Speech Recognition Engines

Flat
Quickstart vocabulary limit
User: Tom J.
Date: 10/21/2020 11:38 pm
Views: 3524
Rating: 0

I'm using Julius Quickstart and it's fantastic, it's light and runs quickly on an SBC for an A.I. robotics project I'm working on.

The problem I'm having is that when I make a .voca file with more than 42 words it won't find a match on the second pass.

I'm assuming this is a time limit in the .jconf file (I'm just modifying all the sample files), but I can't seem to find it.

Any tips on overcoming this limitation so I can get a couple hundred words on it?

Thanks in advance, Tom

--- (Edited on 10/21/2020 11:38 pm [GMT-0500] by ) ---

Re: Quickstart vocabulary limit
User: kmaclean
Date: 10/22/2020 11:04 am
Views: 74
Rating: 0

>when I make a .voca file with more than 42 words it won't find a match on the second pass.

The Julius maintainer would be the best person to ask.

--- (Edited on 10/22/2020 12:04 pm [GMT-0400] by kmaclean) ---

Re: Quickstart vocabulary limit
User: Tom J.
Date: 10/26/2020 1:10 am
Views: 23
Rating: 0

I found a solution in a ten year old thread here and have been testing it, I currently have 257 words in my .voca, it didn't stop there, that's just all I've added so far.Performance hasn't suffered with the larger library, it's very fast and accuracy is not bad.

http://www.voxforge.org/home/forums/message-boards/speech-recognition-engines/adding-words-to-.voca-file---is-there-a-limit?layout=flat

I'm putting a couple things here to help the next person with this issue, this is running with Quick 2018.

----------------

Make sure sample.voca starts like this and then format the words in the same format as the stock sample.voca below of course, I changed my silence return strings to silin and silout because it's easier than the <s> </s> on my filter program:

% NS_B
silin        sil

% NS_E
silout        sil

% WORD

-----------------

sample.grammar:

S: NS_B WORD_LOOP NS_E
WORD_LOOP: WORD_LOOP WORD
WORD_LOOP: WORD

-----------------

Sample.jconf:

#
# Sample Jconf configuration file
# for Julius library rev.4.3
#
# 1) Options can also be specified in command line option.
#    The values are default values in Julius.
# 2) For file name, relative path must be relative to THIS FILE.
# 3) Texts after '#' at each line will be ignored.  If you want to specify
#    '#', use `\#'.
# 4) Each line should be no longer than 512 bytes.
#
#

######################################################################
#### JULIUS APPLICATION OPTION (not a part of JuliusLib)
######################################################################
#-outfile            # save each result in separate file
#-separatescore            # print AM / LM scores separately
#-callbackdebug            # print callback names at call for debug
#-charconv from to        # print with character set conversion
#-nocharconv            # disable "-charconv"
#-module            # start in module mode
#-record dir            # record each inputs into dir
#-logfile file            # redirect logs to file
#-nolog                # disable all logs
#-help                # print help, and exit


######################################################################
#### GLOBAL OPTIONS
######################################################################
####
#### Misc. Options
####
#-C jconffile            # include jconf file at here
#-version            # print version info to stderr, and exit
#-setting            # print version info to stderr, and exit
#-quiet                # output less log
#-debug                # output more log for debug
#-check wchmm            # for debug, enter debug shell mode
#-check trellis            # for debug, enter debug shell mode
#-check triphone        # for debug, enter debug shell mode
#-demo                # same as "-quiet -progout"

####
#### Stream Input
####

## Feature vector input
#-input mfcfile            # feature vector in HTK parameter file
#-input htkparam        # (same as "mfcfile")
#-input outprob            # outprob vector in HTK parameter file
#-input vecet            # feature / outprob vector via network client

## Raw audio input
#-input mic                # live microphone
#-input rawfile            # wavefile
#-input file            # (same as "rawfile")
#-input stdin            # waveform from standard input
#-input adinnet            # waveform via network client
#-input netaudio                # DatLink server
#-input oss            # OSS API input (if available)
#-input alsa            # ALSA API input (if available)
#-input esd            # ESounD daemon input (if available)
#-input portaudio        # PortAudio API
#-input pulseaudio        # PulseAudio API


#-filelist filename        # input file list
#-notypecheck            # does not check parameter type of input
#-48                # 48kHz sampling > 16kHz conv. (16kHz only)
#-NA devname            # hostname for DatLink server
#-adport 5530            # port number for adinnet
#-nostrip            # do not strip zero samples
#-zmean                # remove DC offset (use long input average)
#-nozmean            # disable "-zmean" specified before

