SoundGame/web/detective/gen/gen_sfx_sa3.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""《雨夜专案》Stable Audio 3 small-sfx 音效（批次 2，18-detective.md §8.2 SA3 条目）。

生成（GPU，flock 串行；模型 gated 走 hf-mirror，见 env.d/sfx.sh）：
  python3 gen_sfx_sa3.py --emit   # 打印生成命令，原样执行
生成 6 条原始 wav 到 gen/staging/sfx_sa3/detective_<i>.wav（顺序 = 下表 idx）。
本脚本随后做后处理：修整（签名 earcon 按 max_s 压进 0.3-1.0s 量级——条目切换/
指证选择前置音，长了拖慢导航）/去直流/8ms cos² 淡化/ 响度定档 → ogg q4 +
m4a 128k（highpass=f=30，one-shot 前后垫 60/120ms 静音）→ 解码回读 QC：逐声道
真峰值 <= -3dBFS、|直流| <= 0.001；超限降增益 + 首轮重编加 18kHz 低通，最多 3 轮。

prompt：paper_turn/gavel 直出文档英文原稿；sig_ge/sig_zhou/sig_scene/rain_swell
为文档中文拟声条的英文化（信源：18-detective.md §8.2）。定档：
  gavel -4（裁决仪式拍）、paper_turn -7（导航细节反馈）
  sig_* -6（签名前置音，可辨不抢人声）、rain_swell -10（认罪链垫，压在 bgm/verdict 下）

用法：.venv-tts/bin/python web/detective/gen/gen_sfx_sa3.py [--emit]
"""
import os
import subprocess
import sys

import numpy as np
import soundfile as sf

SR = 44100
HERE = os.path.dirname(os.path.abspath(__file__))
STAGING = os.path.join(HERE, "staging", "sfx_sa3")
DST = os.path.join(HERE, "..", "audio", "sfx")
SA3 = "/home/xsl/SoundGame/tools/stable-audio-3"
FFMPEG = "/home/xsl/SoundGame/tools/bin/ffmpeg"
SEED = 20260909

# (id, gen 秒, prompt, 定档 dBFS, trim 上限 s)
ITEMS = [
    ("paper_turn", 3.0, "single paper page turn on desk, dry, close", -7.0, 0.9),
    ("sig_ge", 3.0, "Enamel mug thumped down on a wooden desk, single dull metallic knock, close", -6.0, 0.7),
    ("sig_zhou", 3.0, "Wooden abacus beads flicked twice, dry clacking, close up", -6.0, 0.9),
    ("sig_scene", 4.0, "Leather shoes stepping through rain puddles twice, then one umbrella shake, night rain, close", -6.0, 1.1),
    ("gavel", 4.0, "wooden gavel single strike, small room reverb", -4.0, 3.0),
    ("rain_swell", 9.0, "Night rain swelling heavier then gradually settling, steady rain on windows and wet street, 6 seconds", -10.0, 7.0),
]

WIN = {}  # 事件窗（raw 定长 latent 床）：首轮自然修剪，实测偏离再回填


def emit():
    durs = " ".join(str(d) for _, d, *_ in ITEMS)
    ps = " ".join(f'"{p}"' for _, _, p, *_ in ITEMS)
    print(f"cd {SA3} && flock /home/xsl/SoundGame/.gpu.lock .venv/bin/stable-audio "
          f"--model small-sfx --steps 8 --seed {SEED} "
          f"-p {ps} --duration {durs} -o {STAGING}/detective.wav")


def trim_oneshot(y, floor_db=-50.0, head_ms=50, tail_ms=140, max_s=None):
    env = np.max(np.abs(y), axis=1)
    act = np.where(env > 10 ** (floor_db / 20))[0]
    a = max(0, act[0] - int(SR * head_ms / 1000))
    b = min(len(y), act[-1] + int(SR * tail_ms / 1000))
    if max_s:
        b = min(b, a + int(SR * max_s))
    return y[a:b]


