prepare/prepare_anchor.py

#!/usr/bin/env python3
"""Rebuild anchor (edited) audio from native audio + replay manifest.

Reproduces, for each record in ``anchor_manifest.jsonl``, the edit that
produced the anchor recording from its native source:

  addition      overlay sound events at SNR-calibrated volume
  deletion      hard silence or blurred masking of regions
  modification  region-scoped volume or speed change

Only ``ffmpeg``/``ffprobe`` are required. Replay is exact except for
blur_mask, whose rendering params were randomized in the original pipeline
and are replaced with fixed defaults (the masked regions stay exact, which
is what the questions depend on).

Usage:
    python prepare_anchor.py [--manifest metadata/media/anchor/anchor_manifest.jsonl]
                             [--root .] [--only OPERATION] [--overwrite] [--dry-run]
"""

import argparse
import json
import math
import os
import re
import shutil
import subprocess
import sys
import tempfile
from pathlib import Path

def ffprobe_format(path: Path) -> tuple[int, int]:
    """Return (sample_rate, channels) of the first audio stream."""
    proc = subprocess.run(
        ["ffprobe", "-v", "error", "-select_streams", "a:0",
         "-show_entries", "stream=sample_rate,channels",
         "-of", "csv=p=0", str(path)],
        stdout=subprocess.PIPE, stderr=subprocess.PIPE,
    )
    rate, ch = proc.stdout.decode().strip().split(",")
    return int(rate), int(ch)


def read_jsonl(path: Path) -> list[dict]:
    with open(path, encoding="utf-8") as fh:
        return [json.loads(line) for line in fh if line.strip()]


def run(cmd: list[str], dry_run: bool = False) -> None:
    if dry_run:
        print("    [dry-run] " + " ".join(cmd[:8]) + " ...")
        return
    proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
    if proc.returncode != 0:
        err = proc.stderr.decode("utf-8", errors="replace").strip()
        raise RuntimeError(f"command failed (rc={proc.returncode}): {' '.join(cmd[:6])}...\n{err[-800:]}")


def ffprobe_duration(path: Path) -> float:
    proc = subprocess.run(
        ["ffprobe", "-v", "error", "-show_entries", "format=duration",
         "-of", "default=noprint_wrappers=1:nokey=1", str(path)],
        stdout=subprocess.PIPE, stderr=subprocess.PIPE,
    )
    return float(proc.stdout.decode().strip())


def analyze_rms_db(path: Path, start: float | None = None, dur: float | None = None) -> float:
    """Return RMS loudness (dB) of a file or a region of it."""
    cmd = ["ffmpeg", "-hide_banner", "-loglevel", "info"]
    if start is not None:
        cmd += ["-ss", f"{start}"]
    if dur is not None:
        cmd += ["-t", f"{dur}"]
    cmd += ["-i", str(path), "-af", "volumedetect", "-f", "null", "-"]
    proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
    text = proc.stderr.decode("utf-8", errors="replace")
    m = re.search(r"mean_volume: ([-\d.]+) dB", text)
    return float(m.group(1)) if m else -30.0


def out_codec_args(rate: int | None = None, channels: int | None = None) -> list[str]:
    args = ["-c:a", "flac"]
    if rate is not None:
        args += ["-ar", str(rate)]
    if channels is not None:
        args += ["-ac", str(channels)]
    return args


def resolve_event(manifest_path: str, events_dir: Path) -> Path:
    """Map a manifest event path to the local audio/events/ file."""
    return events_dir / Path(manifest_path).name


def replay_addition(rec: dict, native: Path, out: Path, events_dir: Path, dry_run: bool,
                    fmt: tuple[int, int] | None = None) -> None:
    positions = rec["params"]["positions"]
    inputs = ["-i", str(native)]
    filter_parts = []
    mix_inputs = ["[0:a]"]

