search

alex-berard / seq2seq

Attention-based sequence to sequence learning
Apache License 2.0
388 stars 122 forks source link

Feature extraction #25

Closed un-lock-me closed 5 years ago

un-lock-me commented 5 years ago

Hi

thanks for sharing your code with us. Is this model good for feature extraction or it has been designed only for the translation purpose?

Thanks:)

alex-berard commented 5 years ago

Hi, It depends on what you mean by "feature extraction". The last layer of an LSTM (which sums up the entire input sequence), or even the attention matrix can be used as features for auxiliary tasks.

If you are talking about low-level audio features, it turns out that the speech-to-text models implemented here take as input pre-computed features (e.g., MFCCs), so they would not be very useful for this purpose.