74806b744b
* disable speed perturbation by default
* minor fixes
* minor updates
* updated bash scripts to incorporate with the `speed-perturb` arg
* minor fixes
1. changed the naming scheme from `speed-perturb` to `perturb-speed` to align with the librispeech recipe
>> 00256a7669/egs/librispeech/ASR/local/compute_fbank_librispeech.py (L65)
2. changed arg type for `perturb-speed` to str2bool
Introduction
This recipe trains multi-domain ASR models for AliMeeting. By multi-domain, we mean that we train a single model on close-talk and far-field conditions. This recipe optionally uses [GSS]-based enhancement for far-field array microphone. We pool data in the following 4 ways and train a single model on the pooled data:
(i) individual headset microphone (IHM) (ii) IHM with simulated reverb (iii) Single distant microphone (SDM) (iv) GSS-enhanced array microphones
This is different from alimeeting/ASR since that recipe trains a model only on the
far-field audio. Additionally, we use text normalization here similar to the original
M2MeT challenge, so the results should be more comparable to those from Table 4 of
the paper.
The following additional packages need to be installed to run this recipe:
pip install jiebapip install paddlepaddlepip install git+https://github.com/desh2608/gss.git
./RESULTS.md contains the latest results.
Performance Record
pruned_transducer_stateless7
The following are decoded using modified_beam_search:
| Evaluation set | eval WER | test WER |
|---|---|---|
| IHM | 9.58 | 11.53 |
| SDM | 23.37 | 25.85 |
| MDM (GSS-enhanced) | 11.82 | 14.22 |
See RESULTS for details.