Name: MELANIE

Text: MELANIE
Topic Spotting
TOP

CONFIDENTIAL
– Subject to–change
prior
notice!prior notice!
CONFIDENTIAL
Subjectwithout
to change
without

1

MELANIE - Topic Spotting, TOP
Production Module for fast and reliable spotting of the topic/category prevailing in
speech signals. Training module for topic spotting parameters.

Functionality
Speech signals can be analysed according to
the topic/category prevailing in the speech. For
each speech signal, an overall score and
probability is estimated. In addition, these
values can also be estimated for any time
interval.

Training of Models
Goal
Training Data

Training is performed in order to optimise the models
on the topic and channel of the application.
The training data must be labeled according to the
topic in the speech signal. The data must be
representative and as similar as possible to the data
of the application. A minimum of one hour of data is
required for each language.

Technical Data
OperatingSystem

Available on Windows XP and Linux

Processing Time

• On a P4; 3GHz, 1GB RAM computer
• Classifying 3 different classes/ topics
• 1/14 of real time; i.e. a signal of 14 minutes length is
processed in one minute

Recognition Rate

The recognition rate depends on the number of
topics, on the similarities and the overlap among
topics, and on the speech quality. For example, when
distinguishing 2 topics in radio quality, an accuracy of
87 % can be obtained.

Products
Mel-Pr-TOP-site
Mel-Pr-TOP-PC
Mel-Tr-TOP

Production environment for MELANIE TOP (site
license)
Production environment for MELANIE TOP (PC
license)
Training environment for MELANIE TOP

CONFIDENTIAL – Subject to change without prior notice!

2

Consequently, this means that a recording with a signal length of 20
minutes can be fully classified in one minute.

CONFIDENTIAL – Subject to change without prior notice!

3

Document Path: ["melanie-topic-spotting-top.pdf"]

e-Highlighter

Click to send permalink to address bar, or right-click to copy permalink.

Un-highlight all Un-highlight selectionu Highlight selectionh