feat: add two-pass EBU R128 loudness normalization for audio extraction

config.toml

@@ -6,7 +6,7 @@
 file_a = "/home/honney/Videos/Drachenzähmen/Drachenzaehmen.mp4"
 file_b = "/home/honney/Videos/Drachenzähmen/How to Train Your Dragon.mkv"
 # Window to extract (seconds). This is the ONLY duration needed.
-duration = 90
+duration = 30
 # Start offset in file (seconds, 0 = file start). ffmpeg -ss before -i for fast seek.
 start = 0.0
 out_dir = "extracted"
@@ -20,6 +20,11 @@ channels = 1
 # If extracted wavs already exist, sync test can load them directly:
 # file_a_wav = "extracted/Drachenzaehmen_audio.wav"
 # file_b_wav = "extracted/How to Train Your Dragon_audio.wav"
+# EBU R128 loudness normalization (two-pass loudnorm)
+normalize = true
+target_lufs = -14       # -14 streaming, -23 broadcast
+target_tp = -1.0        # true peak limit (dBTP)
+target_lra = 11         # loudness range
 
 [sync]
 # Analysis window: null = use full extracted duration; otherwise limit for quick test

extract.py

@@ -11,6 +11,7 @@ import subprocess
 import os
 import sys
 import shutil
+import json
 
 # ─── Input / config (config.toml is base config, fallback to hardcoded) ───
 CONFIG_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "config.toml")
@@ -20,6 +21,10 @@ _cfg_start = 0.0
 _cfg_out_dir = os.path.join(os.path.dirname(os.path.abspath(__file__)), "extracted")
 _cfg_target_fps = "24000/1001"
 _cfg_eps = 0.02
+_cfg_normalize = True
+_cfg_target_lufs = -14
+_cfg_target_tp = -1.0
+_cfg_target_lra = 11
 try:
     try:
         import tomllib  # py3.11+
@@ -41,6 +46,24 @@ try:
                 _cfg_eps = float(_ex["target_fps_eps"])
             except Exception:
                 pass
+        _au = _cfg.get("audio", {})
+        if _au.get("normalize") is not None:
+            _cfg_normalize = bool(_au["normalize"])
+        if _au.get("target_lufs") is not None:
+            try:
+                _cfg_target_lufs = float(_au["target_lufs"])
+            except Exception:
+                pass
+        if _au.get("target_tp") is not None:
+            try:
+                _cfg_target_tp = float(_au["target_tp"])
+            except Exception:
+                pass
+        if _au.get("target_lra") is not None:
+            try:
+                _cfg_target_lra = float(_au["target_lra"])
+            except Exception:
+                pass
         if _cfg_inputs:
             print(f"[config] loaded {CONFIG_PATH}: duration={_cfg_duration}s start={_cfg_start}s target_fps={_cfg_target_fps}")
 except Exception as _e:
@@ -84,6 +107,10 @@ try:
 except Exception:
     TARGET_FPS_FLOAT = 24000 / 1001
 FPS_EPS = _cfg_eps  # tolerance for float comparison
+NORMALIZE = _cfg_normalize
+TARGET_LUFS = _cfg_target_lufs
+TARGET_TP = _cfg_target_tp
+TARGET_LRA = _cfg_target_lra
 
 def _parse_rate(s: str) -> float | None:
     s = s.strip()
@@ -123,6 +150,61 @@ def probe_fps(path: str) -> float | None:
 def fps_matches(a: float | None, b: float, eps: float = FPS_EPS) -> bool:
     return a is not None and abs(a - b) < eps
 
