Is there anyway we can get the (approximate) time/frame of the ouput

[samshipengs]

Given a clip of 5s speech, is there anyway to get the rough time when each word was spoken?
e.g. the output might look like this
"hello world here is a question"
0s 2s 3s 4s 4.5s 5s
(or, the frame numbers from input waveform)

[snakers4]

There is an example for PyTorch producing timestamps in the colab

[samshipengs]

I see, thanks for pointing out. By turn on the word_align flag in decoder there will be time alignment outputs. But it seems like the timing of each word is all the same i.e. wav_len / len(argm) I think, wouldn't there be issue for prolonged pronunciation (someone saying a word very slowly)? and another question is, I didn't quite understand how the decoder function code calculates the beginning time of each word, would you mind sharing some high level intuition on that? (sorry im a newbie to audios...)

[snakers4]

In this particular case basic lm is integrated into the model so it knows when words start and end

Longer pronunciation to a reasonable extent should not be a problem since this was accounted for during training