ESPnet-SPk: major update

[Jungjee]

What?

This PR combines several previous working PRs with new functions.

• ESPnet-spk: add more model architectures #5385, ESPnet-SPK: add inference #5398
  This PR newly introduces:

• Speed perturbation as speaker augmentation

• Frontend design modification -> we now can train the model with frozen/free SSL front-ends through S3PRL

• Additional Speaker models: ECAPA-TDNN and MFA-Conformer

• Adds novel on-going research on representation learning

• Inference stage, a separate inference stage for evaluating models with EER and minDCF.

• (updated) Supported configurations

  • RawNet3 - EER 0.73% on Vox1-O
  • ECAPA-TDNN w/ mel-spectrogram - EER 0.96% on Vox1-O
  • ECAPA-TDNN w/ fixed WavLM-Large frontend - EER 0.60% on Vox1-O

Why?

We have been adding several new features.
Several ongoing PRs have conflicts in places.

See also

[Jungjee]

test_python seems to have timeout errors randomly (maybe better to increase timeout threshold?).
ci on windows / check_installable_on_windows suddenly started to fail.

currently I'm just closing and reopening..

[sw005320]

@Emrys365, we still observe an issue in https://github.com/espnet/espnet/actions/runs/6738909252/job/18319405795?pr=5408#step:9:694
Can you check it?

[Jungjee]

@Emrys365, we still observe an issue in https://github.com/espnet/espnet/actions/runs/6738909252/job/18319405795?pr=5408#step:9:694 Can you check it?

These errors seem to be coming from the files/unittests that wasn't edited in this pr.

Maybe the update here (4c96187) is the reason?

[Emrys365]

Interesting. I thought fixing the input in my PR should avoid this issue. I will look at it later.