The data is exactly the same as in homework 1 except we add one file 'phones/48_idx_chr.map' which provides the mapping from 48 phoneme to the corresponding index and English character. You may download the file '48_idx_chr.map' directly if you already have the training data in homework 1.
\r\nAlso, we provide you the posterior gram of both testing and training set after a trained DNN. If you didn't get good result in homework 1, then you may try to train you RNN with this given posterior gram.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "1. Introduction
\r\n In this homework, you and your teammates are asked to implement the recurrent neural network(RNN) as a classifier. The main task here is to re-evaluate the transition probabilities between phone sequences. Thus, we are going to use the edit distance evaluation method to evaluate your answer, i.e. phone sequences.
Again, you may need the domain knowledge about Automatic Speech Recognition(ASR) to get better comprehension in this work, but little is needed to complete the actual homework.
2. Problem Description
Goal: Given the input feature of each time frame, utilize recurrent neural network to predict the phone sequence.
Input: The posterior gram of phones, the MFCC features, the raw wave files, or any other derivations.
Output: 48 phone label sequences after trimming.
You are asked to implement at least but not limited the following methods:
\r\n rnn.init(): initialize model parameters
rnn.forward(): forward layer-by-layer to obtain all activation values in the neurons.
rnn.calculate_error(): calculate the error signal according to the RNN output and the target function
rnn.backpropagation(): backpropagation the error signal back through layers.
rnn.bptt(): backpropagation through time through layers.
rnn.update(): update the RNN parameters.
rnn.predict(): predict the output labels
rnn.save_model() / rnn.load_model(): basic model io that can store the parameters you trained.
3.Implementation:
Different from Homework 1, RNN is recurrent in time domain and thus have dependencies on each frame. You may try to implement either a CPU or a GPU version.
We recommend 3 types of GPU programming, CUDA C/C++, Python Theano, Matlab GPU Toolbox. We expect you to explore any of them by yourselves and utilize then in your RNN implementation. However, the difficulty is quite different, and you can use any programming language with any open-source matrix manipulation package to achieve your goal. Yet, the choice is up to you.
However, the same as Homework 1, we strictly forbid any form of NN-based toolkit utilization. Please remember that the best way to learn is always to implement by yourselves, although the procedure may be quite harsh.
The evaluation metric for this competition is Levenshtein distance(a.k.a. Edit-distance)(The lower the better).
\r\nSubmission files should contain two columns: id and phone_sequence.
\r\nid is the utterance name (eg. fadg0_si1279)
\r\nphone_sequence is the corresponding trimmed phone sequence (eg. HrLAJarDeBLMrDcLMwU).
\r\nThe file should contain a header and have the following format:
\r\nid,phone_sequence", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "\r\n\t\r\n
fadg0_si1279,HrLAJarDeBLMrDcLMwU
fadg0_si1909,vbLAFKnLhyUwJmrBJLAwLSyLAwKr
fadg0_si649,lwLJctryJvrCaBgLHwDLKyDwLHJywDLHrLHwJnDryJLAbLMtrBwLABsmrQKn
fadg0_sx109,SJKyJBnLMwDJLHIyDyCrFLABSaDwDJLMyLAJBc
...
You cannot sign up to Kaggle from multiple accounts and therefore you cannot submit from multiple accounts.
\r\n\r\n Privately sharing code or data outside of teams is not permitted.\r\n It's okay to share code if made available to all participants on the forums.\r\n
\r\nTeam mergers are not allowed in this competition.
\r\n\r\nThere is no maximum team size.
\r\nYou may submit a maximum of 10 entries per day.
\r\nYou may select up to 10 final submissions for judging.
\r\nStart Date: 10/26/2015 9:48 PM UTC
\r\nMerger Deadline: None
\r\nEntry Deadline: None
\r\nEnd Date: 11/20/2015 7:00 AM UTC
\r\n\t\r\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }