MAGIC-TTS Demo
使用同一条默认音色，在四个实用场景下对比 baseline、关键停顿控制、局部内容拉长与 spontaneous 生成效果。
Compare baseline timing, critical pause control, local content stretching, and spontaneous generation under the same default voice.
Controlled vs Spontaneous
Scene 01
前方路口，左转。
目标：动作词需要被稳定听清，单靠自然合成不够。
Goal: the action phrase should remain clearly audible instead of being left entirely to natural timing.
所有内容字统一设为 170 ms。
All content tokens are fixed at 170 ms.
保持内容字 170 ms，把动作前逗号停顿设为 260 ms。
Keep content at 170 ms and set the comma pause before the action phrase to 260 ms.
保留 260 ms 停顿，并把“左 / 转”拉长到 300 ms。
Keep the 260 ms pause and stretch “左 / 转” to 300 ms.
Scene 02
请跟我读，苹果。
目标：教学场景里，目标词音节需要被明确拉出来。
Goal: the target word should be clearly stretched in a guided-reading setting.
所有内容字统一设为 170 ms。
All content tokens are fixed at 170 ms.
目标词前逗号停顿设为 260 ms。
Set the comma pause before the target word to 260 ms.
保留 260 ms 停顿，并把“苹 / 果”拉长到 300 ms。
Keep the 260 ms pause and stretch “苹 / 果” to 300 ms.
Scene 03
验证码是三七九，二一八。
目标：数字串播报必须控制分组和重点数字时长。
Goal: grouped digits and key numbers should be controlled explicitly rather than left to fluency alone.
所有内容字统一设为 170 ms。
All content tokens are fixed at 170 ms.
把 3+3 分组边界处的逗号停顿设为 260 ms。
Set the boundary pause between the two 3-digit groups to 260 ms.
保留 260 ms 停顿，并把六个数字全部拉长到 300 ms。
Keep the 260 ms pause and stretch all six digits to 300 ms.
Scene 04
前方到站，五山站。
目标：站名前缀和站名本体需要分离，站名本体要被突出。
Goal: the station prefix and station name should be separated, and the station name should be emphasized.
所有内容字统一设为 170 ms。
All content tokens are fixed at 170 ms.
站名前逗号停顿设为 260 ms。
Set the comma pause before the station name to 260 ms.
保留 260 ms 停顿，并把“五 / 山 / 站”拉长到 300 ms。
Keep the 260 ms pause and stretch “五 / 山 / 站” to 300 ms.
Spontaneous Duration Modeling
不提供任何 target-side 显式时长，模型自行建模节奏。
No explicit target-side duration is provided; the model predicts timing on its own.
不提供任何 target-side 显式时长，模型自行建模节奏。
No explicit target-side duration is provided; the model predicts timing on its own.
不提供任何 target-side 显式时长，模型自行建模节奏。
No explicit target-side duration is provided; the model predicts timing on its own.
不提供任何 target-side 显式时长，模型自行建模节奏。
No explicit target-side duration is provided; the model predicts timing on its own.