Sie sind auf Seite 1von 15




# Names Registration Number Signature






We would like to thank the staff of National Agricultural Research Organisation for providing all
the necessary information to the BSE20-29 group about agriculture to aid in the development of
the radio data keyword spotting system for crop related information.

We would also like to thank our academic supervisor, Mr. Alex Mwotil and the AI lab at Makerere
university for their guidance towards the success of this research.

In societies with well-developed internet infrastructure, social media is the leading medium of
communication for various social issues especially for breaking news situations. In rural Uganda
however, public community radio is still a dominant means for news dissemination. Community
radio gives audience to the general public especially to individuals living in rural areas, and thus
plays an important role in giving a voice to those living in the broadcast area. It is an avenue for
participatory communication and a tool relevant in both economic and social development. This is
supported by the rise to ubiquity of mobile phones providing access to phone-in or text-in talk
shows. In this paper, we describe an approach to analysing the readily available community radio
data with machine learning-based speech keyword spotting techniques. We identify the keywords
of interest related to agriculture and build models to automatically identify these keywords from
audio streams. Our contribution through these techniques is a cost-efficient and effective way to
monitor food security concerns particularly in rural areas. Through keyword spotting and radio
talk show analysis, issues such as crop diseases, pests, drought and famine can be captured and fed
into an early warning system for stakeholders and policy makers. This report contains five chapters
that explains our findings about Radio data keyword spotting information system for crop
related information Project. The content of all the chapters is broadly explained and it is
constructed from the practical basis of the site. The chapters covered are introduction,
Methodology, Results, Discussion, and Recommendations.

NARO National Agricultural Research Organisation

AI Artificial Intelligence
KWS Keyword Spotting
LAB Laboratory
CNN Convolutional Neural Network
1 Introduction
1.1 Background

Many farmers in Uganda use radios as the easiest means of getting information. At least more than
70% of Ugandan farmers have radios in their homes which helps them to listen to agricultural talk
shows on those radios [1]. Most radios in Uganda air agricultural talk shows or programmes which
usually help the farmers to get much information about their crops.

Keyword Spotting systems (KWS) is a classification task that aims at detection and retrieving of
a series of words from a database of audio streams. The advantage of using a KWS is that unlike
full automatic speech recognition systems, they can be developed without sufficient labelled data.
This is common especially in low resourced languages [2]. In this paper, we discuss an
implementation of a Keyword Spotting model that we use to mine local community radio content
using specific keywords for a low resourced language in Uganda. We evaluate our approach on
the Luganda language which is a low-resource language that is currently spoken and used in many
of the Agricultural communities in Uganda.

This research was conducted to get information about how crops are usually talked about on those
radio stations that usually have programmes to talk about agriculture and how people usually listen
to those programmes. This can help us to develop a system will be able to spot those keywords
related to crops.

At the end, we are to develop the “Radio Data keyword spotting system for crop related
information” that will filter the words that are related to crops from the voice recordings from those
radio stations, these keywords can then be used to help understand any issue related to such crops
which were talked about in the programme.

1.2 Objectives

The objectives of this research are to:

 To find out how information about crops is got.

 To investigate the problems faced by farmers with their crops.
 To find out how often farmers are sensitized about their crops.
 To find out the methods used to reach out to farmers.
 To find out the challenges faced while trying to know issues about crops.
 To find out whether agricultural talk shows or programmes help to reveal issues with
farmer’s crops
2 Literature Review
Keyword spotting for low resource languages has been implemented before Menon et al. [3] The
approaches include used include CNNs, Siamese CNNs Bromley et al. [2] and autoencoders. Other
models are designed for low computational resources such as Tang & Lin [4] Coucke et al [5] due
to the popularity of using keyword spotting to identify commands in smartphones, battery concerns
from CPU requirements comes into play. Low resource languages pose a problem as models that
consume a lot of data in training fail to converge due to low volume of text corpora and speech

Luganda is an almost zero-resource Bantu language, spoken in the central region of Uganda. Work
on Luganda remains small compared to larger languages such as Kiswahili, Zulu and Hausa.
Research on Luganda exists in machine translation Nandutu [6], keyword spotting Menon et al.
[3]. Prior work on Luganda keyword spotting and radio monitoring of Luganda community radio
has been initiated by Menon et al [3] to generate insights concerning humanitarian aid and
development. While radio content is publicly available and accessible with seemingly no
data/privacy restrictions, there have been few interventions seeking to mine this data for
surveillance purposes particularly for crop pests and diseases.

3 Model Architecture and Design

In this study, we use 1-dimensional convolutional neural network(CNN), a Siamese convolutional
neural network(CNN), and time delay networks to implement keyword spotters.