+def _parse_loudnorm_json(stderr: str) -> dict | None:
+    """Extract JSON block from loudnorm stderr output."""
+    try:
+        start = stderr.rindex("{")
+        end = stderr.rindex("}") + 1
+        return json.loads(stderr[start:end])
+    except (ValueError, json.JSONDecodeError):
+        return None
+
+def extract_audio_loudnorm(input_file: str, duration: int, start_sec: float,
+                           out_path: str, stem: str) -> None:
+    """Two-pass EBU R128 loudnorm extraction: measure then apply."""
+    ss_args = ["-ss", str(start_sec)] if start_sec and start_sec != 0 else []
+    base = ["ffmpeg", "-y", *ss_args, "-i", input_file, "-t", str(duration), "-vn"]
+    measure_af = f"loudnorm=I={TARGET_LUFS}:TP={TARGET_TP}:LRA={TARGET_LRA}:print_format=json"
+
+    # ─── Pass 1: measure ───
+    print(f"[loudness] {stem}: pass 1/2 measuring…")
+    cmd_measure = base + ["-af", measure_af, "-f", "null", "-"]
+    r = subprocess.run(cmd_measure, capture_output=True, text=True)
+    if r.returncode != 0:
+        print(f"[loudness] {stem}: pass 1 failed ({r.stderr.strip()[-80:]}), falling back to simple extract", file=sys.stderr)
+        cmd_fallback = base + [*AUDIO_CODEC, out_path]
+        r2 = subprocess.run(cmd_fallback, capture_output=True, text=True)
+        if r2.returncode != 0:
+            sys.exit(f"ffmpeg (audio) failed:\n{r2.stderr}")
+        return
+
+    measured = _parse_loudnorm_json(r.stderr)
+    if measured is None:
+        print(f"[loudness] {stem}: could not parse measurement, falling back to simple extract", file=sys.stderr)
+        cmd_fallback = base + [*AUDIO_CODEC, out_path]
+        r2 = subprocess.run(cmd_fallback, capture_output=True, text=True)
+        if r2.returncode != 0:
+            sys.exit(f"ffmpeg (audio) failed:\n{r2.stderr}")
+        return
+
+    mi = measured.get("input_i", "-24")
+    mtp = measured.get("input_tp", "-2")
+    mlra = measured.get("input_lra", "7")
+    mthresh = measured.get("input_thresh", "-34")
+    offset = measured.get("target_offset", "0")
+    print(f"[loudness] {stem}: measured I={mi} LUFS, TP={mtp} dBTP, LRA={mlra}, offset={offset} dB")
+
+    # ─── Pass 2: apply with measured values ───
+    print(f"[loudness] {stem}: pass 2/2 applying loudnorm → {TARGET_LUFS} LUFS…")
+    apply_af = (f"loudnorm=I={TARGET_LUFS}:TP={TARGET_TP}:LRA={TARGET_LRA}"
+                f":measured_I={mi}:measured_TP={mtp}:measured_LRA={mlra}"
+                f":measured_thresh={mthresh}:offset={offset}:linear=true")
+    cmd_apply = base + ["-af", apply_af, *AUDIO_CODEC, out_path]
+    r = subprocess.run(cmd_apply, capture_output=True, text=True)
+    if r.returncode != 0:
+        sys.exit(f"ffmpeg (audio) pass 2 failed:\n{r.stderr}")
+    print(f"[loudness] {stem}: normalized to {TARGET_LUFS} LUFS ✔")
+
 # ─── Check ffmpeg ───
 if shutil.which(f"ffmpeg") is None:
     sys.exit(f"ERROR: 'ffmpeg' not found on PATH. Install it and try again.")
@@ -194,7 +276,11 @@ for INPUT_FILE in INPUT_FILES:
         sys.exit(f"ffmpeg (video) failed:\n{r.stderr}")
     print(f"[video: {stem}] done ✔")
 
-    # ─── Extract audio-only (first 90s, no video, CPU PCM re-encode) ───
+    # ─── Extract audio-only (duration s, no video, CPU PCM re-encode) ───
+    if NORMALIZE:
+        print(f"[audio: {stem}] extracting with two-pass loudnorm…")
+        extract_audio_loudnorm(INPUT_FILE, DURATION, START_SEC, audio_out, stem)
+    else:
         cmd_audio = [
             "ffmpeg",
             "-y",
@@ -205,7 +291,7 @@ for INPUT_FILE in INPUT_FILES:
             *AUDIO_CODEC,       # pcm_f32le 48kHz mono, CPU (no GPU equiv.)
             audio_out
         ]
-    print(f"[audio: {stem}] running ffmpeg…")
+        print(f"[audio: {stem}] running ffmpeg (no normalization)…")
         r = subprocess.run(cmd_audio, capture_output=True, text=True)
         if r.returncode != 0:
             sys.exit(f"ffmpeg (audio) failed:\n{r.stderr}")