Files
uj-mastering-master/metrics.py
T
Mikkeli Matlock a322f08d0c Add spectrogram, native-rate loading, and always-labelled axis extremes
- SpectrogramMetric: log-frequency STFT power heatmap over time, magma
  colormap, -80 dB floor. Adaptive hop caps time bins at ~4000 so long
  tracks stay responsive on redraw; N_FFT=4096 keeps low-freq resolution.
- master_core: load audio at native sample rate (librosa.load sr=None)
  instead of librosa's 22050 Hz default, so the full band up to the
  file's own nyquist (~22 kHz at 44.1 kHz) is analysed. ~2x heavier on
  44.1/48 kHz files, by design.
- metrics: shared _show_axis_extents helper forces each axis's exact
  min/max onto the tick list with compact labels (_fmt_tick), so the
  true range is always readable -- notably the spectrogram's 22 kHz top,
  which otherwise sits unlabelled between log-scale decade ticks. Applied
  to all metric renders.
- Docs: README + CLAUDE updated for the new metric, native-rate loading,
  and axis-readability behaviour.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-06-07 00:06:50 +09:00

573 lines
19 KiB
Python

"""
Pluggable analysis metrics.
A `Metric` knows how to compute a series from an `AudioFile` and how to render
that series into a matplotlib `Figure`. Compute is the heavy step (runs on the
worker thread); render is cheap and reruns on font / refresh.
To add a metric: subclass `Metric`, implement `compute` and `render`, and
register the instance in `METRICS` at the bottom of this file.
"""
from __future__ import annotations
import os
import warnings
from abc import ABC, abstractmethod
from typing import Any
import numpy as np
import matplotlib.colors as mcolors
import matplotlib.cm as cm
from matplotlib.figure import Figure
from matplotlib.ticker import FuncFormatter, NullFormatter
import librosa
import pyloudnorm as pyln
from scipy import signal as scipy_signal
from font_manager import safe_title
from master_core import AudioFile
# Small constant to keep 20*log10(...) from blowing up on perfect silence.
_EPS = 1e-12
def _to_dbfs(linear: np.ndarray | float) -> np.ndarray | float:
"""Convert a linear magnitude to dBFS, floored at _EPS."""
return 20.0 * np.log10(np.maximum(linear, _EPS))
def _fmt_tick(v, _pos=None) -> str:
"""Compact tick label: integer for big/whole values, trimmed decimals else."""
av = abs(v)
if v == 0 or av >= 100:
return f"{v:.0f}"
if av >= 1:
return f"{v:.1f}".rstrip("0").rstrip(".")
return f"{v:.3f}".rstrip("0").rstrip(".")
def _show_axis_extents(ax) -> None:
"""Force the exact min/max of each axis onto the tick list.
Matplotlib's locators often omit the extreme values — most visibly on a log
frequency axis, where the top (e.g. 22050 Hz) falls between decade ticks and
goes unlabelled. Union the endpoints into the existing in-range ticks so you
can always read where a plot actually starts and stops.
"""
fmt = FuncFormatter(_fmt_tick)
for is_log, get_lim, set_lim, get_ticks, set_ticks, mpl_axis in (
(ax.get_xscale() == "log", ax.get_xlim, ax.set_xlim, ax.get_xticks, ax.set_xticks, ax.xaxis),
(ax.get_yscale() == "log", ax.get_ylim, ax.set_ylim, ax.get_yticks, ax.set_yticks, ax.yaxis),
):
lo, hi = get_lim()
inside = [t for t in get_ticks() if lo <= t <= hi]
mpl_axis.set_major_formatter(fmt)
if is_log:
mpl_axis.set_minor_formatter(NullFormatter()) # keep minor marks unlabelled
set_ticks(sorted(set(inside) | {lo, hi}))
set_lim(lo, hi) # set_ticks can nudge the view; restore exact limits
class Metric(ABC):
"""A pluggable analysis metric."""
id: str
display_name: str
@abstractmethod
def compute(self, audio_file: AudioFile) -> Any:
"""Compute and return the metric's data from a loaded AudioFile.
The returned object is cached and later passed to `render`. This is the
heavy step and runs on the worker thread.
