Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[scripts] Nnet3 segmentation: fix default params #3051

Merged
merged 2 commits into from
Feb 26, 2019
Merged

[scripts] Nnet3 segmentation: fix default params #3051

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
19355bf
[scripts] Add missing option as default in nnet3 model-cleanup script
danpovey Feb 24, 2019
974fa81
[scripts] Make sure nnet3 segmentation opts set right for chain systems
danpovey Feb 24, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 30 additions & 13 deletions egs/wsj/s5/steps/cleanup/segment_long_utterances_nnet3.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,8 @@
# 2016 Vimal Manohar
# Apache 2.0

# This script is similar to steps/cleanup/segment_long_utterances.sh, but

# This script is similar to steps/cleanup/segment_long_utterances.sh, but
# uses nnet3 acoustic model instead of GMM acoustic model for decoding.
# This script performs segmentation of the input data based on the transcription
# and outputs segmented data along with the corresponding aligned transcription.
Expand All @@ -13,7 +14,7 @@
# are of manageable length for further processing, along with the portion of the
# transcript that seems to match (aligns with) each segment.
# This the light-supervised training scenario where the input transcription is
# not expected to be completely clean and may have significant errors.
# not expected to be completely clean and may have significant errors.
# See "JHU Kaldi System for Arabic MGB-3 ASR Challenge using Diarization,
# Audio-transcript Alignment and Transfer Learning": Vimal Manohar, Daniel
# Povey, Sanjeev Khudanpur, ASRU 2017
Expand All @@ -39,24 +40,22 @@ seconds_per_spk_max=30

# Decode options
graph_opts=
scale_opts= # for making the graphs
beam=15.0
lattice_beam=1.0
lmwt=10

acwt=0.1 # Just a default value, used for adaptation and beam-pruning..
post_decode_acwt=1.0 # can be used in 'chain' systems to scale acoustics by 10 so the
# regular scoring script works.

# Contexts must ideally match training
extra_left_context=0 # Set to some large value, typically 40 for LSTM (must match training)
extra_right_context=0
extra_right_context=0
extra_left_context_initial=-1
extra_right_context_final=-1
frames_per_chunk=150

# i-vector options
extractor= # i-Vector extractor. If provided, will extract i-vectors.
# Required if the network was trained with i-vector extractor.
extractor= # i-Vector extractor. If provided, will extract i-vectors.
# Required if the network was trained with i-vector extractor.
use_vad=false # Use energy-based VAD for i-vector extraction

# TF-IDF similarity search options
Expand Down Expand Up @@ -116,12 +115,12 @@ it and eliminate data where the transcript doesn't seem to match.
--segmentation-extra-opts 'opts' # Additional options to segment_ctm_edits_mild.py.
# Please run steps/cleanup/internal/segment_ctm_edits_mild.py
# without arguments to see allowed options.
--align-full-hyp <true|false> # If true, align full hypothesis
i.e. trackback from the end to get the alignment.
This is different from the normal
--align-full-hyp <true|false> # If true, align full hypothesis
i.e. trackback from the end to get the alignment.
This is different from the normal
Smith-Waterman alignment, where the
traceback will be from the maximum score.
--extractor <extractor> # i-vector extractor directory if i-vector is
--extractor <extractor> # i-vector extractor directory if i-vector is
# to be used during decoding. Must match
# the extractor used for training neural-network.
--use-vad <true|false> # If true, uses energy-based VAD to apply frame weights
Expand Down Expand Up @@ -168,6 +167,23 @@ cp $srcdir/cmvn_opts $dir
cp $srcdir/{splice_opts,delta_opts,final.mat,final.alimdl} $dir 2>/dev/null || true
cp $srcdir/frame_subsampling_factor $dir 2>/dev/null || true

if [ -f $srcdir/frame_subsampling_factor ]; then
echo "$0: guessing that this is a chain system, checking parameters."
if [ -z $scale_opts ]; then
echo "$0: setting scale_opts"
scale_opts="--self-loop-scale=1.0 --transition-scale=1.0"
fi
if [ $acwt == 0.1 ]; then
echo "$0: setting acwt=1.0"
acwt=1.0
fi
if [ $lmwt == 10 ]; then
echo "$0: setting lmwt=1.0"
lmwt=1
fi
fi


utils/lang/check_phones_compatible.sh $lang/phones.txt $srcdir/phones.txt
cp $lang/phones.txt $dir

Expand Down Expand Up @@ -221,6 +237,7 @@ if [ $stage -le 3 ]; then

# Make graphs w.r.t. to the original text (usually recording-level)
steps/cleanup/make_biased_lm_graphs.sh $graph_opts \
--scale-opts "$scale_opts" \
--nj $nj --cmd "$cmd" $text \
$lang $dir $dir/graphs
if [ -z "$utt2text" ]; then
Expand Down Expand Up @@ -267,7 +284,7 @@ if [ $stage -le 5 ]; then
echo "$0: Decoding with biased language models..."

steps/cleanup/decode_segmentation_nnet3.sh \
--acwt $acwt --post-decode-acwt $post_decode_acwt \
--acwt $acwt \
--beam $beam --lattice-beam $lattice_beam --nj $nj --cmd "$cmd --mem 4G" \
--skip-scoring true --allow-partial false \
--extra-left-context $extra_left_context \
Expand Down