    for idx, pos in enumerate(positions, start=1):
        event_path = resolve_event(pos["event_audio_path"], events_dir)
        if not event_path.is_file() and not dry_run:
            raise FileNotFoundError(f"event audio missing: {event_path}")
        inputs += ["-i", str(event_path)]

        offset_ms = int(float(pos["position_sec"]) * 1000)
        volume_db = pos.get("volume_db")
        if volume_db is None:
            snr = float(pos["snr_db"])
            region_dur = float(pos.get("duration_sec") or 3.0)
            bg_rms = analyze_rms_db(native, float(pos["position_sec"]), region_dur) if not dry_run else -25.0
            ev_rms = analyze_rms_db(event_path) if not dry_run else -25.0
            volume_db = round((bg_rms + snr) - ev_rms, 1)
            volume_db = max(-30.0, min(30.0, volume_db))

        label = f"evt{idx}"
        filter_parts.append(f"[{idx}:a]adelay={offset_ms}|{offset_ms},volume={volume_db}dB[{label}]")
        mix_inputs.append(f"[{label}]")

    weights = " ".join(["1"] * len(mix_inputs))
    filter_parts.append(
        f"{''.join(mix_inputs)}amix=inputs={len(mix_inputs)}:duration=first"
        f":dropout_transition=0:normalize=0:weights={weights}[mixed]"
    )
    rate, channels = fmt if fmt else (None, None)
    cmd = ["ffmpeg", "-hide_banner", "-loglevel", "error", "-y", *inputs,
           "-filter_complex", ";".join(filter_parts), "-map", "[mixed]",
           *out_codec_args(rate, channels), str(out)]
    run(cmd, dry_run)


def replay_deletion(rec: dict, native: Path, out: Path, dry_run: bool) -> None:
    mask_type = rec["params"].get("mask_type", "hard_mask")
    regions = rec["params"]["regions"]

    if mask_type == "hard_mask":
        filters = [f"volume=enable='between(t,{s},{e})':volume=0" for s, e in regions]
        cmd = ["ffmpeg", "-hide_banner", "-loglevel", "error", "-y", "-i", str(native),
               "-af", ",".join(filters), *out_codec_args(), str(out)]
        run(cmd, dry_run)
        return

    if mask_type == "blur_mask":
        _replay_blur_mask(native, out, regions, dry_run)
        return

    raise ValueError(f"unsupported mask_type: {mask_type}")


def _replay_blur_mask(native: Path, out: Path, regions: list, dry_run: bool) -> None:
    """Deterministic blur mask (original used randomized params; regions are exact)."""
    lowpass_freq = 300
    blur_attenuation_db = -6.0
    noise_level_db = -8.0
    fade_sec = 0.5
    noise_color = "pink"
    dynamic_noise = True
    fluct_db = 3.0
    fluct_rate = 0.8

    blur_gain = 10 ** (blur_attenuation_db / 20)
    noise_gain = 10 ** (noise_level_db / 20)

    if dynamic_noise:
        lo = noise_gain * (10 ** (-fluct_db / 20))
        hi = noise_gain * (10 ** (+fluct_db / 20))
        rng = hi - lo
        r1 = 2 * math.pi * fluct_rate
        r2 = 2 * math.pi * (fluct_rate * 2.3)
        noise_expr = f"{lo:.8f}+({rng:.8f})*(0.5+0.3*sin({r1}*t)+0.2*sin({r2}*t))"
    else:
        noise_expr = f"{noise_gain:.6f}"

    valid = [(float(s), float(e)) for s, e in regions if float(e) > float(s)]
    total_duration = ffprobe_duration(native) if not dry_run else 300.0

