📎End-to-End Speech Recognition by Following my Development History - Guest Lecturer Shinji Watanabe

以前的筆記

End-to-End Speech Recognition by Following my Development History | Guest Lecturer Shinji Watanabe - YouTube

Impression in ~2015!

• Attention based encoder decoder
• No conditional independence assumption
• Attention mechanism allows too flexible alignments(太難train)

$\arg \max_W p(W|X) = \arg \max_W \prod_j p(w_j|w_{<j}, X)$

Implementation 2016

CTC

• Relying on conditional independent assumptions(similar to HMM)
• Output sequence is not well modeled(no LM) LAS
• Attention allows too flexible alignments
  • Too hard to train from scratch

Input/output Alignment desired ASR system is monotonic (一條線) 但 End-to-end train 的模型很難是一條線

Use both benefits of CTC and attention

Hybrid CTC/Attention network working well!

• Attention based 的模型要 train 起來要花太多時間才能收斂，其平行訓練帶來的好處其實沒有比 RNN 多。

+ joint decoding

Discussions

Check

• Attention pattern!
• Learning curves!

Speech recognition pipeline

傳統的 pipeline: 太多獨立假設才能解