"""
@abstractmethod
def render(self, data: Any, file_path: str, figsize=(10, 4)) -> Figure:
"""Render a Figure from precomputed data. Cheap; runs on the GUI thread."""
class RMSPowerMetric(Metric):
id = "rms_power"
display_name = "RMS Power"
def __init__(self, window: int = 10, hop: int = 2):
self.window = window
self.hop = hop
def compute(self, audio_file: AudioFile):
audio_file.get_energy_levels_over_time(window=self.window, hop=self.hop)
return {
"times": audio_file.get_times(),
"rms_array": audio_file.rms_array,
}
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
times = data["times"]
rms_array = data["rms_array"]
# Adaptive colour scale: bump headroom for loud masters.
maxpower = 0.6 if np.max(rms_array) > 0.3 else 0.3
norm = mcolors.Normalize(vmin=0, vmax=maxpower)
cmap = cm.autumn
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
ax.set_ylim(0., maxpower)
for i in range(len(times) - 1):
ax.fill_between(
times[i:i + 2], 0, rms_array[0][i],
color=cmap(norm(rms_array[0][i])), edgecolor="none",
)
sm = cm.ScalarMappable(cmap=cmap, norm=norm)
sm.set_array([])
fig.colorbar(sm, ax=ax, label="RMS Power")
ax.set_ylabel("Power")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
_show_axis_extents(ax)
fig.tight_layout()
return fig
class WaveformMetric(Metric):
"""Raw mono waveform with a min/max envelope downsample for plotting speed."""
id = "waveform"
display_name = "Waveform"
def __init__(self, target_columns: int = 4000):
self.target_columns = target_columns
def compute(self, audio_file: AudioFile):
y = audio_file.y_mono
sr = audio_file.sr
n = len(y)
if n <= self.target_columns:
times = np.arange(n) / sr
return {"times": times, "lo": y, "hi": y}
chunk = n // self.target_columns
trimmed = y[: chunk * self.target_columns]
reshaped = trimmed.reshape(self.target_columns, chunk)
lo = reshaped.min(axis=1)
hi = reshaped.max(axis=1)
times = (np.arange(self.target_columns) * chunk + chunk / 2) / sr
return {"times": times, "lo": lo, "hi": hi}
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
times = data["times"]
lo = data["lo"]
hi = data["hi"]
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
ax.fill_between(times, lo, hi, color="#3a7ad6", linewidth=0)
ax.axhline(0, color="black", linewidth=0.5, alpha=0.3)
# Fixed full-scale range with a touch of headroom for float-wav signals.
ax.set_ylim(-1.1, 1.1)
ax.set_xlim(times[0], times[-1])
ax.set_ylabel("Amplitude")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
_show_axis_extents(ax)
fig.tight_layout()
return fig
class LUFSMetric(Metric):
"""ITU-R BS.1770 loudness: short-term (3 s) time series + integrated + LRA.
Powered by pyloudnorm. The time series slides `meter.integrated_loudness`
across the track because pyloudnorm doesn't expose a per-block series.
Slightly redundant work, but the per-call cost is small.
"""
id = "lufs"
display_name = "LUFS"
# Short-term as defined by EBU R128 / BS.1770: 3-second window.
WINDOW_S = 3.0
HOP_S = 0.5
SILENCE_FLOOR = -70.0 # BS.1770 absolute gate
def compute(self, audio_file: AudioFile):
y = audio_file.y_mono.astype(np.float64, copy=False)
sr = audio_file.sr
meter = pyln.Meter(sr)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
integrated = self._safe_integrated(meter, y)
window_n = int(self.WINDOW_S * sr)
hop_n = int(self.HOP_S * sr)
if len(y) < window_n:
times = np.array([len(y) / (2.0 * sr)])
lufs = np.array([integrated if np.isfinite(integrated) else self.SILENCE_FLOOR])
lra = float("nan")
else:
n_windows = 1 + (len(y) - window_n) // hop_n
lufs = np.empty(n_windows)
for i in range(n_windows):
start = i * hop_n
lufs[i] = self._safe_integrated(meter, y[start:start + window_n])
times = (np.arange(n_windows) * hop_n + window_n / 2.0) / sr
try:
lra = float(meter.loudness_range(y))
except (ValueError, FloatingPointError):
lra = float("nan")
lufs = np.where(np.isfinite(lufs), lufs, self.SILENCE_FLOOR)
lufs = np.clip(lufs, self.SILENCE_FLOOR, 0.0)
return {
"times": times,
"lufs": lufs,
"integrated": float(integrated),
"lra": lra,
}
@staticmethod
def _safe_integrated(meter: "pyln.Meter", segment: np.ndarray) -> float:
try:
return float(meter.integrated_loudness(segment))
except (ValueError, FloatingPointError):
return float("-inf")
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
times = data["times"]
lufs = data["lufs"]
integrated = data["integrated"]
lra = data.get("lra", float("nan"))
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
ax.plot(times, lufs, color="#2a9d8f", linewidth=1.4, label="Short-term (3 s)")
if np.isfinite(integrated):
ax.axhline(
integrated, color="#e76f51", linestyle="--", linewidth=1.5,
label=f"Integrated: {integrated:.1f} LUFS",
)
if np.isfinite(lra):
# Invisible plot entry to surface LRA in the legend without adding a line.
ax.plot([], [], " ", label=f"LRA: {lra:.1f} LU")
# Streaming target reference (Spotify normalises to -14 LUFS).
ax.axhline(-14.0, color="gray", linestyle=":", linewidth=0.8, alpha=0.6)
ax.text(
times[-1], -14.0, " -14 LUFS (streaming target)",
va="center", ha="left", fontsize=8, alpha=0.6,
)
ax.set_ylim(-50.0, 0.0)
ax.set_xlim(times[0], times[-1])
ax.set_ylabel("LUFS")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
ax.grid(True, alpha=0.3)
ax.legend(loc="lower right", fontsize=8)
_show_axis_extents(ax)
fig.tight_layout()
return fig
class CrestFactorMetric(Metric):
"""Crest factor = 20*log10(peak / RMS) per sliding window, in dB."""
id = "crest_factor"
display_name = "Crest Factor"
WINDOW_S = 1.0
HOP_S = 0.25
def compute(self, audio_file: AudioFile):
y = audio_file.y_mono.astype(np.float64, copy=False)
sr = audio_file.sr
window_n = int(self.WINDOW_S * sr)
hop_n = int(self.HOP_S * sr)
if len(y) < window_n:
times = np.array([len(y) / (2.0 * sr)])
peak = float(np.max(np.abs(y))) if len(y) else 0.0
rms = float(np.sqrt(np.mean(y * y))) if len(y) else 0.0
crest = 20.0 * np.log10(max(peak, _EPS) / max(rms, _EPS))
return {"times": times, "crest_db": np.array([crest])}
# RMS via cumulative-sum-of-squares (O(N)); peaks via sliding window view.
y2 = y * y
cumsum = np.concatenate(([0.0], np.cumsum(y2)))
n_windows = 1 + (len(y) - window_n) // hop_n
starts = np.arange(n_windows) * hop_n
ends = starts + window_n
mean_sq = (cumsum[ends] - cumsum[starts]) / window_n
rms = np.sqrt(np.maximum(mean_sq, _EPS))
abs_y = np.abs(y)
peaks = np.empty(n_windows)
for i in range(n_windows):
peaks[i] = np.max(abs_y[starts[i]:ends[i]])
crest_db = 20.0 * np.log10(np.maximum(peaks, _EPS) / rms)
times = (starts + window_n / 2.0) / sr
return {"times": times, "crest_db": crest_db}
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
times = data["times"]
crest_db = data["crest_db"]
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
ax.plot(times, crest_db, color="#e09f3e", linewidth=1.4, label=f"Crest factor (1 s)")
# Rules of thumb: ~12 dB = roomy, ~6 dB = heavily limited.
ax.axhline(12.0, color="gray", linestyle=":", linewidth=0.8, alpha=0.6)
ax.text(times[-1], 12.0, " 12 dB", va="center", ha="left", fontsize=8, alpha=0.6)
ax.axhline(6.0, color="gray", linestyle=":", linewidth=0.8, alpha=0.6)
ax.text(times[-1], 6.0, " 6 dB (squashed)", va="center", ha="left", fontsize=8, alpha=0.6)
ax.set_ylim(0.0, 25.0)
ax.set_xlim(times[0], times[-1])
ax.set_ylabel("Crest factor (dB)")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
ax.grid(True, alpha=0.3)
ax.legend(loc="lower right", fontsize=8)
_show_axis_extents(ax)
fig.tight_layout()
return fig
class PSRMetric(Metric):
"""Peak-to-Short-term LUFS Ratio (sample-peak variant), in LU.