3.1 1-cd CNN

The 1-cd CNN takes in a input the processed raw audio data. The input to the model is an array
representing the audio wave form(X). The network is designed to learn the set of parameters (θ)
to map the input to a prediction (T) according to a hierarchical feature extraction given by equation

T=F(X|Θ) = fL(...f2(f1(X|θ1)|θ2)|θL)
where L is the number of hidden layers in the network. The final architecture we created is a 14-
layer deep neural network with five 1-d convolutional layers with each with intermediate pooling
layers. A dropout of 0.5 was also applied after the two successive dense layers. The final layer is
a softmax activation function to map to only 10 target keywords selected randomly from the
Luganda/English corpus. The model will be trained using batch gradient descent with the Adam
Optimizer and a learning rate of 0.001.

4 Methodology
4.1 Instruments

We used questionnaires which we issued out to people at NARO to obtain quantitative data.

Notebooks and pens were also used to note down all the additional information that we needed
from the field.

See Appendix 7.1 for the questionnaire.

4.2 Sample

Our sample size was 30 agricultural experts and officials. The sample was randomly selected in
order to cover a satisfactory number of agriculture experts. Due to the fact that the number of
agriculture experts is small we had to consider some other top individuals in an organisation to be
part of our sample, this is due to the fact that those people have got broad knowledge about crops
and agriculture at large.

4.3 Data Collection

We collected the data as a group of four people who are working on the project. We went to the
field with questionnaires which we distributed among different people at NARO. We noted down
any additional information which was not covered by the questionnaires.
4.4 Data Analysis

Data was manually entered into the computer using Microsoft Excel and then converted to a csv
file which we then analysed it using R and R studio in order to visualise it and get different plots
like bar graphs and density plots.

4.5 Limitations

We faced the challenge of going through a lot of protocol that is to say we had to go to a number
of offices in order to get the rights to talk to the people we wanted to talk to. Another challenge is
that most of our meetings were postponed several times which negatively affected our normal
schedule of the project.

The choice for the sample size gave us a hard time due to the fact that the population of agriculture
experts was relatively small.
5 Results
We got the following results after analysing findings from the field. This helped us to get a clear
view on our key questions
How do you get information about crops in different parts of the country?
Which problems do farmers face with their crops?
Do you think agricultural talk shows on radio stations talk about crops?

The relationship between crop issues that are likely talked on radio stations on those programmes
and how relevant are those agriculture programmes on radio talk about crops.
6 Discussion
Basing on the results obtained after analysing the findings, we found out that the radio stations are
the means commonly used in getting information about crops. As per the results we are focused
on using those audio stream from the radio stations in implementing our project. The results
indicate that people rarely go directly to farms to get information about crops.

The radios are commonly used based on the number of reasons, but based on the responses
collected we found out that the major reason why its commonly used is because it is very cheap to
own a radio at least in every family.

The results also indicate that the most experienced problem with crops is pests and diseases. This
means that our system will help the users to identify those areas which are affected by diseases so
easily after getting any information related to diseases from the keywords spotted from the audio
From the results, we can see that most of the agriculture talk shows on radio talk about crops in
particular, almost every programme on the radio about agriculture has to talk about crops. This
will give us a big dataset where we can easily spot those keywords related to crops.

Agriculture experts or officials face different challenges in trying to know which problems are
faced by farmers with their crops, but from the analysis made, it shows that these officials lack
enough information about those problems in different areas. So by developing this system, all the
crop information will be available and it can be easily retrieved.

7 Recommendations
Since radios are the mostly used means to broadcast agricultural programmes, the Crop-based
keyword spotting system will use the audio streams from those radio stations as the input in order
to get those words related to crops. Due to the fact that crops are mainly affected by pests and
diseases, these keywords filtered by the system can then be used to filter out those ones pointing
to crop diseases if there is any.


[1] W. Bank, “Making farming more productive and profitable for Ugandan farmers,” 19 June 2018.
[Online]. Available:

[2] F. International, “barzawire,” 25 09 2019. [Online]. Available:


[3] C. O. M. Daniel Mutembesa, “Crowdsourcing real-time viral disease and pest information: A Case
of nation-wide cassava disease surveillance in a developing country,” In a sixth AAAI Coference
on Human Computation and crowdsourcing, 2018.
[4] I. G. L. Jane Bromley, “Signature verification using a "siamese" time delay neural network,” in In
Advances in Neural information processing systems, 1994, pp. 737-774.

[5] J. L. Raphael Tang, “Deep Residual Learning for small-footprint keyword spotting,” 2018.
[Online]. Available:

[6] I. Nandutu, “Luganda text-to-speech machine,” School Computing and Engineering,Uganda

Technology and Management University, 2016.

[7] M. C. G. L. P. L. Alice Coucke, “Efficient keyword spotting using dilated convolutions and
gating,” May 2019. [Online]. Available: