Common component configuration

In RASR many components appear again and again in different tools / context. The configuration parameters for these components are documented here:

Acoustic Model Configuration

Within this section, all common parameters which are part of the acoustic modeling are presented. These parameters are configured using the standard Configuration mechanism, e.g. a parameter concerning hidden markov models is configured [*.acoustic-model.hmm].

Allophone configuration

add-all (boolean) : All possible allophones w.r.t the phoneme list from the lexicon are generated and added.
add-from-lexicon (boolean) :  Only those allophones, which actually occur within the given lexicon are generated and added.
store-to-file (string) : Write all created allophones to the given file.
add-from-file (string) : Read allophones from the given file.

An example for a common allophone configuration would be:

[*.allophones]
add-all = false
add-from-lexicon = true

State-tying configuration

type (enum): Specify the type of state tying used (some settings have multiple names) :
none / no-tying : no state tying
monophone : no phoneme context
lut / lookup : use a [[state tying file formats#Lookup_Table|lookup table]]
cart / decision-tree : use CART in [[CART_file#New_CART_file_format|new file format]].
file (string): state tying definition file

HMM configuration

The configuration of the Hidden Markov Models is divided in two parts: a more general hmm part, where the structure of the HMM is defined and a tdp part, where the time distortion penalties for the individual transitions can be modified.

hmm

states-per-phone (int): We usually model each phone with three states, one for start, middle and end of the phone. The number of states per phone can be set with this parameter

state-repetitions (int): Each state of the phone may be repeated several times. With this parameter, the number of repetitions per state of the phone is set. Usually, the value 2 is chosen. With 3 states per phone, this results in the commonly used 6 state HMM.

across-word-model (bool): With this parameter, you can chose between across-word modeling (yes) or within-word modeling (no).

early-recombination (bool) : if you want to have a smaller state tree, you may chose yes for early recombination. But this may result in worse performance, so no is the recommended parameter choice.
tdp

scale (float): The global time distortion penalty scale for the score calculation is set with this parameter.

*.loop (float): The tdp for a loop transition between two phone states. For the silence phoneme, this penalty can be given separately via silence.loop. The penalty for the transition into the first state of a word can be set via the entry-m1 parameter. For compatibility reasons, there also is the parameter entry-m2, which sets the tdp for an incoming transition into the repetition of the first state of a word.

*.forward (float): The tdp for a forward transition between two phone states. For the silence phoneme, this penalty can be given separately via silence.forward.

*.skip (float): The tdp for a skip transition between two phone states. For the silence phoneme, this penalty can be given separately via silence.skip.

*.exit (float): For the end of a word, a special exit transition penalty may be given with this parameter. It controls the average length of the recognized words. For the silence phoneme, this penalty can be given separately via silence.exit.

An example HMM configuration may look like this:

[*.acoustic-model.hmm]
states-per-phone                = 3
state-repetitions               = 2
across-word-model               = yes
early-recombination             = no

[*.acoustic-model.tdp]
scale                                   = 1.0

*.loop                                  = 3.0
*.forward                               = 0.0
*.skip                                  = 3.0
*.exit                                  = 0.0

entry-m1.loop                           = infinity
entry-m2.loop                           = infinity

silence.loop                            = 0.0
silence.forward                         = 3.0
silence.skip                            = infinity
silence.exit                            = 6.0

Mixture Set

Common

feature-scorer-type (enum): There are several feature scorers available, depending on the used RWTH ASR version. Default is the SIMD-diagonal-maximum feature scorer, which is used for GMMs. For more details see below.
file (string): The location of the acoustic model is given with this parameter. See also Mixture File
scale (float): The acoustic model scale is set with this parameter. It is usually set to 1.
normalize-mixture-weights (bool): normalize mixture weights after split, not required in the Viterbi approximation (default False)

GMM

The default feature-scorer for GMM setups is SIMD-diagonal-maximum which uses a maximum approximation and a diagonal covariance matrix to calculate the scores. Additionally, the Single Instruction Multiple Data (SIMD) technique is used to achieve data level parallelism and thus improve the computing performance.

The additional parameters for this feature-scorer are:

covariance-tying (enum): Type of covariance matrix is used, it can be “pooled-covariance” (default) or “mixture-specific-covariance”.
minimum-variance (float): Floor variance value (it can be necessary if the number of mixture is very high).
reduced-mixture-set-dimension (int): Clip the mean and variance vectors to the given dimension.

An example configuration:

[*.mixture-set]
file                            = mixture.file
feature-scorer-type             = SIMD-diagonal-maximum
scale                           = 1.0
#covariance-tying                = pooled-covariance
#minimum-variance                = 0.01

Neural Networks

Neural network acoustic models are handled by the tf-fwd node in the flow file. The feature-scorer itself just needs to add the prior. The prefered featur-scorer for this is nn-precomputed-hybrid.

The additional parameters are:

prior-file (string) (optional): A prior specified in xml format. If no prior-file is given then the prior is computed from the mixture weigths.

priori-scale (float): (log-linear) scale to apply to the prior probabilites

[*.mixture-set]
feature-scorer-type       = nn-precomputed-hybrid
file                      = small.mix
normalize-mixture-weights = no
prior-file                = prior.xml
priori-scale              = 0.7
scale                     = 1.0

Channel Configuration

see Channel

Corpus Configuration

See also Bliss Corpus

audio-dir (string): Usually, the path entries within the corpus to the audio-files are relative to a certain path. The main directory to the audio files is given with this parameter.
capitalize-transcriptions (bool): It is possible to have the transcriptions mapped to only upper case letters, if you want to perform a case-insensitive recognition. Then, the value of this parameter should be yes, otherwise it should be no.
gemenize-transcriptions (bool): Convert transcription to lower case (overwrites capitalize-transcriptions).
file (string): The location of the BLISS Bliss Corpus file which should be recognized
partition (int) : Divide corpus into partitions with (approximately) equal number of segments.
recording-based-partition (bool) : create corpus partitions based on recordings instead of segments
segments.file (string) : include only segments in this file (# can be used to comment out segments to skip)
segment-order (string): file defining the order of processed segments (one segment identifier per line).
segment-order-look-up-short-name (bool): use short names in segment-order file (segment name only)
segments-to-skip (string list) : exclude segments in this list (space separated)
select-partition (int) : Select partition of the corpus
skip-first-segments (int) : skip the first N segments (counted after partitioning)
warn-about-unexpected-elements (bool): By default a warning is issued whenever an unknown XML element is ignored or flattened. Use this flag to turn this off.

Note: segment format in segments.file or segments-to-skip should contain the full-segment names, i.e. `` <corpusname>/<recordingname>/<segmentname> ``

Example:

[*.corpus]
file                            = /home/corpus/test.corpus.gz
audio-dir                       = /home/audio
warn-about-unexpected-elements  = no
capitalize-transcriptions       = no
#segments.file                  = segments.debug
#segments-to-skip               = coretex-de/tagesthemen/1 coretex-de/tagesthemen/2 coretex-de/tagesthemen/3
</pre>

Example for ``segments.debug`` file:
<pre>
coretex-de/tagesthemen/1
coretex-de/tagesthemen/2
coretex-de/tagesthemen/3
# coretex-de/tagesthemen/4

Lexicon Configuration

file (string): The location of the XML Bliss Lexicon to be used for recognition.

normalize-pronunciation (bool) : If there are no pronunciation weights given, a uniform distribution of the weights among all pronunciations of a lemma is enforced by setting this to true (default).

A typical lexicon configuration:

[*.lexicon]
file                            = $(LEXICON)
normalize-pronunciation         = false