    orig_parts, blur_parts = [], []
    for s, e in valid:
        seg_dur = e - s
        eff_fade = min(fade_sec, seg_dur / 2.5)
        fie, fos = s + eff_fade, e - eff_fade
        orig_parts.append(
            f"if(between(t,{s},{fie}),(({fie}-t)/{eff_fade:.4f}),"
            f"if(between(t,{fie},{fos}),0,"
            f"if(between(t,{fos},{e}),((t-{fos})/{eff_fade:.4f}),1)))"
        )
        blur_parts.append(
            f"if(between(t,{s},{fie}),((t-{s})/{eff_fade:.4f}),"
            f"if(between(t,{fie},{fos}),1,"
            f"if(between(t,{fos},{e}),(({e}-t)/{eff_fade:.4f}),0)))"
        )

    if len(valid) == 1:
        orig_env, blur_env = orig_parts[0], blur_parts[0]
    else:
        orig_env, blur_env = "1", "0"
        for i in range(len(valid) - 1, -1, -1):
            s, e = valid[i]
            orig_env = f"if(between(t,{s},{e}),{orig_parts[i]},{orig_env})"
            blur_env = f"if(between(t,{s},{e}),{blur_parts[i]},{blur_env})"

    filter_complex = (
        f"[0:a]asplit=2[orig][to_blur];"
        f"[to_blur]lowpass=f={lowpass_freq},volume={blur_gain:.6f}[blurred];"
        f"[1:a]volume='{noise_expr}':eval=frame[noise_scaled];"
        f"[blurred][noise_scaled]amix=inputs=2:duration=first:weights=3 1[blur_noise];"
        f"[orig]volume='{orig_env}':eval=frame[orig_env];"
        f"[blur_noise]volume='{blur_env}':eval=frame[blur_env];"
        f"[orig_env][blur_env]amix=inputs=2:duration=first:normalize=0[out]"
    )
    cmd = ["ffmpeg", "-hide_banner", "-loglevel", "error", "-y",
           "-i", str(native),
           "-f", "lavfi", "-i",
           f"anoisesrc=color={noise_color}:duration={total_duration:.2f}:sample_rate=16000",
           "-filter_complex", filter_complex, "-map", "[out]",
           *out_codec_args(), str(out)]
    run(cmd, dry_run)


def _atempo_chain(factor: float) -> str:
    parts = []
    remaining = factor
    while remaining > 2.0:
        parts.append("atempo=2.0")
        remaining /= 2.0
    while remaining < 0.5:
        parts.append("atempo=0.5")
        remaining /= 0.5
    parts.append(f"atempo={remaining:.4f}")
    return ",".join(parts)


def replay_modification(rec: dict, native: Path, out: Path, dry_run: bool) -> None:
    kind = rec["params"]["kind"]
    regions = rec["params"]["regions"]

    if kind == "volume":
        filters = []
        for region in regions:
            s, e = region["region_sec"]
            gain = 10 ** (float(region["delta_db"]) / 20)
            filters.append(f"volume=enable='between(t,{s},{e})':volume={gain}")
        cmd = ["ffmpeg", "-hide_banner", "-loglevel", "error", "-y", "-i", str(native),
               "-af", ",".join(filters), *out_codec_args(), str(out)]
        run(cmd, dry_run)
        return

    if kind == "speed":
        _replay_speed(native, out, regions, dry_run)
        return

    raise ValueError(f"unsupported modification kind: {kind}")


def _replay_speed(native: Path, out: Path, regions: list, dry_run: bool) -> None:
    """Apply region-scoped atempo and re-concatenate. Regions use original timeline."""
    total = ffprobe_duration(native) if not dry_run else 600.0
    segs = sorted(regions, key=lambda r: r["region_sec"][0])

