SoundGame/tools/gen_cipher_syn.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""《暗码》程序合成音效（19-cipher.md §6.2，12 件 SYN 全部程序合成兜底）。

落盘：staging wav → tools/gen_cipher_syn_wav/<id>.wav（44.1kHz 立体声）；
      成品     → web/cipher/audio/sfx/<id>.{ogg,m4a}（tools/bin/ffmpeg 双格式）。

one-shot QC（10 件）：去直流 + 首尾 8ms cos² 淡化 + 编码后解码回读真峰值 ≤ -3dBFS
（MDCT 过冲坑：以 ebur128 peak=true 实测为准，超限自动降档重编）。
loop QC（2 件：static_bed 4s / amb_room 8s）：只去直流（首尾禁淡化），事件按
循环周期网格化保证循环点相位连续，另做真峰值校验只报告不强制；无缝循环由
游戏层 buffer.loop=true 承担（19-cipher.md §3.4，radio _ringSrc 先例）。

19-cipher.md §6.2 原标 SA3 的 10 件因「文件方案在 70-160ms 短音上 ms 级抖动」
同理由与管线一致性裁决改 SYN（§修订记录 R1：SFX 全 SYN 兜底，prompt 转设计
参考）；tune_seek/key_send 等节奏型音效时长以真实事件为准（§6.2 标注时长）。

用法：.venv-tts/bin/python tools/gen_cipher_syn.py
"""
import os
import subprocess
import sys

import numpy as np
import soundfile as sf

ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
FFMPEG = os.path.join(ROOT, "tools", "bin", "ffmpeg")
OUT = os.path.join(ROOT, "tools", "gen_cipher_syn_wav")
DST = os.path.join(ROOT, "web", "cipher", "audio", "sfx")
os.makedirs(OUT, exist_ok=True)
os.makedirs(DST, exist_ok=True)

sr = 44100
ONE_SHOT_TP_CEIL = -3.0   # dBFS
SAMPLE_PEAK = 0.66        # ≈ -3.6dBFS，给 MDCT 过冲留余量
FADE_N = int(sr * 0.008)

_rng = np.random.default_rng(1940)


def dc(y):
    return y - np.mean(y)


def fade(y):
    w = np.sin(np.linspace(0, np.pi / 2, FADE_N)) ** 2
    y = y.copy()
    y[:FADE_N] *= w
    y[-FADE_N:] *= w[::-1]
    return y


def norm(y, peak=SAMPLE_PEAK):
    m = np.max(np.abs(y))
    return (y / m * peak) if m > 0 else y


def stereo(y):
    return np.stack([y, y], axis=1).astype(np.float32)


def place(y, start_s, sig):
    s = int(start_s * sr)
    n = min(len(sig), len(y) - s)
    if n > 0:
        y[s:s + n] += sig[:n]


def bp_noise(n, lo, hi):
    x = _rng.standard_normal(n)
    X = np.fft.rfft(x)
    freqs = np.fft.rfftfreq(n, 1 / sr)
    mask = ((freqs >= lo) & (freqs <= hi)).astype(float)
    y = np.fft.irfft(X * mask, n)
    return y / (np.max(np.abs(y)) + 1e-9)


def bell(n, f, partials=((1, 1.0), (2.4, 0.4), (3.9, 0.18), (5.4, 0.08)), dec=6.0):
    t = np.arange(n) / sr
    y = np.zeros(n)
    for r, g in partials:
        y += g * np.sin(2 * np.pi * f * r * t) * np.exp(-t * (dec / r))
    return y


def footstep(n_off, heavy=1.0):
    """皮靴脚步：55-90Hz 低频冲击 + 皮革摩擦带通噪声。"""
    tt = np.arange(n_off) / sr
    th = np.sin(2 * np.pi * (62 + _rng.uniform(-6, 6)) * tt) * np.exp(-tt * 22) * heavy
    th += bp_noise(n_off, 150, 700) * np.exp(-tt * 30) * 0.5 * heavy
    return th * (1 - np.exp(-tt * 300))


