xsl/SoundGame
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(assets): SA3 音效补齐 radio 11+1 与 dungeon 13源+15pan（token 走 hf-mirror 实测可用）

Browse Source 
- 模型 17/17 文件经镜像全量缓存（gated+token，直连本机不通；sfx.md/setup.md/asset-pipeline.md 镜像结论同步更新）
- 修复迁移遗留：stable_audio_3/models/ 源码子包被 .gitignore models/ 规则误伤丢失，从上游恢复并加白名单
- radio：arg_muffled/brake_screech 等叙事件落地（接线原有 L111/L154 零改动）+ phone_ring SYN→SA3 升级；dungeon 13 源 A/B 对齐（RMS 偏差 ≤0.3dB）全量同名替换，SYN 原版备份 tools/asset_fix_backup/sa3_ab_20260902/
- 关键玩法语义达标：echo 距离档 0.83<1.43<2.18<2.98s 严格递增；one-shot 真峰值全 ≤-3dBFS（解码回读法）
- check_all_assets 8 游戏全绿（radio 254/254、dungeon 303/303）；playtest radio 221/dungeon 1610 全绿
This commit is contained in:
xsl
2026-09-02 19:35:38 +08:00
parent d870a75a26
commit 24c526af68
181 changed files with 5324 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
web/dungeon/gen/gen_sfx2.py
+2
View File
@@ -5,6 +5,8 @@
SA3 无 HF token 跳过（research/new8/batch1-assets-report.md §4 明示：
「dungeon 回声/脚步是核心动词，开发联调前建议先补产…或临时用 SYN 兜底」），本脚本即该兜底。
若日后补产 SA3 版，直接覆盖同名文件即可（id 不变、游戏层零改动）。
※ 2026-09-02 更新：SA3 版已由 gen_sfx_sa3.py 补产并 A-B 后同名替换本脚本产出（28 文件；
本脚本重跑即回退 SYN 兜底版，原版备份 tools/asset_fix_backup/sa3_ab_20260902/dungeon/）。
25 个 id / 25×2 文件（mob_step_*/tel_* 的 l/r 为 mono 源经 ffmpeg pan 派生预 pan 立体声）：
echo_1..4 正前回声 4 距离档：贴身亮短 → 远处暗长（stone 反射；门/水材质现有生成器
web/dungeon/gen/gen_sfx_sa3.py
+210
View File
@@ -0,0 +1,210 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""《回声地窟》Stable Audio 3 small-sfx 音效后处理（SA3 补产批次，16-dungeon.md §9.2）。
批次 2 的 25 个 id 此前为 SYN 兜底（gen_sfx2.py，§9.2 降级条款）；本脚本产 SA3 候选版，
与在役 SYN 版做 A-B（响度/时长/包络对齐），质量明显更好才同名晋级 audio/sfx/（游戏层零改动）。
prompt 出自 §9.2 表（echo 距离档/tel 前摇的英文稿按中文说明补全）。
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 13 条 mono 源原始 wav 到 gen/staging/sfx_sa3/dungeon_<i>.wav（顺序 = 下表 idx）。
后处理：修整/去直流/8ms cos² 淡化/源真峰值 -4dBFS（同在役 SYN 档）→
mob_step_*/tel_* 派生预 pan 立体声（l=能量只在 L / r=只在 R / c=双声道同相复制，
工艺同 gen_sfx2.py pan_derive/stereo_dup）→
ogg libvorbis q4 + m4a aac 128k（highpass=f=30，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 轮补偿。候选落 gen/staging/sfx_sa3/out/（不入库），A-B 后晋级。
用法：.venv-tts/bin/python web/dungeon/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")
OUT = os.path.join(STAGING, "out")
DST = os.path.join(HERE, "..", "audio", "sfx")
SA3 = "/home/xsl/SoundGame/tools/stable-audio-3"
FFMPEG = "/home/xsl/SoundGame/tools/bin/ffmpeg"
SEED = 20260902
# (id, gen 秒, prompt) —— §9.2 标注 SA3 的 13 个 mono 源；定档全部 -4dBFS（同在役 SYN）
ITEMS = [
("echo_1", 2.5, "close stone wall echo slap-back, bright, short, dry"),