def prep_oneshot(y, target_db, win=None, max_s=None):
    if win:
        a, b = int(SR * win[0]), int(SR * win[1])
        y = y[a:b]
        nf2 = max(1, int(SR * 0.050))
        y[-nf2:] *= (0.5 * (1 + np.cos(np.pi * np.arange(nf2) / nf2)))[:, None]
    else:
        y = trim_oneshot(y, max_s=max_s)
    y = y - np.mean(y, axis=0, keepdims=True)
    nf = max(1, int(SR * 0.008))
    fade = 0.5 * (1 - np.cos(np.pi * np.arange(nf) / nf))
    y[:nf] *= fade[:, None]
    y[-nf:] *= fade[::-1][:, None]
    y = y / (np.max(np.abs(y)) + 1e-9) * 10 ** (target_db / 20)
    pad_h = np.zeros((int(SR * 0.06), y.shape[1]))
    pad_t = np.zeros((int(SR * 0.12), y.shape[1]))
    return np.concatenate([pad_h, y, pad_t])


def encode(name, sig):
    base = sig.astype(np.float64)
    vol, lpf, dc_adj = 0.0, False, 0.0
    tmp = os.path.join(STAGING, "_enc.wav")
    chk = os.path.join(STAGING, "_qc.wav")
    res = {}
    for attempt in range(4):
        x = base - dc_adj
        if vol < 0:
            x = x * (10 ** (vol / 20))
        sf.write(tmp, x.astype(np.float32), SR)
        af = "highpass=f=30,lowpass=f=18000" if lpf else "highpass=f=30"
        for ext, cmd in [("ogg", ["-c:a", "libvorbis", "-q:a", "4"]),
                         ("m4a", ["-c:a", "aac", "-b:a", "128k"])]:
            out = os.path.join(DST, f"{name}.{ext}")
            subprocess.run([FFMPEG, "-y", "-v", "error", "-i", tmp, "-af", af,
                            "-ar", "44100", *cmd, out], check=True)
            subprocess.run([FFMPEG, "-y", "-v", "error", "-i", out, "-c:a", "pcm_f32le", chk], check=True)
            y, sr = sf.read(chk, always_2d=True)
            tp = 20 * np.log10(np.max(np.abs(y)) + 1e-9)
            dcv = float(np.mean(y.mean(axis=1)))
            rms = 20 * np.log10(np.sqrt(np.mean(y ** 2)) + 1e-9)
            ok_tp, ok_dc = tp <= -3.0, abs(dcv) <= 0.001
            res[ext] = (tp, dcv, rms, y.shape[0] / sr)
            if not (ok_tp and ok_dc):
                if not ok_tp:
                    vol -= (tp + 3.0) + 0.25
                    lpf = True
                if not ok_dc:
                    dc_adj -= dcv
                break
        else:
            break
    else:
        raise AssertionError(f"{name}: 补偿 3 轮后仍超限 {res}")
    return res


def main():
    os.makedirs(STAGING, exist_ok=True)
    if "--emit" in sys.argv:
        emit()
        return
    print(f"| {'id':12s} | {'时长s':>6s} | {'ogg峰dBFS':>9s} | {'m4a峰dBFS':>9s} | {'RMS dBFS':>8s} | 直流 |")
    for i, (name, _d, _p, tgt, max_s) in enumerate(ITEMS):
        raw = os.path.join(STAGING, f"detective_{i}.wav")
        y, sr = sf.read(raw, always_2d=True, dtype="float64")
        assert sr == SR, f"{name}: 采样率 {sr}"
        if y.shape[1] == 1:
            y = np.stack([y[:, 0], y[:, 0]], axis=1)
        sig = prep_oneshot(y, tgt, win=WIN.get(name), max_s=max_s)
        res = encode(name, sig)
        tp_o, dc, rms, dur = res["ogg"]
        print(f"| {name:12s} | {dur:6.2f} | {tp_o:9.2f} | {res['m4a'][0]:9.2f} | {rms:8.2f} | {dc:+.4f} |")
    print("detective SA3 sfx done -> web/detective/audio/sfx/")


if __name__ == "__main__":
    main()