# ---------------------------------------------------------------- one-shot 渲染
def r_tune_seek():  # 调谐找频 3s：扫频静电 + 两处鬼影摩尔斯 + 落定静音点
    n = int(sr * 3.0)
    t = np.arange(n) / sr
    y = bp_noise(n, 300, 4000) * 0.28
    # 谐振峰在频段内来回扫（电台调谐手感）
    f0 = 900 + 1900 * np.sin(2 * np.pi * 0.55 * t) ** 3
    ph = 2 * np.pi * np.cumsum(f0) / sr
    y += (np.sin(ph) + 0.4 * np.sin(2 * ph)) * 0.10
    # 两处鬼影滴答（微弱 750Hz dit/dah，§2.1 电文音色同族，弱化远台）
    for st, pats in [(0.9, [1, 3, 1]), (2.0, [1, 1, 3])]:
        for i, u in enumerate(pats):
            L = int(sr * 0.08 * u)
            tt = np.arange(L) / sr
            mk = np.sin(2 * np.pi * 750 * tt) * np.exp(-tt * 6) * 0.16
            place(y, st + i * 0.16, mk)
    env = np.minimum(1, t / 0.15) * np.minimum(1, (3.0 - t) / 0.3)
    return dc(y * env)


def r_paper_page():  # 密码本翻页 0.5s：干纸摩擦（3-8kHz 噪声抖动）+ 指尖拍页轻响
    n = int(sr * 0.50)
    y = bp_noise(n, 2600, 9000) * 0.5
    t = np.arange(n) / sr
    flut = 0.5 + 0.5 * np.sin(2 * np.pi * 13 * t) * np.sin(2 * np.pi * 2.1 * t)
    env = (1 - np.exp(-t * 60)) * np.exp(-np.maximum(t - 0.28, 0) * 16)
    y *= flut * env
    tap_n = int(sr * 0.05)
    tt = np.arange(tap_n) / sr
    tap = bp_noise(tap_n, 900, 2600) * np.exp(-tt * 90) * 0.8
    place(y, 0.30, tap)
    return dc(y)


def r_key_send():  # 电键拍发 4s：快手七手法——不均匀节奏的黄铜键击串（句读有呼吸）
    n = int(sr * 4.0)
    y = np.zeros(n)
    # (起秒, 击长单位 u=80ms 倍数) —— 有连发有停顿的「手法」节奏
    seq = [(0.20, 1), (0.36, 3), (0.68, 1), (0.84, 1), (1.16, 3), (1.48, 1),
           (1.80, 1), (2.12, 3), (2.44, 1), (2.60, 1), (2.92, 3), (3.24, 1), (3.40, 3)]
    for st, u in seq:
        L = int(sr * 0.08 * u)
        tt = np.arange(L) / sr
        click = bp_noise(L, 1400, 5200) * np.exp(-tt * 110)      # 触点脆响
        body = np.sin(2 * np.pi * 2100 * tt) * np.exp(-tt * 40) * 0.35  # 簧片余振
        th = np.sin(2 * np.pi * 180 * tt) * np.exp(-tt * 80) * 0.25     # 底座轻顿
        place(y, st, (click + body + th) * 0.8)
    return dc(y)


def r_pen_scratch():  # 处长记事笔 1.5s：4 段 2.6kHz 带通沙写，顿笔收尾
    n = int(sr * 1.50)
    y = np.zeros(n)
    for i, st in enumerate([0.05, 0.38, 0.72, 1.06]):
        L = int(sr * (0.20 + 0.05 * (i % 2)))
        tt = np.arange(L) / sr
        scr = bp_noise(L, 2200, 5600) * (0.55 + 0.45 * np.sin(2 * np.pi * 7 * tt))
        scr *= (1 - np.exp(-tt * 250)) * np.exp(-np.maximum(tt - 0.14, 0) * 30)
        place(y, st, scr * 0.5)
    return dc(y)


def r_tea_cup():  # 茶杯落桌 0.5s：瓷碰 2.8/3.7/4.6kHz 衰减 + 木桌低频轻顿
    n = int(sr * 0.50)
    y = np.zeros(n)
    clink = bell(n, 2820, partials=((1, 1.0), (1.32, 0.5), (1.63, 0.3), (2.1, 0.12)), dec=28)
    tt = np.arange(n) / sr
    th = np.sin(2 * np.pi * 130 * tt) * np.exp(-tt * 45) * 0.5
    return dc((clink * 0.5 + th) * (1 - np.exp(-tt * 900)))


def r_step_halt():  # 门外脚步停住 3s（T4 danger 拍 L2 前置 2s 渐强）：4 步逼近→第 4 步后死寂
    n = int(sr * 3.0)
    y = np.zeros(n)
    t = np.arange(n) / sr
    for i, st in enumerate([0.10, 0.78, 1.46, 2.14]):
        L = int(sr * 0.22)
        y[int(st * sr):int(st * sr) + L] += footstep(L, heavy=0.75 + 0.22 * i)
    # 渐强窗：0→2s 缓慢抬头（L2 预警），脚步落停后房间音沉下去
    env = np.minimum(1, 0.35 + 0.65 * np.clip((t - 0.4) / 1.8, 0, 1))
    tail = bp_noise(n, 90, 400) * 0.05 * np.exp(-np.maximum(t - 2.3, 0) * 8)
    return dc((y + tail) * env)