("echo_2", 3.0, "mid distance cave echo, moderate tail"),
("echo_3", 3.5, "far cave echo, darker longer"),
("echo_4", 4.5, "very far cavern echo, long dark tail, huge space"),
("stairs_down", 4.0, "descending old metal stairs, deep cavern reverb"),
("atk_hit", 2.0, "blunt impact on hard insect chitin shell, dull thud hit"),
("enemy_die", 2.5, "monster body collapsing on stone floor, heavy fall thud and debris"),
("mob_step_small", 2.5, "small beast skittering steps, dry"),
("mob_step_bug", 2.5, "insect chitin crawling clicks"),
("mob_step_heavy", 2.5, "heavy armored thudding steps"),
("tel_small", 2.5, "small beast pounce hiss, short aggressive screech"),
("tel_bug", 2.5, "insect legs chitin rubbing scrape, creepy skitter"),
("tel_heavy", 2.5, "huge monster stomp and low growl, heavy rumble"),
]
# pan 派生表（与在役文件集一致：mob_step 有 _c，tel 无 _c）
DERIVE = {}
for _n in ["mob_step_small", "mob_step_bug", "mob_step_heavy"]:
DERIVE[_n] = ["_l", "_c", "_r"]
for _n in ["tel_small", "tel_bug", "tel_heavy"]:
DERIVE[_n] = ["_l", "_r"]
# 事件窗长度（秒）：raw 10.59s 定长 latent 床上按自响起点截取；None=按 -50dBFS 活动窗
# 自然修剪。echo 档长 0.65<1.25<2.0<2.8 保持与在役 SYN 档（0.58/1.08/1.68/2.38）同序递增
# ——回声尾长=墙距是本作核心信息，A-B 不满足递增即不晋级。
WIN = {
"echo_1": 0.65, "echo_2": 1.25, "echo_3": 2.0, "echo_4": 2.8,
"stairs_down": None, "atk_hit": None, "enemy_die": 1.7,
"mob_step_small": 0.45, "mob_step_bug": 0.55, "mob_step_heavy": 0.65,
"tel_small": 1.0, "tel_bug": 1.1, "tel_heavy": 1.3,
}
# 响度对齐：候选按在役 SYN 版（同名 ogg）RMS 定档（-4dBFS 峰值帽不变），保游戏混音参数
SYN_DIR = os.path.join(HERE, "..", "audio", "sfx")
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}/dungeon.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(y, target_db=-4.0, win_len=None, syn_rms_db=None):
if y.ndim == 1:
y = y[:, None]
if win_len is not None:
# 自响起点：10ms 窗包络首个超过 (峰值-20dB) 的位置，提前 20ms 起
w = int(SR * 0.01)
nfr = len(y) // w
eb = 20 * np.log10(np.max(np.abs(y[:nfr * w, 0]).reshape(nfr, w), axis=1) + 1e-9)
pk = 20 * np.log10(np.max(np.abs(y[:, 0])) + 1e-9)
on = np.where(eb > pk - 20)[0][0] * w
a = max(0, on - int(SR * 0.02))
b = min(len(y), a + int(SR * win_len))
y = y[a:b]
nf2 = max(1, int(SR * 0.050)) # 窗切在持续声中，切点 50ms 缓出
y[-nf2:] *= (0.5 * (1 + np.cos(np.pi * np.arange(nf2) / nf2)))[:, None]
else:
y = trim_oneshot(y)
y = y - np.mean(y, axis=0, keepdims=True) # 去直流
nf = max(1, int(SR * 0.008))
fade = 0.5 * (1 - np.cos(np.pi * np.arange(nf) / nf))
y[:nf] *= fade[:, None] # 8ms cos² 淡入淡出
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])) # 编码器软着陆（同 gen_sfx2）