    # Build a clip plan: (kind, start, end, factor)
    plan = []
    cursor = 0.0
    for region in segs:
        s, e = region["region_sec"]
        factor = float(region["speed_factor"])
        if s > cursor + 0.05:
            plan.append(("copy", cursor, s, None))
        plan.append(("atempo", s, e, factor))
        cursor = e
    if cursor < total - 0.05:
        plan.append(("copy", cursor, total, None))

    if dry_run:
        print(f"    [dry-run] speed modification: {len(plan)} clips")
        return

    tmpdir = tempfile.mkdtemp(prefix="anchor_speed_")
    try:
        clip_paths = []
        for i, (kind, s, e, factor) in enumerate(plan):
            seg_path = os.path.join(tmpdir, f"seg_{i:03d}.flac")
            extract = ["ffmpeg", "-hide_banner", "-loglevel", "error", "-y",
                       "-ss", f"{s}", "-to", f"{e}", "-i", str(native),
                       *out_codec_args(), seg_path]
            run(extract)
            if kind == "atempo":
                tempo_path = os.path.join(tmpdir, f"seg_{i:03d}_tempo.flac")
                run(["ffmpeg", "-hide_banner", "-loglevel", "error", "-y",
                     "-i", seg_path, "-af", _atempo_chain(factor),
                     *out_codec_args(), tempo_path])
                clip_paths.append(tempo_path)
            else:
                clip_paths.append(seg_path)

        concat_list = os.path.join(tmpdir, "concat.txt")
        with open(concat_list, "w", encoding="utf-8") as fh:
            for p in clip_paths:
                fh.write(f"file '{p}'\n")
        run(["ffmpeg", "-hide_banner", "-loglevel", "error", "-y",
             "-f", "concat", "-safe", "0", "-i", concat_list,
             *out_codec_args(), str(out)])
    finally:
        shutil.rmtree(tmpdir, ignore_errors=True)


def process_record(rec: dict, root: Path, overwrite: bool, dry_run: bool) -> str:
    media_id = rec["media_id"]
    operation = rec["operation"]
    native = root / "audio" / "native" / Path(rec["native_audio_path"]).name
    out = root / "audio" / "anchor" / Path(rec["target_audio_path"]).name
    events_dir = root / "audio" / "events"

    if out.is_file() and not overwrite and not dry_run:
        return "skip"
    if not native.is_file() and not dry_run:
        return "missing-native"

    out.parent.mkdir(parents=True, exist_ok=True)
    if operation == "addition":
        fmt = None if dry_run else ffprobe_format(native)
        replay_addition(rec, native, out, events_dir, dry_run, fmt)
    elif operation == "deletion":
        replay_deletion(rec, native, out, dry_run)
    elif operation == "modification":
        replay_modification(rec, native, out, dry_run)
    else:
        return f"unknown-operation:{operation}"
    return "ok"


def main():
    parser = argparse.ArgumentParser(description="Rebuild anchor audio from manifest")
    parser.add_argument("--manifest", default="metadata/media/anchor/anchor_manifest.jsonl")
    parser.add_argument("--root", default=".", help="release root (contains audio/ + prepare/)")
    parser.add_argument("--only", choices=["addition", "deletion", "modification"],
                        help="process only this operation")
    parser.add_argument("--overwrite", action="store_true")
    parser.add_argument("--dry-run", action="store_true")
    parser.add_argument("--limit", type=int, default=0, help="process at most N records (debug)")
    args = parser.parse_args()

    root = Path(args.root).resolve()
    manifest_path = root / args.manifest
    if not manifest_path.is_file():
        print(f"ERROR: manifest not found: {manifest_path}", file=sys.stderr)
        sys.exit(1)

    records = read_jsonl(manifest_path)
    if args.only:
        records = [r for r in records if r["operation"] == args.only]
    if args.limit:
        records = records[:args.limit]

    print(f"Processing {len(records)} anchor records (root={root})")
    counts = {}
    for rec in records:
        status = process_record(rec, root, args.overwrite, args.dry_run)
        counts[status] = counts.get(status, 0) + 1
        if status not in ("ok", "skip"):
            print(f"  [{status}] {rec['media_id']} ({rec['operation']})")

    print("Summary:", counts)


if __name__ == "__main__":
    main()