def r_lock_click():  # 门锁/枪机 1s：锁舌过槽两响 + 低频金属余振
    n = int(sr * 1.0)
    y = np.zeros(n)
    for st, f, dec in [(0.06, 3400, 60), (0.22, 2400, 45)]:
        L = int(sr * 0.16)
        tt = np.arange(L) / sr
        c = bp_noise(L, 1200, 5600) * np.exp(-tt * dec)
        c += np.sin(2 * np.pi * f * tt) * np.exp(-tt * dec * 0.5) * 0.4
        place(y, st, c)
    tt = np.arange(n) / sr
    y += np.sin(2 * np.pi * 95 * tt) * np.exp(-tt * 12) * 0.3
    return dc(y)


def r_dawn_bell():  # 黎明远钟 4s：低音钟（E3 系）远播 + 雨底残响；e1/e2 结算锚点
    n = int(sr * 4.0)
    y = np.zeros(n)
    for st in [0.0, 1.9]:
        L = n - int(st * sr)
        tt = np.arange(L) / sr
        b = bell(L, 165, partials=((1, 1.0), (2.02, 0.42), (2.94, 0.22), (4.1, 0.08)), dec=3.2)
        place(y, st, b * 0.55)
    t = np.arange(n) / sr
    rain = bp_noise(n, 900, 6000) * 0.03 * (1 + 0.3 * np.sin(2 * np.pi * 0.3 * t))
    return dc(y + rain * np.minimum(1, (4.0 - t) / 0.8))


def r_ui_confirm():  # 全局先例同款（storyline/radio/abyss gen_sfx.py：880->1320Hz 双音）
    a = np.sin(2 * np.pi * 880 * np.arange(int(sr * 0.09)) / sr) * np.exp(-np.arange(int(sr * 0.09)) / sr * 60)
    b = np.sin(2 * np.pi * 1320 * np.arange(int(sr * 0.14)) / sr) * np.exp(-np.arange(int(sr * 0.14)) / sr * 40)
    return dc(np.concatenate([a, np.zeros(int(sr * 0.02)), b, np.zeros(int(sr * 0.05))]))


def r_ui_menu():  # 全局先例同款（330+495Hz 双谐波 0.35s）
    d = 0.35
    x = np.arange(int(sr * d)) / sr
    return dc(np.concatenate([
        (np.sin(2 * np.pi * 330 * x) + 0.5 * np.sin(2 * np.pi * 495 * x)) * np.exp(-x * 20),
        np.zeros(int(sr * 0.05))]))


ONE_SHOTS = {
    "tune_seek": r_tune_seek,
    "paper_page": r_paper_page,
    "key_send": r_key_send,
    "pen_scratch": r_pen_scratch,
    "tea_cup": r_tea_cup,
    "step_halt": r_step_halt,
    "lock_click": r_lock_click,
    "dawn_bell": r_dawn_bell,
    "ui_confirm": r_ui_confirm,
    "ui_menu": r_ui_menu,
}


# ---------------------------------------------------------------- loop 渲染
def circular_smooth(y, k_s=0.05):
    """首尾 k 秒等功率交叉淡化（噪声类 loop 循环点去咔哒；相位类成分已网格化）。"""
    k = int(sr * k_s)
    w = np.sin(np.linspace(0, np.pi / 2, k)) ** 2
    out = y[:-k].copy()
    out[:k] = out[:k] * w + y[-k:] * (1 - w)
    return out


def r_static_bed():  # 短波静电底噪 4s 无缝：碎裂噪声 + 电源哼 60Hz 整周期 + 随机噼啪网格化
    n = int(sr * 4.0)
    y = bp_noise(n, 250, 5200) * 0.13
    t = np.arange(n) / sr
    y += np.sin(2 * np.pi * 60 * t) * 0.012          # 4s = 240 整周期，相位对齐
    y += np.sin(2 * np.pi * 120 * t) * 0.006
    # 噼啪：16 个落点整除网格，尾 30ms 不出界（n=4s 网格无需回卷）
    for i in range(16):
        st = i * 0.25 + _rng.uniform(0, 0.14)
        L = int(sr * 0.03)
        tt = np.arange(L) / sr
        pk = bp_noise(L, 1800, 8000) * np.exp(-tt * 130) * _rng.uniform(0.1, 0.3)
        place(y, st, pk)
    return dc(circular_smooth(dc(y)))


