Use kaldi trained model in mobvoihotwords

[shipleyxie]

I have trained a model in kaldi(mobvoihotwords), and i want use it in kaldidecoder(kaldidecoder-hark_3.1.0-openblas.tar.xz).
Following the instruction of kaldidecoder, I run the below cmd, but got some cnfused *ERROR*, Could you help me to solve this?

kaldidecoder  --filename-features-list=wav.scp --nnet-type=3 --filename-words=../tree/graph_online/words.txt --filename-align-lexicon=../tree/graph_online/phones/align_lexicon.int --filename-mdl=final.mdl --filename-fst=../tree/graph_online/HCLG.fst

The log messages are posted below;

[[ Configuration of UI-Registered options ]]
acoustic-scale = 0.1
add-pitch = false
beam = 16
beam-delta = 0.5
computation.debug = false
config = ''
debug-computation = false
delta = 0.000976562
determinize-lattice = true
enable-debug = false
extra-left-context-initial = 0
fbank-config = ''
feature-type = 'mfcc'
filename-align-lexicon = '../tree/graph_online/phones/align_lexicon.int'
filename-class-frame-counts = ''
filename-feature-transform = ''
filename-features-list = 'wav.scp'
filename-fst = '../tree/graph_online/HCLG.fst'
filename-mdl = 'final.mdl'
filename-nnet = ''
filename-phones = ''
filename-words = '../tree/graph_online/words.txt'
frame-subsampling-factor = 1
frames-per-chunk = 20
hash-ratio = 2
help = false
host-mfcnet = 'localhost'
host-result = 'localhost'
ivector-extraction-config = ''
ivector-silence-weighting.max-state-duration = -1
ivector-silence-weighting.silence-phones = ''
ivector-silence-weighting.silence-weight = 1
lattice-beam = 10
lm-name = ''
max-active = 2147483647
max-mem = 50000000
max-tasks = 4
mfcc-config = ''
min-active = 200
minimize = false
nbest_count = 1
nnet-type = 3
online-pitch-config = ''
optimization.allocate-from-other = true
optimization.allow-left-merge = true
optimization.allow-right-merge = true
optimization.backprop-in-place = true
optimization.consolidate-model-update = true
optimization.convert-addition = true
optimization.extend-matrices = true
optimization.initialize-undefined = true
optimization.max-deriv-time = 2147483647
optimization.max-deriv-time-relative = 2147483647
optimization.memory-compression-level = 1
optimization.min-deriv-time = -2147483648
optimization.move-sizing-commands = true
optimization.optimize = true
optimization.optimize-row-ops = true
optimization.propagate-in-place = true
optimization.remove-assignments = true
optimization.snip-row-ops = true
optimization.split-row-ops = true
phone-determinize = true
plp-config = ''
port-mfcnet = 5530
port-result = 10500
print-args = true
print-kaldi-license = false
print-openblas-license = false
print-openfst-license = false
progress-output-interval = 0
prune-interval = 25
splice = 3
verbose = 0
word-determinize = true

[ERROR] Expected token "<Nnet3>", got instead "<DIMENSION>".
terminate called after throwing an instance of 'std::runtime_error'
  what():
[1]    48813 abort (core dumped)  ~/project/kaldidecoder3/build/src/kaldidecoder  --nnet-type=3

• This topic was modified 3 years, 9 months ago by shipleyxie.

[Masayuki Takigahira]

Thank you for your inquiry.

[ERROR] Expected token "<Nnet3>", got instead "<DIMENSION>".

According to the content of the error message, it seems that a file that is not a model in nnet3 format was entered. First, please check the your recipe used for learning.
From the content of the error message, it looks like you’ve probably given a model in nnet1 format.
Is there a file called final.nnet other than final.mdl? If final.nnet exists, you should specify –nnet-type=1 instead of –nnet-type=3 because it is a learning recipe for nnet1.

For details, refer to the following items in HARK-Document.
https://www.hark.jp/document/hark-document-en/subsec-KaldiDecoder.html

Best regards,
HARK Support Team.

• This reply was modified 3 years, 9 months ago by Masayuki Takigahira.

[Author]

Posts