Lecture 18 - End-to-end systems
This lecture mainly discussed end-to-end or HMM-Free approaches to speech recognition, in particular the RNN encoder-decoder architecture.
At the strart of the lecture we covered current approaches to robust speech recognition, f
continuing lecture 15. The recordings for that part of the lecture are in lecture 15
We discussed the idea of HMM-Free systems for speech recognition in which the temporal modelling is not by by an HMM, but by a recurrent network mapping from acoustic features to a word sequence directly.
RNN Encoder-Decoder Architecture
- The basic idea of mapping between variable-length sequences, using an encoder to create a hidden representation (context) from input sequence, and a decoder to write out the output sequence from the hidden representation
- RNN decoder architecture
- RNN encoder architecture and the attention model to create the context from the sequence of encoder hidden states
- Conclusions and some experiments
This page maintained by Steve Renals.
Last updated: 2017/03/28 12:58:24UTC
|Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, Scotland, UK
Tel: +44 131 651 5661, Fax: +44 131 651 1426, E-mail:
Please contact our webadmin with
any comments or corrections. Logging and Cookies
Unless explicitly stated otherwise, all material is copyright ©
The University of Edinburgh