Pat York - University of Nevada, Reno - Iridium Technolgy

 

GE Database Management

Video

Speech Recognition and Automatic Speech Data Curation (Speech & Transription Chunking)

During the Spring Semester, 2015, at UNR, I took CS491Q "Topics in Machine Learning" (unofficially entitled "Deep Learning") taught by Dr. Richard Kelley. We were assigned a semester project in a topic of our choice, provided that the project was based on a deep learning method. I chose to implement the paper "Deep Speech: Scaling up end-to-end speech recognition" by Ng et al. [arXiv:1412.5567v2].

Read more about my implementation of the Deep Speech paper for speech recognition.

In training the deep model for speech recognition, I was greatly pleased to find thousands of hours of audio with corresponding text. I was then dismayed to find the length of the individual audio files and transcriptions. Connectionist Temporal Classification (CTC) [Graves et al. 2006] has a tendency to underflow when the input to the model and/or expect outputs to the model are too large. With copious amounts of data in a poor, temporal format, I set out to create a system by which long duration audio files (and their transcripts) can be "chunked" into smaller audio files of 5-10 seconds, while retaining the text that is contained in the smaller slices of audio. The Deep Speech paper utilized a proprietary dataset of 5,000 hours of audio; manually "chunking" 5,000 hours of audio would require about 5,000 man hours of labor. However, with my approach, 10,000 hours of audio/text can be automatically chunked to 5,000 hours of smaller fragments in approximately 10 hours of human effort (yet to be proven fully).

Read more about my Automatic Speech Data Curation ("Chunking") system.
And, coming soon, download a dataset of X,XXX hours of high-quality tagged speech data, for your own use!

Twitter Sentiment Analysis

During the Spring Semester, 2014, at UNR, I took CS491H officially entitled "Database Systems" and unofficially entitled "Data Science & Big Data" taught by Dr. Richard Kelley. We had a semester project which I completed with Tommy Avant. We decided on using Twitter to predict the stock market. To accomplish this, we looked for a correlation between the sentiment of Twitter and the prices of various stocks. We ended up limiting the scope of our project for time restrictions, but we are continuing to approach this idea.

Read more about Twitter Sentiment Analysis.

Home Lab via VMWare ESXi

An interesting problem that fell out of the Twitter Sentiment Analysis project was the need for a series of always on servers, both Windows and Linux. Luckily, in the Spring semester, 2014, I was taking CS447 officially entitled "Systems Administration" and unofficially entitled "Cloud Computing" taught by Nancy LaTourrette. In this class we were exposed to the ins and outs of virtualizing environments. We were also given access to UNR's Acedemic Cloud Services in its pre-rollout phase, and I was given access to a set of VMs for the Twitter Sentiment Analysis project as well.

At the conlcusion of the Spring semester, 2014, the ACS went offline for the summer in preparation for its rollout in the Fall. Because of this I was in need of a new set of servers or VMs.

Enter VMWare ESXi. Using my newfangled virtualization knowledge and the omnipotent knowledge of Google, I ordered a set of hardware and installed ESXi. Below is some documentation of my process as well as tutorials and advice for those building their own ESXi home lab from scratch (work in progress; adding as I discover/require more).

Current Hardware:

Current Software and VMs:

  • VMWare vSphere Hypervisor Version: ESXi 5.0 Update 3 (ESXi 5.0u3)
  • VMs:
    • Always-on:
      • Ubuntu 12.04 LTS x64 (2)
      • Windows 7 Home Premium x64 (1)
    • Intermittent-on:
      • Windows Server 2008R2 x64
      • Windows Server 2012 x64

Project Overview and Impetus
Hardware Considerations
Cooling and Noise Issues
Video Card Passthrough
USB Passthrough (as a PCI Device)
FTP and SSH to the ESXi Host
Multiple Datastores and Physical Disks
VM Duplication, Templating, Moving, Deletion, and Snapshotting
Tested Hardware, Peripherals, and OSs