model.props

来自「It is the Speech recognition software. 」· PROPS 代码 · 共 56 行

# Copyright 1999-2004 Carnegie Mellon University.  
# Portions Copyright 2004 Sun Microsystems, Inc.  
# Portions Copyright 2004 Mitsubishi Electronic Research Laboratories.
# All Rights Reserved.  Use is subject to license terms.
# 
# See the file "license.terms" for information on usage and
# redistribution of this file, and for a DISCLAIMER OF ALL 
# WARRANTIES.
#
# Trained using entire channel 1 training data from WSJ (~32K utterances).  
# All the training data was downsampled to 8k:
#
#   sph2pipe -f raw ldc/$foo.wv1 > raw/$foo.raw
#   sox -t sw -r 16000 raw/$foo.raw -t sw -r 8000 raw_8k/$foo.raw
#
# The feature files were created as follows:
#
#   wave2feat \
#        -i raw_8k/$foo.raw -raw \
#        -srate 8000    \
#        -wlen 0.025625  \
#        -alpha 0.97     \
#        -nfft 512       \
#        -nfilt 31       \
#        -lowerf 200     \
#        -upperf 3500    \
#        -ncep 13        \
#        -o mfc_8k/$foo.mfc
#
# Training was done as follows using SphinxTrain:
#
#    1) First train full models using 15 iterations per step.
#    2) Force align the training data against models from step #1.
#    3) Use aligned transcripts from step #2 to train new models;
#       set convergence ratio to 0.02.  This resulted in around
#       5-7 iterations per step.
#
description = Wall Street Journal 8kHz acoustic models
modelClass = edu.cmu.sphinx.model.acoustic.WSJ_8gau_13dCep_8kHz_31mel_200Hz_3500Hz.Model
modelLoader = edu.cmu.sphinx.model.acoustic.WSJ_8gau_13dCep_8kHz_31mel_200Hz_3500Hz.ModelLoader

isBinary = true
featureType = 1s_c_d_dd
vectorLength = 39
sparseForm = false

numberFftPoints = 512
numberFilters = 31
gaussians = 8
minimumFrequency = 200
maximumFrequency = 3500
sampleRate = 8000

dataLocation = cd_continuous_8gau
modelDefinition = etc/WSJ_8gau_13dCep_8kHz_31mel_200Hz_3500Hz.4000.mdef