如何为 Seed Audio 写提示词

最重要的规则

1. 1

   先设定场景，再写台词

   从地点和它的声音写起：下雨的站台、喧闹的体育场、安静的录音棚。只有引号里的文字会被说出来，其余都是指示说明。

2. 2

   每个声音只描述一次，并描述完整

   角色第一次开口时,写清他的性别、年龄、口音、音色、情绪和语速:Leo(四十多岁的男人,声音低沉沙哑,疲惫,语速缓慢)。之后只需写名字和一句简短的说话方式提示即可。

3. 3

   按所选时长来写

   先选好时长。英语语速约为每秒 2.5 个单词,所以短片段只能放下一句短台词。字数太多,单词会被截断;字数太少,结尾会留下空白。

4. 4

   提示词要用台词所用的语言来写

   西班牙语场景就用西班牙语提示词,从场景描述到对白都是如此。英文提示词搭配其他语言的台词,会让演绎效果变得混乱。

5. 5

   多生成几个版本

   即使种子相同,每次生成也都是一次全新的演绎。同一句台词生成两三次,挑最好的留下。

可参考的提示词

• 环境音中的旁白

  A quiet forest at dawn, birdsong slowly building, a stream running somewhere close. A narrator (a man in his sixties, standard British accent, warm and slightly gravelly, unhurried) says softly: "Every spring, the river remembers its way back to the sea." The birdsong continues to the end.

  为什么有效: 先写场景,声音用选角说明来描述,鸟鸣持续到最后,让片段在声音中结束,而不是戛然而止。时长选 10 秒正合适。

• 两个声音,精确到秒

  A small radio studio late at night, the hum of equipment, rain against the window. Theo (a man in his late twenties, bright, quick, a little nervous) says: "[0.3s:3.2s] Okay, we're live. Say hi to everyone." June (a woman in her fifties, low, smoky, completely calm, slow) answers: "[3.8s:8.6s] Good evening, night owls. Pour something warm and stay a while." A soft jingle chimes; the rain continues to the end.

  为什么有效: 每句台词开头都有自己的时间窗口,长度与字数相称,说话人之间留有短暂的间隔,结尾还有安静的一秒。请把时长设为 10 秒。

• 不带人声的背景音乐

  Generate pure background music and sound effects only. No voice, no dialogue, no narration, no lyrics.
  Style: warm, hopeful acoustic folk for a travel video. Create original music only.
  Background music: fingerpicked acoustic guitar, soft hand claps, a gentle upright piano, a light shaker.
  Sound effects: a car door closing, gravel under tires, wind on an open road. Keep effects clear but balanced under the music.
  Duration: about 30 seconds.
  Structure:
  [0:00-0:08] Guitar alone, light and curious; a car door closes.
  [0:08-0:20] Piano and shaker join, gravel under the tires, the tempo lifts.
  [0:20-0:27] The full band with hand claps, wind rushing past.
  [0:27-0:30] The music settles on one ringing guitar chord.
  Final mix: No vocals or spoken words. Bright and open, ready to sit under a travel montage.

  为什么有效: 在开头和最后的混音说明里都写上“无人声”,就能保证是纯器乐,各段落加起来正好是设置中的 30 秒。

• 只有音效,没有音乐

  Inside a blacksmith's forge at night. The bellows wheeze and the coals roar up bright; a hammer rings on hot iron in a steady rhythm, three strikes and a pause, then the blade hisses as it plunges into a bucket of water and steam bursts up. Rain drums on the tin roof throughout. No music, no voices.

  为什么有效: 每个声音都有明确的声源和出现的时间点,最后一句则排除掉你不想要的内容。

• 保存给 Character 用的干净人声

  Clean, close-miked studio voice recording. No music, no background sound, no effects. Mara (a woman in her early thirties, standard American accent, low, warm, slightly husky, calm and assured) says at an even pace: "[0.2s:8.8s] I grew up on this coast. The tide always tells you something, if you are patient enough to listen."

  为什么有效: 只有一个说话人,没有背景声,从第一刻起就有台词:这是最适合保存为声音并赋予 Character 的片段。请把时长设为 10 秒。

• 由图片生成声音

  The old fisherman in the picture says warmly, a little amused, unhurried: "Well, look who finally showed up."

  为什么有效: 上传一张肖像作为图片,Seed Audio 会设计一个与面孔相配的声音,所以提示词里只需写说话方式和台词。5 秒时长适合一句短台词。