# Description

HMM GMM Binary English speech models (262K dictionary, 32 bit LM)

To use these models you need a modified version of Julius from:

https://github.com/palles77/julius

You can use a precompiled version of Julius for Windows experiments 'julius-dnn.exe' (you might need to install latest VC REDIST x86 version from https://support.microsoft.com/en-gb/help/2977003/the-latest-supported-visual-c-downloads). Otherwise you need to download source code from the address above (https://github.com/palles77/julius), configure it and compile it using either VS 2017 or GCC.

When compiling using GCC you can do it for example as:

$ bash configure --enable-words-int --enable-sp-segment

$ make all

$ make install

# How to run

julius-dnn -C julius.jconf -dnnconf dnn.jconf > julius-dnn-output.txt

Note that:

1. This compilation has WORDS_INT precompiler directive set as 1 (which enables 32 bit based Language Models).

2. A special directive DNN is set to 1 in libsent/include/sent/stddefs.h to enable compatibility with DNN HTK based acoustic models MFCC features calculation.

# Files in this directory

dnn.jconf - DNN config file

ENVR-v5.3.am - 16 Gaussians general triphone model, 32 gaussians silence model, binary format

ENVR-v5.3.dct - dictionary

ENVR-v5.3.layer2_bias.npy - layer 2 bias

ENVR-v5.3.layer2_weight.npy - layer 2 weights

ENVR-v5.3.layer3_bias.npy - layer 3 bias

ENVR-v5.3.layer3_weight.npy - layer 3 weights

ENVR-v5.3.layer4_bias.npy - layer 4 bias

ENVR-v5.3.layer4_weight.npy - layer 4 weights

ENVR-v5.3.layer5_bias.npy - layer 5 bias

ENVR-v5.3.layer5_weight.npy - layer 5 weights

ENVR-v5.3.layer6_bias.npy - layer 6 bias

ENVR-v5.3.layer6_weight.npy - layer 6 weights

ENVR-v5.3.layerout_bias.npy - output layer bias

ENVR-v5.3.layerout_weight.npy - output layer weights

ENVR-v5.3.lm - binary language model

ENVR-v5.3.mfc - mfc standardisation data

ENVR-v5.3.norm - normalisation data

ENVR-v5.3.phn - binary phones list

ENVR-v5.3.prior - probabilities of output layer for calculating phone classification

julius-dnn-output.txt - sample decoding output

julius-dnn.exe - executable modified Julius compiled for Windows

julius.jconf - julius configuration

mozilla.wav - sample wave file

README.md - this file

test.dbl - list of files to decode

wav_config - input file format for decoding (MFCC config data)

# Additional info

A detailed documentation can be obtained from Julius decoder home page

https://github.com/julius-speech/julius

File format accepted is 16kHz, RIFF WAV PCM 16bit, mono, signed integer.

Decoded file can be changed by editing list.dbl

The sample file is borrowed from Mozilla Open Source English DataSet:

https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/

# Licence

https://en.wikipedia.org/wiki/Mozilla_Public_License

# Contact

silesiaresearch at gmail dot com