rickychanhoyin

minor correction

better regularization / generalization, and avoid nan value during training (for audio file with all zero samples)

Conversion fail for subsampling > 1

3

python converter/convert.py --input=text.mdl --output=final.onnx --trans-model=final.trans --conf=final.conf --chunk-size=50 --left-context=39 --right-context=39 --modulus=3 --subsample-factor=3 --nnet-type=3 2021-02-25 13:42:05,502 root INFO frames per chunk: 51, left-context: 39, right-context: 39, modulus: 3 2021-02-25 13:42:43,973 parser INFO...

Very minor change in alimeeting recipe (revised)

11

No need to load_audio. alimeeting audio data is wav format, default export_to_webdataset uses "flac". If load_audio uses default (True), webdataset would show "[Suppressed TypeError] Error message: save() got an unexpected...