####
#### Audio Input Scaling
####
#-lvscale 1.0            # input level scaling factor (1.0 = disable)

####
#### Speech segment detection by level and zero-cross
####
#### default: on for microphone, off for other sources
####
#-cutsilence            # detection on
#-nocutsilence            # detection off
-lv 4000            # level threshold (0-32767)
#-zc 60                # zero-cross threshold (times in sec.)
#-headmargin 300        # head silence margin (msec)
#-tailmargin 400        # tail silence margin (msec)
#-chunk_size 1000        # processing segment unit length in samples
#-rejectshort 0            # reject shorter input (msec)
#-rejectlong -1            # reject longer input (msec) -1 to disable

####
#### Input rejection by average power (EXPERIMENTAL)
####
####  This will be enabled by "--enable-power-reject" on compilation.
####  Should be used with Decoder VAD or GMM VAD.
####  Valid for real-time input only.
####
#-powerthres 9.0        # reject input by avg. energy

####
#### Gaussian Mixture Model
####
####  GMM will be used for input rejection by accumurated score, or
####  for GMM-based frontend VAD when "--enable-gmm-vad" specified.
####
####  NOTE: If you use MFCC for the GMM which is different from AM, you
####  should also set the parameters like other AM with an option "-AM_GMM".
####  If "-AM_GMM" is not used, Julius assume GMM to use the same
####  parameter as the first AM.
####
#-gmm hmmdefs            # GMM definitions in HTK format
#-gmmnum 10            # num of Gaussians to be computed per GMM
#-gmmreject string        # comma-separated list of GMM name to reject
#### GMM_VAD
#-gmmmargin 20            # head margin for GMM based VAD in frames

####
#### Decoding option
####
#### Real-time processing means concurrent processing of MFCC computation
#### and 1st pass decoding.  By default, real-time processing on the
#### 1st pass is on for microphone / adinnet / netaudio input, and
#### off for others.
####
#-realtime            # force real-time processing
#-norealtime            # force non real-time processing

####
#### Plug-in
####
#### See plugin/00readme.txt for detail
####
#-plugindir ./plugin:/usr/local/share/julius/plugins


######################################################################
#### INSTANCE DEFINITION FOR MULTI DECODING
######################################################################
####
#### The following three argument will create a new configuration set
#### with default parameters, and switch current set to it.  Jconf
#### parameters specified after the option will be set into the current
#### set.
####
#### To do multi-model decoding, these argument should be specified at
#### the first of each model / search instances with different names.
#### Any options before the first instance definition will be IGNORED.
####
#### When no instance definition is found (as older version of Julius),
#### all the options are assigned to a default instance named "_default".
####
#### Please note that decoding with a single LM and multiple AMs is not
#### fully supported.  For example, you may want to construct the
#### jconf file as this:
####
####  -AM am_1 -AM am_2
####  -LM lm (LM spec..)
####  -SR search1 am_1 lm
####  -SR search2 am_2 lm
####
#### This type of model sharing is not supported yet, since some part
#### of LM processing depends on the assigned AM.  Instead, you can
#### get the same result by defining the same LMs for each AM, like this:
####
####  -AM am_1 -AM am_2
####  -LM lm_1 (LM spec..)
####  -LM lm_2 (same LM spec..)
####  -SR search1 am_1 lm_1
####  -SR search2 am_2 lm_2
####

## Create a new AM configuration set, and switch current to it.
## You should give a unique name.
#-AM name           

## Create a new LM configuration set, and switch current to it.
## You should give a unique name.
#-LM name           

## Create a new Search configuration set with AM and LM, and switch
## current to it.  AM and LM name can be either name or ID number.
#-SR name am_name_or_id lm_name_or_id

## Switch current AM to special one reserved for GMM, to specify
## analysis parameter for GMM.  Be sure not to confuse with normal AM
## configuration.
# -AM_GMM

## When using instance declarations, global options should be placed
## at top before any instance declaration, or after this option below.
## This option is only a switcher and can be used anywhere anytime.
# -GLOBAL

## This option disables the strict section checkings and back to 4.0
# -nosectioncheck

######################################################################
#### LANGUAGE MODEL (-LM)
######################################################################
####
#### Only one type of LM can be specified for a LM configuration.
####

####
#### N-gram
####
#-d binary_ngram_file        # N-gram in Julius binary format
#-nlr ngram            # forward (left-to-right) N-gram
#-nrl rev_ngram            # backward (right-to-left) N-gram
#-v dictfile            # word dictionary
## param.
#-silhead "<s>"            # beginning-of-sentence (silence) word
#-siltail "</s>"        # end-of-sentence (silence) word
#-mapunk "<unk>"        # word to which unknown words should be mapped
#-iwspword            # add a pause word to the dictionary
#-iwspentry "<UNK> [sp] sp sp"    # word that will be added by "-iwspword"
#-sepnum 150            # num of high freq words to linearize
#-adddict dictfile              # append additional word dictionary
#-addword entry                 # append additional word entry

