Skip to content

Latest commit

 

History

History
171 lines (120 loc) · 4.45 KB

README.adoc

File metadata and controls

171 lines (120 loc) · 4.45 KB
Raw

Zamia AI

The Zamia AI project is a framework that provides a set of components needed to build completely free, open source end-to-end speech and natural language processing A.I. systems.

For more information about the Zamia AI project check out our homepage at:

Getting Started

Try Our Example Applications

The easiest way to see Zamia AI in action is to download one of our precompiled examples:

Chatbot

This is a simple text chat bot built on top of Zamia AI’s knowledge base. You can chat with the model by running the command:

zaicli chat

Here is a sample session giving you some ideas to get your conversation started:

[bofh@centos7 chatbot]$ zaicli chat
...
ai> okay computer
INFO:root:RESP: [   0.5] Greetings! | action: attention(on)
ai> what is your name
INFO:root:RESP: [   0.5] My name is "HAL 9000"
ai> who was the predecessor of barack obama
INFO:root:RESP: [   0.2] The predecessor of "Barack Obama" as "President of the United States of America" was "George W. Bush".
ai> what time is it
INFO:root:RESP: [   0.1] It is 6 minutes past 0
ai> are you a computer
INFO:root:RESP: [   0.1] Yes, I am a Computer. Are you knowledgeable about Computers?
ai> what is your favorite movie
INFO:root:RESP: [   1.0] 2001: A Space Odyssey by Stanley Kubrick
ai>

You can also run the chatbot application itself which shoud also give you a starting point on how to embed Zamia AI into you own applications:

./chatbot.py

Voice Assistant

This example show how Zamia AI can be combined with Zamia Speech Kaldi ASR Models and eSpeak NG to create a completely free, offline capable voice assistant.

Compiling Zamia AI Models

This tutorial uses the chatbot example that comes with Zamia AI. You can also find it online here:

to try it, create a directory and put all the source files there.

Now, the first step is to compile the skills that are used in this examples - i.e. Zamia AI will will compute all sample utterances plus their corresponding responses from the skills our toplevel.py skill depends on into an SQLite DB. To do this, we run the command

zaicli compile all

you can check out the number of utterances extracted from each skill using the command

zaicli stats

also, you can check the utterances extracted from all or specific skills using the command

zaicli utterances [-s skill]

Train word vectors using fastText:

wget 'http://goofy.zamia.org/zamia-speech/misc/sentences-en.txt.xz'
unxz sentences-en.txt.xz

zaicli utterances >utts.txt

cat utts.txt sentences-en.txt > corpus.txt
mkdir model
fasttext skipgram -input corpus.txt -output model/word_embeddings

Once we are satisfied it is time to train our NLP model which will allow Zamia AI to handle utterances that have no exact match in our DB:

zaicli train

Utterance Classification

Utterance classification is used to predict which of the existing skills a new utterance fits best. Can also be used on a text file containing one utterance per line.

Train the classifier:

zaicli utt_class_train

Run the prediction:

zaicli utt_class "Who is the current US President?"

Adapt the Kaldi Speech Recognition Model

Note
 This is only needed for application that use voice input. 
cat utts.txt utts.txt utts.txt utts.txt utts.txt sentences-en.txt >corpus5.txt

MODEL="/opt/kaldi/model/kaldi-generic-en-tdnn_sp"

cut -f 1 -d ' ' ${MODEL}/data/local/dict/lexicon.txt >vocab.txt
lmplz -o 4 --prune 0 1 2 3 --limit_vocab_file vocab.txt --interpolate_unigrams 0 <corpus5.txt >lm.arpa

rm -rf work
kaldi-adapt-lm ${MODEL} lm.arpa en-tdnn_sp
tar xfJ work/kaldi-en-tdnn_sp-adapt.tar.xz

Further Documentation

For other technical documentation, check out

License

My own scripts as well as the data I create is Apache-2.0 licensed unless otherwise noted in the script’s copyright headers.

Some scripts and files are based on works of others, in those cases it is my intention to keep the original license intact. Please make sure to check the copyright headers inside for more information.

Author

Guenter Bartsch <[email protected]>