Issue Tracking
Please raise any issues that you might have running the below scripts by opening an issue here. I will track this repo and reply to any open issues
Install kaldi
https://github.com/kaldi-asr/kaldi
Setup:
After installing kaldi, change the KALDI_ROOT in the path.sh to point to your local kaldi
Also, run the following commands
ln -sf <your_kaldi_folder>/egs/wsj/s5/steps .
ln -sf <your_kaldi_folder>/egs/wsj/s5/utils .
Data prep:
Create a folder, say, data/dsat. Create three files in the folder as follows:
wav.scp : Each line is of the form id audio_file_path
utt2spk: Each line is of the form id id
spk2utt: same as spk2utt. To create this file just run cp utt2spk spk2utt
For example, see data/test_data, a folder that I created
Download Pre-packaged Stuff
Download the folder from google drive https://drive.google.com/open?id=1w6PKskKrFNMfumkLDqvAul5fPjHDXHO5
inside this repo, run tar -xzvf exp.tar.gz, this will create a folder exp.
The exp folder contains all the necessary items used to run decoding and transcribe speech. Also, it contains the pre-trained Arabic ASR model
Before running the scripts below
Change the path.sh according to your local kaldi directory location
run source ./path.sh
Extract features:
- MFCC
Extract MFCC features using the command below:
Usage: steps/make_mfcc.sh [options] <data-dir> [<log-dir> [<mfcc-dir>] ]
e.g.: steps/make_mfcc.sh data/train exp/make_mfcc/train mfcc
Note: <log-dir> defaults to <data-dir>/log, and <mfccdir> defaults to <data-dir>/data
Options:
--mfcc-config <config-file> # config passed to compute-mfcc-feats
--nj <nj> # number of parallel jobs
--cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs.
--write-utt2num-frames <true|false> # If true, write utt2num_frames file
Concretely, I ran:
steps/make_mfcc.sh --cmd "run.pl" --nj 1 --mfcc-config conf/mfcc_hires.conf data/test_data exp/make_hires/test_data mfcc
Expected output on Command Line:
utils/validate_data_dir.sh: Successfully validated data-directory data/test_data
steps/make_mfcc.sh: [info]: no segments file exists: assuming wav.scp indexed by utterance.
Succeeded creating MFCC features for test_data
- Ivectors
Extract the i-vectors using the command below:
steps/online/nnet2/extract_ivectors_online.sh
Usage: steps/online/nnet2/extract_ivectors_online.sh [options] <data> <extractor-dir> <ivector-dir>
e.g.: steps/online/nnet2/extract_ivectors_online.sh data/train exp/nnet2_online/extractor exp/nnet2_online/ivectors_train
main options (for others, see top of script file)
--config <config-file> # config containing options
--cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs.
--nj <n|10> # Number of jobs
--stage <stage|0> # To control partial reruns
--num-gselect <n|5> # Number of Gaussians to select using
# diagonal model.
--min-post <float;default=0.025> # Pruning threshold for posteriors
--ivector-period <int;default=10> # How often to extract an iVector (frames)
Concretely, I ran:
steps/online/nnet2/extract_ivectors_online.sh --cmd “run.pl" --nj 1 data/test_data exp/nnet3/extractor exp/nnet3/ivectors_test_data
Expected command line output:
steps/online/nnet2/extract_ivectors_online.sh: extracting iVectors
steps/online/nnet2/extract_ivectors_online.sh: combining iVectors across jobs
steps/online/nnet2/extract_ivectors_online.sh: done extracting (online) iVectors to exp/nnet3/ivectors_test_data using the extractor in exp/nnet3/extractor
Decode:
Now, we run decoding:
steps/nnet3/decode.sh [options] <graph-dir> <data-dir> <decode-dir>
e.g.: steps/nnet3/decode.sh --nj 8 \
--online-ivector-dir exp/nnet2_online/ivectors_test_eval92 \
exp/tri4b/graph_bg data/test_eval92_hires /decode_bg_eval92
main options (for others, see top of script file)
--config <config-file> # config containing options
--nj <nj> # number of parallel jobs
--cmd <cmd> # Command to run in parallel with
--beam <beam> # Decoding beam; default 15.0
--iter <iter> # Iteration of model to decode; default is final.
--scoring-opts <string> # options to local/score.sh
--num-threads <n> # number of threads to use, default 1.
--use-gpu <true|false> # default: false. If true, we recommend
# to use large --num-threads as the graph
# search becomes the limiting factor.
Concretely,I ran the following:
steps/nnet3/decode.sh --nj 1 --cmd run.pl --acwt 1.0 --post-decode-acwt 10.0 --online-ivector-dir exp/nnet3/ivectors_test_data exp/mer80/chain/tdnn_7b/graph_tg data/test_data exp/mer80/chain/tdnn_7b/decode_test_data
Expected command line output:
steps/nnet2/check_ivectors_compatible.sh: WARNING: One of the directories do not contain iVector ID.
steps/nnet2/check_ivectors_compatible.sh: WARNING: That means it's you who's reponsible for keeping
steps/nnet2/check_ivectors_compatible.sh: WARNING: the directories compatible
steps/nnet3/decode.sh: feature type is raw
Transcribe:
lattice-scale --inv-acoustic-scale=8.0 "ark:gunzip -c exp/mer80/chain/tdnn_7b/decode_test_data/lat.*.gz|" ark:- | lattice-add-penalty --word-ins-penalty=0.0 ark:- ark:- | lattice-prune --beam=8 ark:- ark:- | lattice-mbr-decode --word-symbol-table=exp/mer80/chain/tdnn_7b/graph_tg/words.txt ark:- ark,t:- | utils/int2sym.pl -f 2- exp/mer80/chain/tdnn_7b/graph_tg/words.txt > transcript.txt
You should see transcribed speech output in the transcripts.txt