Chinese voice to animated text: inspect the recognition before export

Source and method

An original Mandarin script was spoken by eSpeak NG, then processed with the same local Whisper small recognizer and video renderer used by this website. This is synthesized speech, not a recording of a person and not a singing benchmark.

These case studies use original scripts, synthetic speech, generated tones and a generated notice image. They contain no user uploads. Source media and recognized text keep their original language when the interface language changes.

Download original input ↓ · Download settings ↓ · Download source text ↓

Compare the source, recognition and review

Source text

今天我们用一段中文录音制作文字动画。
先检查识别结果。
再调整字幕的出现时间。
最后下载视频，分享你的故事。

Unedited recognition result

今天我们用一般中文录音制作文
字动画,
先天查识别结果,
再调整字幕的出现时间,
最后下彩视频,
分享你的故事。

Reviewed result

今天我们用一段中文录音制作文字动画。
先检查识别结果。
再调整字幕的出现时间。
最后下载视频，分享你的故事。

The reviewed captions restore the original script and use evenly timed phrases across the audio. Their timing is approximate and is not forced alignment. Both the unedited model output and the reviewed export are available; compare them with the sound.

Download video with unedited captions ↓ · Download reviewed video ↓

What to check before sharing

In this run, the source 一段 was recognized as 一般, 检查 as 天查, and 下载 as 下彩. Review the Chinese words against the recording before adjusting phrase boundaries. These are results from this synthetic sample, not an error rate for Chinese speech.

Measurement and limitations

Times describe a single build run, not a service-level promise. Audio cases are measured on the build computer; image OCR is measured in the Linux deployment image. Upload time and the shared processing queue are excluded.

Recognition time

6.41 s

Render time

2.16 s

Processing engine

faster-whisper small / CPU int8

This is a product demonstration, not an accuracy benchmark or customer testimonial.