#!/usr/bin/env python3 """生成火箭点火启动倒数(10→1 → Lift off)的英文男声语音,合成成一段约 10 秒的音频。 依赖系统自带的 `say`(macOS)与 `ffmpeg`。每个数字被放在倒数轴上的对应时刻: "10" 在 t=0,"9" 在 t=1s … "1" 在 t=9s,"Lift off" 在 t=10s(即起飞那一刻)。 物理侧 COUNTDOWN_SECONDS=10,正好对上。 用法:python3 scripts/gen_countdown_voice.py 输出:backend/upload/assets/countdown.m4a(随 git 分发)。 """ import subprocess import sys import tempfile from pathlib import Path ASSETS = Path(__file__).resolve().parent.parent / "backend" / "upload" / "assets" OUT = ASSETS / "countdown.m4a" VOICE = "Daniel" # macOS 内置的英文男声 COUNTDOWN_SECONDS = 10 def run(cmd): subprocess.run(cmd, check=True, capture_output=True) def main() -> None: ASSETS.mkdir(parents=True, exist_ok=True) with tempfile.TemporaryDirectory() as tmp: tmp = Path(tmp) # 每句独立合成成 PCM(aiff 是 AIFF-C 压缩,交给 ffmpeg 统一转码更稳) lines = [str(n) for n in range(COUNTDOWN_SECONDS, 0, -1)] + ["Lift off"] clips = [] for i, text in enumerate(lines): aiff = tmp / f"{i:02d}.aiff" run(["say", "-v", VOICE, "-o", str(aiff), text]) clips.append((i, aiff)) # 用 adelay 把每句推到它该说的时刻(第 i 句在 i 秒处),再 amix 叠在一起。 inputs = [] filters = [] for i, aiff in clips: delay = i * 1000 inputs += ["-i", str(aiff)] filters.append(f"[{i}:a]aresample=44100,adelay={delay}|{delay},apad[a{i}]") # normalize=0:各路互不重叠,直接相加即可,别被 amix 按路数摊薄 mix = "".join(f"[a{i}]" for i, _ in clips) + f"amix=inputs={len(clips)}:duration=longest:normalize=0[aout]" filters.append(mix) # 直接 aac 编码成 m4a(浏览器支持一致,体积也小) run(["ffmpeg", "-y", *inputs, "-filter_complex", ";".join(filters), "-map", "[aout]", "-t", str(COUNTDOWN_SECONDS + 2), "-c:a", "aac", "-b:a", "128k", str(OUT)]) print(f"{OUT.name}: {OUT.stat().st_size // 1024} KB") if __name__ == "__main__": try: main() except subprocess.CalledProcessError as err: sys.exit(f"audio 生成失败:\n{err.stderr.decode(errors='replace')}")