Sound textures

import json

import matplotlib.pyplot as plt
import numpy as np

import sonore as so
from sonore.texture import TextureStats

plt.rcParams.update({"font.size": 9, "axes.titlesize": 10, "figure.dpi": 100})
FS = 44100


def finish(snd):
    """How every sound in the gallery is played: 5 ms ramps, RMS 0.1, peak at most 0.95."""
    snd = snd.ramp(5e-3).normalize(rms=0.1)
    return snd.normalize(peak=0.95) if snd.peak > 0.95 else snd


def original(name):
    """A recording; docs/textures/SOURCES.md says where each comes from and which excerpt."""
    return so.load(f"docs/textures/{name}.flac")


def synthesized(name, imposed=""):
    """A synthesis precomputed by tools/make_texture_synths.py (below)."""
    suffix = f"__{imposed}" if imposed else ""
    return so.load(f"docs/textures/synth/{name}{suffix}.flac").resample(FS)


def report(name):
    """How closely the synthesis matches: the statistic SNR averaged over classes."""
    info = json.loads(open(f"docs/textures/synth/{name}.json").read())
    snr = np.mean(list(info["snr_all_classes"].values()))
    return f"Average statistic SNR {snr:.0f} dB after {info['best_iteration']} iterations."


def texture_fig(snd, original, synthetic=False):
    """General-purpose plots (spectrogram, long-term spectrum) plus the texture
    model's own statistics drawn with plain matplotlib: modulation power by
    band and rate (the model's modulation spectrum), envelope sparsity, and
    band-averaged modulation power. Dashed black: the original."""
    fig = plt.figure(figsize=(10, 6.6), layout="constrained")
    gs = fig.add_gridspec(2, 3)
    ax_s = fig.add_subplot(gs[0, :2])
    ax_l, ax_m, ax_v, ax_p = (fig.add_subplot(gs[r, c]) for r, c in ((0, 2), (1, 0), (1, 1), (1, 2)))
    so.STFT(snd, 20e-3).plot(ax_s, fmax=10000, db_range=70, colorbar=False)
    ax_s.set_title("Spectrogram")
    so.long_term_spectrum(snd).plot(ax_l, lw=1.2, label="this sound")
    if synthetic:
        so.long_term_spectrum(original).plot(ax_l, color="k", ls="--", lw=1, label="original")
        ax_l.legend(fontsize=7)
    ax_l.set(xlim=(20, 10000), ylim=(-70, 3), title="Long-term spectrum")
    target = TextureStats.measure(original)
    this = TextureStats.measure(snd, window="uniform" if synthetic else "ramped")  # syntheses are circular
    cfs = target.model.filterbank.cfs[1:-1]
    ok = target.channel_mask()
    # The model's modulation spectrum: power at each modulation rate, in each cochlear band.
    mcf = target.model.mod_bank.cfs
    lev = 10 * np.log10(np.maximum(this.mod_power[1:-1], 1e-6))
    ax_m.pcolormesh(mcf, cfs, lev, cmap="magma", vmin=lev.max() - 25, vmax=lev.max(), shading="nearest")
    ax_m.set(
        xscale="log",
        yscale="log",
        xlabel="Modulation rate [Hz]",
        ylabel="Band center [Hz]",
        title="Modulation spectrum by band",
    )
    ax_v.semilogx(cfs, this.env_var[1:-1], lw=1.2)
    ax_p.loglog(target.model.mod_bank.cfs, this.mod_power[ok].mean(0), lw=1.2)
    if synthetic:
        ax_v.semilogx(cfs, target.env_var[1:-1], "k--", lw=1)
        ax_p.loglog(target.model.mod_bank.cfs, target.mod_power[ok].mean(0), "k--", lw=1)
    ax_v.set(xlabel="Band center [Hz]", ylabel="var / mean²", title="Envelope sparsity by band")
    ax_p.set(
        xlabel="Modulation rate [Hz]", ylabel="Power / variance", title="Modulation power (band average)"
    )
    for ax in (ax_v, ax_p):
        ax.grid(ls=":", which="both", lw=0.5)
    return fig, [ax_s]  # the playhead follows the spectrogram

from sonore.texture import PAPER_CLASSES  # noqa: E402
from sonore.texture.synth import synthesize  # noqa: E402


def synthesize_like(name, classes=PAPER_CLASSES, seconds=5.0, iterations=30):
    """What tools/make_texture_synths.py does for each recording (not run here)."""
    target = TextureStats.measure(original(name))
    new, rep = synthesize(target, duration=seconds, classes=classes, rng=1, max_iter=iterations)
    return new * (0.05 / new.rms)