Provide .srt with backend TensorRT

[PatLemaire]

Hi,
After several attempts to modify the client, I notice that the 'segments' returned by the TensorRT backend are not in the same form as those in faster_whisper. There is no start, end or completed. As a result, it is not possible to obtain a transcription for this backend.
Is it possible to modify the segments so that they are compatible?

[makaveli10]

Yes, because tensorrt backend has a different implementation for the token decoding i.e. it doesnt decode with time-stamps. We have a closed issue on TensorRT-LLM => NVIDIA/TensorRT-LLM#647

Feel free to open a PR

[igarzonpres]

@makaveli10 Do you have any update on this? Some implementation ideas?

[makaveli10]

@igarzonpres Refer to NVIDIA/TensorRT-LLM#647 (comment) for using Forced aligner.