####
#### Grammar
####
#### "-gram", "-gramlist" can be used multiple times
#-gram gramprefix        # (comma-separated list of) grammar file prefix
#-gramlist txtfile        # text file containing grammar prefixes
#-dfa dfafile -v dictfile    # specify DFA and dictionary separately
#-nogram            # reset all grammar list already specified

####
#### Isolated Word
####
#-w dictfile            # word dictionary
#-wlist txtfile            # text file containing dictionaries
#-nogram            # reset all dictfiles already specified
## param.
#-wsil silB silE NULL        # head / tail silence models to be appended

####
#### User-defined LM
####
#-userlm            # declare to use user LM defined in program

####
#### misc LM option
####
#-forcedict            # skip error word entries (no stop on error)

######################################################################
#### ACOUSTIC MODEL (-AM)  (-AM_GMM)
######################################################################
####
#### Acoustic analysis parameters are included in this section, since
#### the AM defines the required parameter.  You can use different MFCC
#### type for each AM.
#### For GMM, the same parameter should be specified after "-AM_GMM"
####
#### When using multiple AM,  the values of "-smpPeriod", "-smpFreq",
#### "-fsize" and "-fshift" should be the same among all AM.
####

## Acoustic model
#-h hmmfile            # acoustic HMM (ascii or Julius binary)
#-hlist logicaltri        # HMMList to map logical phone to physical
#-tmix 2            # # of mixture to compute in a mixture PDF
#-spmodel "sp"            # name of a short-pause silence model
#-multipath            # force enable MULTI-PATH model handling
#-gprune {safe|heuristic|beam|none|default} # Gaussian pruning method
-iwcd1 max
#-iwcd1 {max|avg|best 3}    # Inter-word triphone approximation method
#-iwsppenalty -1.0        # pause insertion penalty for "-iwsp"
#-gshmm hmmfile         # HMM for Gaussian mixture selection
#-gsnum 24            # Threshold number of HMM for gshmm

## Analysis
#-smpPeriod 625            # sampling period (ns) (= 10000000 / smpFreq)
#-smpFreq 16000            # sampling rate (Hz)
#-fsize 400            # window size (samples)
#-fshift 160            # frame shift (samples)
#-preemph 0.97            # pre-emphasis coef.
#-fbank 24            # number of filterbank channels
#-ceplif 22            # cepstral liftering coef.
#-rawe                # use raw energy
#-norawe            # disable "-rawe" (this is default)
#-enormal            # normalize log energy
#-noenormal            # disable "-enormal" (this is default)
#-escale 1.0            # scaling log energy for enormal
#-silfloor 50.0            # energy silence floor in dB for enormal
#-delwin 2            # delta window (frames)
#-accwin 2            # acceleration window (frames)
#-hifreq -1            # cut-off hi frequency (Hz) (-1: disable)
#-lofreq -1            # cut-off low frequency (Hz) (-1: disable)
#-zmeanframe            # frame-wise DC offset removal (same as HTK)
#-nozmeanframe            # disable "-zmeanframe" (this is default)

## Cepstral mean / variance normalization
#-cmnload filename        # load initial cep. mean / variance on startup
#-cmnsave filename        # save cep. mean / variance at each input end
#-cmnupdate            # update beginning cep. data at each input
#-cmnnoupdate            # keep initial mean, disable "-cmnupdate"
#-cmnmapweight 100.0        # weight for MAP-CMN
#-cvn                # enable variance normalization

## Vocal tract length normalization (VTLN)
#-vtln 1.0 300 4800        # enable VTLN (alpha, lowerfreq, upperfreq)

## Spectral subtraction (default: disabled)
#-sscalc            # do SS, estimate noise from head sil
#-sscalclen 300            # length of head silence for "-sscalc" (msec)
#-ssload filename        # do SS, load noise spectrum saved by "mkss"
#-ssalpha 2.0            # alpha coef. for spectral subtraction
#-ssfloor 0.5            # spectral floor coef.