pad_t = np.zeros((int(SR * 0.12), y.shape[1]))
y = np.concatenate([pad_h, y, pad_t])
if syn_rms_db is not None: # A-B 响度对齐（峰值帽 -4 不变）。
rms = 20 * np.log10(np.sqrt(np.mean(y ** 2)) + 1e-9) # 垫后计算：与解码回读同口径
s = min(10 ** ((syn_rms_db - rms) / 20), 1.0)
y = y * s
return y
def _syn_metrics(name):
"""在役 SYN 版 ogg 解码回读（时长, RMS）——响度对齐基准。"""
p = os.path.join(SYN_DIR, name + ".ogg")
subprocess.run([FFMPEG, "-y", "-v", "error", "-i", p, "-c:a", "pcm_f32le", "/tmp/_syn_qc.wav"], check=True)
y, sr = sf.read("/tmp/_syn_qc.wav", always_2d=True)
return y.shape[0] / sr, 20 * np.log10(np.sqrt(np.mean(y ** 2)) + 1e-9)
def derive(name, mono):
"""mono -> 预 pan 立体声（能量只在 L 或 R；c = 双声道同相复制），工艺同 gen_sfx2。"""
outs = []
for suf in DERIVE.get(name, []):
if suf == "_l":
outs.append((name + suf, np.stack([mono[:, 0], np.zeros(len(mono))], axis=1)))
elif suf == "_r":
outs.append((name + suf, np.stack([np.zeros(len(mono)), mono[:, 0]], axis=1)))
elif suf == "_c":
outs.append((name + suf, np.stack([mono[:, 0], mono[:, 0]], axis=1)))
return outs
def encode(name, sig):
base = sig.astype(np.float64)
vol, lpf, dc_adj = 0.0, False, 0.0
tmp = os.path.join(OUT, "_enc.wav")
chk = os.path.join(OUT, "_qc.wav")
res = {}
for attempt in range(4): # 首发 + 最多 3 轮补偿（首轮重编加 18kHz 低通）
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(OUT, 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 # 首轮重编起加 18kHz 低通压 MDCT 过冲
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)
os.makedirs(OUT, exist_ok=True)
if "--emit" in sys.argv:
emit()
return
print(f"| {'id':18s} | {'时长s':>6s} | {'ogg峰dBFS':>9s} | {'m4a峰dBFS':>9s} | {'RMS dBFS':>8s} | {'SYN时长':>7s} | {'SYN RMS':>8s} | 直流 | 声道 |")
for i, (name, _d, _p) in enumerate(ITEMS):
raw = os.path.join(STAGING, f"dungeon_{i}.wav")
y, sr = sf.read(raw, always_2d=True, dtype="float64")
assert sr == SR, f"{name}: 采样率 {sr}"
mono = y[:, :1]
syn_dur, syn_rms_db = _syn_metrics(name)
# 响度对齐只做一次（按 mono 基准），派生 l/r/c 直接复制基准样本——
# 与在役 SYN 派生工艺一致：l/r 文件 RMS = 基准 -3dB（能量单声道承载），逐声道振幅不变
base_sig = prep(mono, win_len=WIN.get(name), syn_rms_db=syn_rms_db)
for eid, sig in [(name, base_sig)] + derive(name, base_sig):
res = encode(eid, sig)
tp_o, dc, rms, dur = res["ogg"]
tp_m = res["m4a"][0]
ch = 1 if eid == name else 2
sd = f"{syn_dur:6.2f}" if eid == name else " -"
print(f"| {eid:18s} | {dur:6.2f} | {tp_o:9.2f} | {tp_m:9.2f} | {rms:8.2f} | {sd} | {syn_rms_db:7.2f} | {dc:+.4f} | {ch:d} |")
print("dungeon SA3 候选完成 -> gen/staging/sfx_sa3/out/（A-B 复核后晋级）")
if __name__ == "__main__":
main()