timestamp
#39
by Sh1man - opened
How can I get timestamps?
You can use the position of the [STREAMING_WORD] token as a proxy for the end timestamp of an emission group. Specifically, since you know the transcription-delay in frames, you can offset the idx of the [STREAMING_WORD] token by this value to get the proxy position for the end timestamp in the token space. Then convert this to the time space to get the value in seconds.
Sure you can always build something, but it would be better to have the word level and segment level timestamp back in a json file so that we can work with this. When is it planned? I believe you can do this with the API already so why not with the vLLM?