#51 · open · 2 comments
I'm seeing the timestamp drift by i.e. almost by 10s. See following [audio.mp3](https://github.com/user-attachments/files/31648756/audio_example.mp3): > Así es, así es, también músico, estaban yendo a ese [...] That happens in the audio at 02:05 (mm:ss), while the model predicts at `01:54` the start of that phrase: <img width="1098" height="126" alt="Image" src="https://github.com/user-attachments/assets/2ad1be83-2d80-4c2b-8a6b-139b378aee80" />
Me too. When the audio doesn't start immediately, there's a "silence" segment before it, and the first few transcriptions will be shifted by 10 seconds. 00:00 ~ 05:39：No output, correct. (Music playing and no one speaking.) 05:39 ~ 07:30：Transcriptions be 10 seconds slower. As transcription progresses, the 07:30 mark is the dividing line. 7:30 ~ 23:36：From this point onwards, the model output is very accurate until the end. Audio from podcast： [ VTuber學｜關於我推的學問 ft. 貓頭鷹出版社 王正緯](https://www.youtube.com/live/N9boWvU-KkA)
got it, we will fix this issue.