Open grig-guz opened 6 years ago
I can do this one
@AhmedAbdelmoneim make sure to assign yourself when you choose a task
Here are some papers I found:
https://www.isca-speech.org/archive/archive_papers/interspeech_2011/i11_0437.pdf - used CD-DNN-HMMs, achieved a WER (word error rate) of 18.5%
http://proceedings.mlr.press/v32/graves14.pdf - used LSTM RNN, achieved a bunch of WERs ranging from 6.7% to 27.3%
https://www.isca-speech.org/archive/archive_papers/interspeech_2010/i10_1045.pdf - used RNN Language Model, achieved a WER of 9.5%
https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=7472621&tag=1 - I don't understand the architecture completely, but basically divided it into two components, a listener and a speller, achieved a WER of 14.1%. Also cool because they have a github repo with most of their work. Link to repo: https://github.com/v0lta/Listen-attend-and-spell
Deep Speech paper: https://arxiv.org/pdf/1412.5567.pdf Repo: https://github.com/mozilla/DeepSpeech
Research :man_scientist:
Before starting, we need to take a look at what people already did. Possible sources:
Task
Find at least 3 approaches, and discuss/evaluate them in terms of accuracy/complexity/need for data