PSR = sample_peak_dBFS - short_term_LUFS over the same 3 s windows used by
LUFSMetric. High PSR = punchy transients; low PSR = heavily limited.
"""
id = "psr"
display_name = "PSR"
WINDOW_S = 3.0
HOP_S = 0.5
SILENCE_FLOOR = -70.0
def compute(self, audio_file: AudioFile):
y = audio_file.y_mono.astype(np.float64, copy=False)
sr = audio_file.sr
meter = pyln.Meter(sr)
window_n = int(self.WINDOW_S * sr)
hop_n = int(self.HOP_S * sr)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
if len(y) < window_n:
times = np.array([len(y) / (2.0 * sr)])
peak_db = _to_dbfs(np.max(np.abs(y))) if len(y) else self.SILENCE_FLOOR
lufs = LUFSMetric._safe_integrated(meter, y)
psr = peak_db - lufs if np.isfinite(lufs) else 0.0
return {"times": times, "psr": np.array([psr])}
n_windows = 1 + (len(y) - window_n) // hop_n
abs_y = np.abs(y)
lufs_series = np.empty(n_windows)
peaks_db = np.empty(n_windows)
for i in range(n_windows):
start = i * hop_n
end = start + window_n
peaks_db[i] = _to_dbfs(np.max(abs_y[start:end]))
lufs_series[i] = LUFSMetric._safe_integrated(meter, y[start:end])
times = (np.arange(n_windows) * hop_n + window_n / 2.0) / sr
# PSR is meaningless where the loudness reading is below the absolute gate.
valid = np.isfinite(lufs_series) & (lufs_series > self.SILENCE_FLOOR)
psr = np.where(valid, peaks_db - lufs_series, np.nan)
return {"times": times, "psr": psr}
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
times = data["times"]
psr = data["psr"]
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
ax.plot(times, psr, color="#7251b5", linewidth=1.4, label="PSR (3 s)")
# Ian Shepherd's rough thresholds.
ax.axhline(10.0, color="gray", linestyle=":", linewidth=0.8, alpha=0.6)
ax.text(times[-1], 10.0, " 10 LU (good punch)", va="center", ha="left", fontsize=8, alpha=0.6)
ax.axhline(4.0, color="gray", linestyle=":", linewidth=0.8, alpha=0.6)
ax.text(times[-1], 4.0, " 4 LU (squashed)", va="center", ha="left", fontsize=8, alpha=0.6)
ax.set_ylim(0.0, 25.0)
ax.set_xlim(times[0], times[-1])
ax.set_ylabel("PSR (LU)")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
ax.grid(True, alpha=0.3)
ax.legend(loc="lower right", fontsize=8)
_show_axis_extents(ax)
fig.tight_layout()
return fig
class TruePeakMetric(Metric):
"""ITU-R BS.1770 true peak via 4x polyphase oversampling, in dBTP.
Per-window true peak with a moderate hop so it renders quickly. Windows are
oversampled independently — slight edge under-detection at window boundaries
is masked by the 60% overlap.
"""
id = "true_peak"
display_name = "True Peak"
WINDOW_S = 0.25
HOP_S = 0.1
OVERSAMPLE = 4
def compute(self, audio_file: AudioFile):
y = audio_file.y_mono.astype(np.float32, copy=False)
sr = audio_file.sr
window_n = int(self.WINDOW_S * sr)
hop_n = int(self.HOP_S * sr)
if len(y) < window_n:
y_up = scipy_signal.resample_poly(y, self.OVERSAMPLE, 1) if len(y) else np.zeros(1, dtype=np.float32)
peak_db = _to_dbfs(np.max(np.abs(y_up))) if len(y_up) else -70.0
return {
"times": np.array([len(y) / (2.0 * sr)]),
"tp_db": np.array([peak_db]),
"integrated_tp_db": float(peak_db),
}
n_windows = 1 + (len(y) - window_n) // hop_n
tp_db = np.empty(n_windows)
for i in range(n_windows):
start = i * hop_n
w = y[start:start + window_n]
w_up = scipy_signal.resample_poly(w, self.OVERSAMPLE, 1)
tp_db[i] = _to_dbfs(np.max(np.abs(w_up)))
times = (np.arange(n_windows) * hop_n + window_n / 2.0) / sr
integrated_tp_db = float(np.max(tp_db))
return {"times": times, "tp_db": tp_db, "integrated_tp_db": integrated_tp_db}
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
times = data["times"]
tp_db = data["tp_db"]
integrated = data.get("integrated_tp_db", float("nan"))
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
ax.plot(times, tp_db, color="#c1121f", linewidth=1.0, label="True Peak (250 ms)")