## Others
#-htkconf configfile        # load analysis settings from HTK Config file

######################################################################
#### RECOGNIZER (-SR)
######################################################################
####
#### Default values for beam width and LM weights will change
#### according to compile-time setup of JuliusLib and model specification.
#### Please see the startup log for the actual values.
####

####
#### parameter (common)
####
#-inactive            # start this process with inactive status
#-1pass                # perform only the 1st pass, omit 2nd pass
#-no_ccd            # switch off the phone context dependency
#-force_ccd            # force on the phone context dependency
#-cmalpha 0.05            # CM alpha value
#-iwsp                # append a skippable sp at all word ends
#-transp 0.0            # transition penalty for transparent words

####
#### parameter (1st pass)
####
#-lmp weight penalty        # LM weight and word insertion penalty (pass1)
#-penalty1 penalty        # word insertion penalty for grammar (pass1)
-b                         # beam width (# of nodes)
#-bs score                      # beam width (score)
#-nlimit 3            # with enable-wpair-nlimit, set max N at nodes
#-progout            # progressive output while decoding
#-proginterval 300        # output interval in msec for "-progout"

####
#### parameter (2nd pass)
####
#-lmp2 weight penalty        # LM weight and word insertion penalty (pass2)
#-penalty2 penalty        # word insertion penalty for grammar (pass2)
-b2            # envelope beam width of 2nd pass (#word)
#-sb 80.0            # envelope score width at 2nd pass
#-s 500                # hypotheses stack size on 2nd pass (#hypo)
#-m 2000            # hypotheses overflow threshold (#hypo)
#-n n                # num of sentences to find
#-output 1            # num of sentences to output as result
#-lookuprange 5            # hypo. lookup range at word expansion (#frame)
#-looktrellis            # expand only trellis words in grammar
#-fallback1pass            # output 1st pass result when 2nd pass fails

####
#### short-pause segmentation (and decoder-based VAD)
####
#-spsegment            # enable sp segmentation (or decoder VAD)
#-spdur 10            # # of frames to detect a short pause
#-pausemodels string        # comma-separated pause model names
#### for decoder-VAD
#-spmargin 40            # backstep margin at trigger up (frame)
#-spdelay 4            # decision delay at trigger up (frame)

####
#### lattice output
####
#-lattice            # output result in word graph (aka -graphout)
#-graphrange 0            # merge same words nearby, -1 to disable merge
#-graphcut 80            # graph depth cut threshold (in depth)
#-graphboundloop 20        # max itertations for boundary adjustment loop
#-graphsearchdelay        # activate an alternate generation algorithm
#-nographsearchdelay        # disable "-graphsearchdelay"

####
#### confusion network output
####
#-confnet            # enable confusion network output
#-noconfnet            # disable confusion network output

####
#### multi-grammar output (for grammar and isolated word)
####
#-multigramout            # output max hypo for each grammar
#-nomultigramout        # disable "-multigramout"

####
#### forced alignment
####
#-walign            # enable alignment for result at word level
#-palign            # enable alignment for result at phoneme level
#-salign            # enable alignment for result at state level

####
#### misc.
####
# !!!!!! VoxForge change
#-outprobout filename        # save computed outprob vectors to HTK file (for debug)
# !!!!!!

# VoxForge configurations:
-dfa grammar/sample.dfa
-v grammar/sample.dict
-h acoustic_model_files/hmmdefs
-hlist acoustic_model_files/tiedlist
-spmodel "sp"        # HMM model name
-multipath
-gprune safe
-iwcd1 max
-iwsppenalty -10.0    # transition penalty for the appended sp models
-smpFreq 16000        # sampling rate (Hz)
-iwsp            # append a skippable sp model at all word ends
-penalty1 5.0       
-penalty2 20.0   
-b2 200                 # beam width on 2nd pass (#words)
-sb 200.0        # score beam envelope threshold
-n 1
# !!!!!!

--- (Edited on 10/26/2020 1:10 am [GMT-0500] by Tom J.) ---

Re: Quickstart vocabulary limit
User: kmaclean
Date: 10/26/2020 11:52 am
Views: 318
Rating: 0

Thanks for the update!

--- (Edited on 10/26/2020 12:52 pm [GMT-0400] by kmaclean) ---

Re: Quickstart vocabulary limit
User: Tom J.
Date: 1/10/2021 11:56 pm
Views: 1280
Rating: 0

Just a quick update, it took months but I assembled a voca file of 31,174 words, Julius is running on a 1.5 ghz ARMv7 microcomputer and returns matches in a second or two despite having such a large dictionary of words.

I've got some more fish to fry first but I'm considering pulling the whole folder into a ramdrive at boot to see if it can return results even faster, I have 2Gigs of ram so a 50 mb ramdrive wouldn't be a tremendous burden.

Oh, I've posted in another part of the forum but I'm very interested in studying any projects written in C++ that utilize Juliuslib api.

--- (Edited on 1/10/2021 11:56 pm [GMT-0600] by Tom J.) ---

PreviousNext