[Tuning] iwslt22 low-resource ST decode configuration tuning

[freddy5566]

Summary

Tune decode penalty for better performance in terms of BLUE score.

Results

clean_set: from 2.6 --> 2.7 (BLUE score) / 0.2 --> 0.4 (penalty)
full_set: from 3.6 --> 3.7 (BLUE score) / 0.2 --> 0.3 (penalty)

dataset	score	verbose_score
decode_st_transformer_st_model_valid.acc.ave/test	2.7	22.1/4.6/1.9/0.8 (BP = 0.756 ratio = 0.782 hyp_len = 17593 ref_len = 22504)

st_full_wav2vec-transformer-warmup-15k

dataset	score	verbose_score
decode_st_full_transformer_st_model_valid.acc.ave/test	3.7	24.4/5.2/2.0/0.9 (BP = 0.928 ratio = 0.931 hyp_len = 20945 ref_len = 22504)

[codecov]

Codecov Report

Merging #5019 (84f8055) into master (b7885c2) will decrease coverage by 0.02%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #5019      +/-   ##
==========================================
- Coverage   77.02%   77.00%   -0.02%     
==========================================
  Files         606      606              
  Lines       53748    53743       -5     
==========================================
- Hits        41399    41386      -13     
- Misses      12349    12357       +8     

Flag	Coverage Δ
test_integration_espnet1	66.15% <ø> (-0.15%)	⬇️
test_integration_espnet2	47.96% <ø> (ø)
test_python	66.84% <ø> (-0.01%)	⬇️
test_utils	23.28% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

see 8 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more