E6893 Big Data Analytics:

Speech Analytics Software Library

Team Members: Kyle White

December 9, 2014
1 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Motivation

• A need exists in the speech processing and recognition community to adapt or create
tools for coping with increasing amounts of available data.

• Example: A speech recognition project can require on the order of 50GB or more of
training audio data for acoustic modeling to realize gains for popular approaches such as
decoding speech via neural network classifiers. Transcribed domain data on the order of
gigabytes is needed for high performance language models. This scenario calls for
speech processing solutions that are able to store, access, and process this volume of
data.

2 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Software Library Goals

• The goal of the speech analytics library is to assemble, adapt, or develop capabilities to
condition, decode, and understand speech content in an archive of speech related data.

• The initial goal of the library, for this course project, is to utilize an open-source language
modeling toolkit in combination with the MapReduce paradigm to enable distributed
training of n-gram language models.

3 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Library Overview

• The Speech Analytics Library is intended to be a collection of tools, and the single tool
that currently inhabits the library allows the distributed training of n-gram language models
for speech recognition. In the first phase, Hadoop is utilized to count n-grams across the
input transcriptions. The intermediate output is an n-gram document in a simple, RSA
internal format. Phase two compiles the intermediate n-gram count documents and
utilizes berkeleylm for the creation of a n-gram language model in the commonly accepted
ARPA format.

Language Model Generation Tool

Transcripts
Phase I N-gram count Phase II ARPA Format
documents Language
Model

4 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Language Model Generation: Phase I

• Phase one processing is centered around the usage of Hadoop distributed computing and
Lucene text analysis libraries.
• The Hadoop Library is utilized to create a Java MapReduce application capable of
taking textual transcriptions as input and producing output n-gram count documents.
Custom Client, Mapper, Reducer, and Partitioner classes are utilized in this step.

• The Lucene core library capabilities for analysis/tokenization are utilized by custom
Mapper classes to perform analysis on input for n-gram creation from transcript
strings. The sequence of Lucene TokenStream filters is shown below.

WhitespaceTokenizer

LowerCaseFilter

ShingleFilter

5 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Language Model Generation: Phase II

• Phase two processing is centered around the usage of the berkeleylm library to generate
a language model. At this time the second phase simply outputs this language model to
an ARPA format language model file. Alternatively, berkeleylm excels at efficiently storing
and querying large language models during runtime for a speech recognizer, and it is for
possible future uses related to this shared focus on large data sets that berkeleylm was
chosen.

Example Input Example Output

6 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Library Dependencies

• The library utilizes the following existing tools to fulfill its goals.

Capability Leveraged Tool

HDFS Apache Hadoop
MapReduce Apache Hadoop
Text Segmentation and Apache Lucene
Tokenization
Language Modeling berkeleylm

• Hadoop – Chosen for solid file system and implementation of MapReduce framework.
Selected for strong usage scenario with batch processing an audio archive of speech
related data.
• Lucene – Top level Apache product with track record for text analysis and tokenization
• berkelylm – Library written for language modeling in the context of large data volumes.
Meant for estimating, storing, and accessing large n-gram language models

7 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Usage

• The library is currently a two phase affair, where each phase must be kicked off
from the command line. Simplification to a single phase is planned. The library
is built via Maven, but a pre-built JAR is included in the repository.
• Example scripts to run the two phases are included in the repository /scripts
folder, and an example of one script is below.

8 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University
Conclusion

• Work completed so far allows distributed training of n-gram language models, and
presents the opportunity for additional work to better utilize berkeleylm as a language
model manager during runtime operations which potentially depend on large language
models.

9 E6893 Big Data Analytics – Final Project Presentation © 2014 CY Lin, Columbia University