populate recording_id for Cut when using Cut.compute_and_store_featur…

[freewym]

…es()

Sometimes we might want to know the corresponding recoding of a specific cut for debugging or evaluation, e.g. reference transcripts are usually provided in the format of <recording_id> <reference>.

Currently we use Cut.compute_and_store_features() to extract features but this API does not fetch Cut.recording.id, so Cut.recording_id will return None. On the other hand, Cut.id is not None but it is a random string, making it a bit difficult to find the corresponding recording of a cut.

This PR is trying to copy Cut.recording.id to Cut.features.recording_id when calling Cut.compute_and_store_features().

@pzelasko LMK what you think.

[freewym]

I realized that a cut may be from multiple recordings... In that case maybe we can leave the recording id as None?

[danpovey]

Why did you want to know that information? The Supervision would always come from a specific recording I think.

…

On Mon, Nov 23, 2020 at 8:50 AM Yiming Wang ***@***.***> wrote: I realized that a cut may be from multiple recordings... In that case maybe we can leave the recording id as None? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#147 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAZFLOZITRCSCHFCGUOTUZDSRGWUBANCNFSM4T62XXFA> .

[freewym]

Why did you want to know that information? The Supervision would always come from a specific recording I think.
…
On Mon, Nov 23, 2020 at 8:50 AM Yiming Wang @.***> wrote: I realized that a cut may be from multiple recordings... In that case maybe we can leave the recording id as None? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#147 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAZFLOZITRCSCHFCGUOTUZDSRGWUBANCNFSM4T62XXFA .

OK, you are right, SupervisionSegment always has non-empty recording_id. I was also thinking of a scenario where the supervisions are not available (e.g. in evaluation), and we may want to save the decoded results in a file to evaluate them later. Maybe an easier fix is switching the if else condition at the following line

lhotse/lhotse/cut.py

Line 259 in fd5d25e

return self.features.recording_id if self.has_features else self.recording.id

as long as Cut.recording is not None.

[pzelasko]

I am fine with that last change you suggested.

[freewym]

I am fine with that last change you suggested.

Done