def r_amb_room():  # 值班室夜境 8s 无缝：挂钟 8  tick（整 1Hz）+ 雨窗噪声 + 暖炉低哼
    n = int(sr * 8.0)
    y = np.zeros(n)
    for i in range(8):  # 每秒一记，尾 60ms 不出界
        L = int(sr * 0.06)
        tt = np.arange(L) / sr
        tick = bp_noise(L, 2000, 7000) * np.exp(-tt * 160) * 0.5
        tick += np.sin(2 * np.pi * 4200 * tt) * np.exp(-tt * 200) * 0.25
        place(y, i * 1.0, tick)
    t = np.arange(n) / sr
    rain = bp_noise(n, 700, 4500) * 0.05 * (0.6 + 0.4 * np.sin(2 * np.pi * 0.25 * t) ** 2)  # 0.25Hz 整 2 周期
    hum = np.sin(2 * np.pi * 100 * t) * 0.012 + np.sin(2 * np.pi * 150 * t) * 0.006  # 整周期
    return dc(circular_smooth(dc(y + rain + hum)))


LOOPS = {
    "static_bed": r_static_bed,
    "amb_room": r_amb_room,
}


# ---------------------------------------------------------------- 编码 + 真峰值校验
def encode(wav_path, base):
    for args, ext in [(["-c:a", "libvorbis", "-q:a", "3"], ".ogg"),
                      (["-c:a", "aac", "-b:a", "128k"], ".m4a")]:
        subprocess.run([FFMPEG, "-y", "-v", "error", "-i", wav_path,
                        "-ar", "44100", *args, base + ext], check=True)


def true_peak_db(path):
    r = subprocess.run([FFMPEG, "-i", path, "-af", "ebur128=peak=true", "-f", "null", "-"],
                       capture_output=True, text=True)
    peaks = [float(line.split("Peak:")[1].strip().split()[0])
             for line in r.stderr.splitlines() if "Peak:" in line and "dBFS" in line]
    return max(peaks) if peaks else float("nan")


def check_tp(base):
    return max(true_peak_db(base + ".ogg"), true_peak_db(base + ".m4a"))


def main():
    report = []
    problems = []

    for fid, fn in ONE_SHOTS.items():
        wav_path = os.path.join(OUT, fid + ".wav")
        base = os.path.join(DST, fid)
        y = fade(norm(dc(fn())))
        sf.write(wav_path, stereo(y), sr)
        encode(wav_path, base)
        tp = check_tp(base)
        if tp > ONE_SHOT_TP_CEIL:  # MDCT 过冲超限：按超出量降档重编一次再验
            gain = 10 ** ((ONE_SHOT_TP_CEIL - tp) / 20) * 0.98
            sf.write(wav_path, stereo(y * gain), sr)
            encode(wav_path, base)
            tp2 = check_tp(base)
            report.append((fid, len(y) / sr, tp, f"降 {20 * np.log10(gain):.1f}dB 重编 -> {tp2:.2f}"))
            tp = tp2
        else:
            report.append((fid, len(y) / sr, tp, ""))
        if tp > ONE_SHOT_TP_CEIL:
            problems.append(f"{fid}: 真峰值 {tp:.2f} dBFS 仍超 {ONE_SHOT_TP_CEIL}")

    for fid, fn in LOOPS.items():
        wav_path = os.path.join(OUT, fid + ".wav")
        base = os.path.join(DST, fid)
        y = norm(fn())  # loop：去直流已在渲染内；无首尾淡化（无缝优先）
        sf.write(wav_path, stereo(y), sr)
        seam = abs(float(y[0]) - float(y[-1]))
        encode(wav_path, base)
        tp = check_tp(base)
        report.append((fid, len(y) / sr, tp, f"loop seam |y0-y1|={seam:.4f}"))
        if tp > ONE_SHOT_TP_CEIL:
            problems.append(f"{fid}: loop 真峰值 {tp:.2f} dBFS 超 {ONE_SHOT_TP_CEIL}（建议降档）")

    print(f"{'id':14s} {'dur(s)':>7s} {'TP(dBFS)':>9s}  note")
    for fid, dur, tp, note in report:
        print(f"{fid:14s} {dur:7.2f} {tp:9.2f}  {note}")
    if problems:
        print("PROBLEMS:")
        for p in problems:
            print(" ", p)
        return 1
    print(f"ALL OK: {len(ONE_SHOTS)} one-shot + {len(LOOPS)} loop -> {DST}")
    return 0


if __name__ == "__main__":
    sys.exit(main())