Even when the speakers starts talking after 10 sec, Whisper make the first timestamp to start at sec 0. How could I change that?

[romain130492]

Hello

I'm using Whisper,
when having a video with a speaker starting his speech at sec 10, I'm getting the first timestamp to be at sec 1. instead of sec 10.
Here is my config:

Config
POST v1/audio/transcriptions

{ 
 model:"whisper-1"
 file:"...mp3"
 response_format:"srt",
 prompt:"Hello, welcome to my lecture"
}

Output:

1
00:00:01,000 --> 00:00:14,000
Why are there both successful and struggling entrepreneurs? 

2
00:00:15,000 --> 00:00:23,000
Many customers prefer to watch videos to enjoy online content.

3
00:00:24,000 --> 00:00:32,000
an other sentences.

• I believe 1 it should be 00:00:10,000 --> 00:00:14,000, since there is no one talking at all for 10 sec.
• Also, the 3, the speakers starts again talking at sec 28, but I'm getting the timestamp to be at sec 24. The silence is simply included in the timestamp with Whisper

Any idea how I could fix that, maybe using a prompt?

Thanks!

[mayeaux]

You can accomplish this by using word_level timestamps and then rebuilding the file yourself. I just finished that code I will publish it pretty quick.

[romain130492]

Can you do that with the API too? let me know if you publish that code, thanks!

[mayeaux]

Can you do that with the API too? let me know if you publish that code, thanks!

It's not possible with the API because of the lack of word timestamping.

With the word timestamps it's pretty trivial to rewrite the srt/vtt files, I'll publish it pretty soon once I go back to work on the related code.

[romain130492]

I see, then you need to deploy that model on your own server right.

[st-nickkimer]

@mayeaux any chance you've been able to get around to this code snippet?