The first time an AI voice sounded like me, I was excited. Then I listened for a little longer. Some versions sounded forced, as if I were holding my voice down. Others slipped back into a robotic sound. Getting the tone of my voice right was only part of the job.
Start with one awkward sentence
A different local voice-generation setup brought the voice closer to mine. Then we worked on how it spoke.
One sentence kept exposing the problem: a voice clip can sound good on its own and still feel awkward in a full video. A pause in the wrong place interrupted the thought. We regrouped the narration around its meaning and adjusted the gaps between passages.
Listen to the whole edit
We generated the narration in short passages, then put it back into the video. With captions and visuals, a clip that sounded fine alone could still drag or stumble. I listened to the complete edit before choosing the final version.
After approving the Chinese video, I tried an English version. The early sample sounded like me, but the delivery was not fluent enough. We tested short passages, adjusted the delivery, and then made the full narration. I approved the result after listening. That is an outcome from this experiment, not a promise that the same setup will work equally well for everyone.
What I would do again
Test a difficult sentence before generating the whole script. Split the narration by meaning. Listen again after editing it into the video. The tool generates the audio; I decide whether it is a voice I want to publish.
The narration is an AI-generated version of my own voice, used with my permission. This note shares the process, not the original recordings, voice model or speaker data.