# 0 dBTP = sample-level clip; -1 dBTP a common mastering ceiling.
ax.axhline(0.0, color="black", linestyle="--", linewidth=1.0, alpha=0.8)
ax.text(times[-1], 0.0, " 0 dBTP (clip)", va="center", ha="left", fontsize=8, alpha=0.7)
ax.axhline(-1.0, color="gray", linestyle=":", linewidth=0.8, alpha=0.6)
ax.text(times[-1], -1.0, " -1 dBTP (typical ceiling)", va="center", ha="left", fontsize=8, alpha=0.6)
if np.isfinite(integrated):
ax.plot([], [], " ", label=f"Max: {integrated:.2f} dBTP")
ax.set_ylim(-30.0, 6.0)
ax.set_xlim(times[0], times[-1])
ax.set_ylabel("dBTP")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
ax.grid(True, alpha=0.3)
ax.legend(loc="lower right", fontsize=8)
_show_axis_extents(ax)
fig.tight_layout()
return fig
class SpectrogramMetric(Metric):
"""Log-frequency STFT spectrogram: frequency power distribution over time.
Each column is the magnitude spectrum of a short window, plotted in serial
as a colour-coded heatmap. The hop is chosen adaptively so long tracks don't
produce tens of thousands of columns (which would stall the GUI redraw): for
typical song lengths the hop lands around 50 ms, coarsening gracefully on
very long files.
"""
id = "spectrogram"
display_name = "Spectrogram"
N_FFT = 4096 # ~11 Hz bins at 44.1 kHz; keeps low-freq detail now
# that sr is native (nyquist ~22 kHz, not 11 kHz)
TARGET_COLUMNS = 4000 # cap on time bins, for render speed
DB_FLOOR = -80.0 # dynamic range shown, relative to peak
F_MIN = 20.0 # log axis can't show DC; clip the low edge here
def compute(self, audio_file: AudioFile):
y = audio_file.y_mono.astype(np.float32, copy=False)
sr = audio_file.sr
# Pick a hop that keeps the column count near TARGET_COLUMNS, but never
# finer than n_fft//4 (the usual 75%-overlap floor).
min_hop = self.N_FFT // 4
hop = max(min_hop, len(y) // self.TARGET_COLUMNS)
stft = librosa.stft(y, n_fft=self.N_FFT, hop_length=hop)
mag = np.abs(stft)
s_db = librosa.amplitude_to_db(mag, ref=np.max)
freqs = librosa.fft_frequencies(sr=sr, n_fft=self.N_FFT)
times = librosa.frames_to_time(
np.arange(s_db.shape[1]), sr=sr, hop_length=hop, n_fft=self.N_FFT
)
# Drop the DC bin (0 Hz) so the log frequency axis has no non-positive coord.
return {
"freqs": freqs[1:],
"times": times,
"s_db": s_db[1:, :],
"nyquist": sr / 2.0,
}
def render(self, data, file_path, figsize=(10, 4)) -> Figure:
freqs = data["freqs"]
times = data["times"]
s_db = data["s_db"]
nyquist = data["nyquist"]
fig = Figure(figsize=figsize, facecolor="white")
ax = fig.add_subplot(111)
mesh = ax.pcolormesh(
times, freqs, s_db,
cmap="magma", vmin=self.DB_FLOOR, vmax=0.0, shading="auto",
)
fig.colorbar(mesh, ax=ax, label="Power (dB)")
ax.set_yscale("log")
ax.set_ylim(self.F_MIN, nyquist)
ax.set_xlim(times[0], times[-1])
ax.set_ylabel("Frequency (Hz)")
ax.set_xlabel("Time (seconds)")
ax.set_title(safe_title(os.path.basename(file_path)))
_show_axis_extents(ax)
fig.tight_layout()
return fig
METRICS: dict[str, Metric] = {
m.id: m for m in (
RMSPowerMetric(),
WaveformMetric(),
LUFSMetric(),
CrestFactorMetric(),
PSRMetric(),
TruePeakMetric(),
SpectrogramMetric(),
)
}
DEFAULT_METRIC_ID = "rms_power"