cloudchenl

Results 2 issues of cloudchenl

output dimension order is random, and lead to wrong synthesis

2 comment

in composition stage: for stream_name in list(out_dimension_dict.keys()): if stream_name not in stream_start_index: stream_start_index[stream_name] = stream_dim_index list(out_dimension_dict.keys()) results in random order, maybe [bap, lf0, vuv, mgc] or [lf0, bap, vuv, mgc]......

tf.exp(log_var) is variance not std

https://github.com/rishikksh20/vae_tacotron2/blob/0ab0fb964be82111a0f57f6cec92d7e682fe20b1/tacotron/models/modules.py#L20