps.
 *
 * ## What it computes
 *
 * A loud tone raises the threshold of audibility around itself: noise that would
 * be obvious in silence is inaudible next to it. The model estimates that
 * threshold per band, and the quantiser is then free to be as coarse as the
 * threshold allows — spending bits where they are heard and saving them where
 * they are not.
 *
 * Four things set the threshold:
 *
 * 1. **Spreading.** Masking is not confined to the band that causes it; it
 *    spreads across the cochlea, further upward in frequency than downward. Band
 *    energies are convolved with a spreading function on the bark scale.
 * 2. **Tonality.** A pure tone masks much less noise than a noise band of the
 *    same energy does — roughly 29 dB against 6 dB. Tonality is estimated per
 *    band from spectral flatness: a flat band is noise-like, a peaky one tonal.
 * 3. **Absolute threshold.** Below the quiet threshold of hearing nothing is
 *    audible whatever else is happening, which matters most at the extremes of
 *    the spectrum where the ear is least sensitive.
 * 4. **Time.** A threshold may not rise faster than the ear's forward masking
 *    allows, or the quantiser would spread a transient's noise backwards in time
 *    into the silence before it — pre-echo, the most recognisable MP3 artefact.
 *
 * ## Why the partitions are finer than the bands
 *
 * Thresholds are wanted per scalefactor band, and computing them there directly
 * is the obvious shortcut. It does not work: a scalefactor band is roughly a
 * bark wide, and the spreading function is almost flat across ±1 bark, so a
 * quiet band adjacent to a loud one inherits nearly all of its energy and comes
 * out with a threshold at its own signal level. Measured that way the model
 * declared every band already masked at 128 kbps, and the encoder did nothing.
 *
 * So energy is partitioned at about a third of a bark, where the spreading
 * function's shape is actually resolved, and each scalefactor band then takes
 * the *lowest* threshold among the partitions it covers. Taking the minimum is
 * the conservative choice, and the right one: a band is only as maskable as its
 * most exposed part.
 *
 * ## Two deliberate simplifications
 *
 * **Tonality from spectral flatness**, not from the unpredictability measure of
 * ISO model 2. The unpredictability measure needs magnitude and phase histories
 * for two previous granules and is markedly better on transients; flatness needs
 * only the current spectrum. Short blocks are what really answers transients,
 * and those arrive with window switching.
 *
 * **Thresholds are used as ratios, not absolute levels.** The model works on an
 * FFT of the input while the quantiser works on MDCT lines, and the two do not
 * share a scale. Rather than calibrate between them — which would be a constant
 * fitted to this implementation, and wrong the moment either side changed — each
 * band's signal-to-mask ratio is applied to that band's *own* MDCT energy. The
 * absolute threshold of hearing is the one part that needs a real level, and it
 * assumes the usual convention that full scale is 96 dB SPL.
 */
/**
 * Marks the granules that contain an attack sharp enough to need short blocks.
 *
 * Pre-echo is a consequence of the transform's length: a long block spreads its
 * quantisation noise across all 576 samples it covers, so noise belonging to a
 * drum hit arrives up to 12 ms *before* the hit does, in what should be silence.
 * The ear notices that far more readily than noise during the hit itself.
 *
 * Detection compares each short segment's energy against a decaying average of
 * what came before. A ratio threshold rather than a fixed level, because the
 * effect is relative: a loud attack in loud music is masked, and a modest one
 * after silence is not.
 *
 * @param out One flag per granule. Sized by the caller.
 */
export declare function detectAttacks(input: Float32Array, granuleCount: number, out: Uint8Array): void;
/**
 * Turns attack flags into a legal sequence of block types.
 *
 * The format does not allow jumping between window shapes: a short block has to
 * be introduced by a start block and left by a stop block, because those are the
 * asymmetric windows whose overlap adds back to unity against their neighbours.
 * Emitting a short block straight after a long one produces a stream that parses
 * and decodes to a click.
 *
 * So a detected attack in granule N forces granule N-1 to become a start block —
 * one granule of lookahead, which is why detection runs over the whole input
 * before any of it is coded.
 *
 * @param attacks One flag per granule, consumed here.
 * @param out Block type per granule: 0 long, 1 start, 2 short, 3 stop.
 */
export declare function sequenceBlockTypes(attacks: Uint8Array, out: Uint8Array): void;
/**
 * Per-granule masking thresholds for one channel.
 *
 * Stateful — it remembers the previous granule's thresholds for the temporal
 * cap — so one instance per channel, never shared.
 */
export declare class PsychoacousticModel {
    #private;
    constructor(sampleRate: number);
    /** Forgets history, e.g. when starting a new stream. */
    reset(): void;
    /**
     * Masking thresholds for one granule, as allowed mean squared error per MDCT
     * line — the units {@link quantizeGranule} expects.
     *
     * @param input The channel's samples.
     * @param startSample First sample of the granule being coded.
     * @param spectrum The granule's 576 MDCT lines, which set the scale.
     * @param out 22 thresholds, one per scalefactor band.
     */
    analyse(input: Float32Array, startSample: number, spectrum: Float32Array, out: Float32Array, short?: boolean): void;
}