Speech Acoustic Modeling Example

This folder contains examples for speech recognition.

lstm_proj.py: Functions for building a LSTM Network with/without projection layer.
io_util.py: Wrapper functions for DataIter over speech data.
train_lstm_proj.py: Script for training LSTM acoustic model.
decode_mxnet.py: Script for decoding LSTMP acoustic model.
default.cfg: Configuration for training on the AMI SDM1 dataset. Can be used as a template for writing other configuration files.
python_wrap: C wrappers for Kaldi C++ code, this is built into a .so. Python code that loads the .so and calls the C wrapper functions in io_func/feat_readers/reader_kaldi.py.

Connect to Kaldi:

decode_mxnet.sh: called by Kaldi to decode a acoustic model trained by mxnet (please select the simple method for decoding).

A full receipt:

run_ami.sh: a full receipt to train and decode acoustic model on AMI. It takes features and alignment from Kaldi to train an acoustic model and decode it.

To reproduce the results, use the following steps.

Build Kaldi

Build Kaldi as shared libraties if you have not already done so.

cd kaldi/src
./configure --shared # and other options that you need
make depend
make

Build Python Wrapper

Copy or link the attached python_wrap folder to kaldi/src.
Compile python_wrap/

cd kaldi/src/python_wrap/
make

Extract Features and Prepare Frame-level Labels

The acoustic models use Mel filter-bank or MFCC as input features. It also need to use Kaldi to do force-alignment to generate frame-level labels from the text transcriptions. For example, if you want to work on the AMI data SDM1. You can run kaldi/egs/ami/s5/run_sdm.sh. You will need to do some configuration of paths in kaldi/egs/ami/s5/cmd.sh and kaldi/egs/ami/s5/run_sdm.sh before you can run the examples. Please refer to Kaldi's document for more details.

The default run_sdm.sh script generates the force-alignment labels in their stage 7, and saves the force-aligned labels in exp/sdm1/tri3a_ali. The default script generates MFCC features (13-dimensional). You can try training with the MFCC features, or you can create Mel filter bank features by your self. For example, a script like this can be used to compute Mel filter bank features using Kaldi.

#!/bin/bash -u

. ./cmd.sh
. ./path.sh

# SDM - Signle Distant Microphone
micid=1 #which mic from array should be used?
mic=sdm$micid

# Set bash to 'debug' mode, it prints the commands (option '-x') and exits on :
# -e 'error', -u 'undefined variable', -o pipefail 'error in pipeline',
set -euxo pipefail

# Path where AMI gets downloaded (or where locally available):
AMI_DIR=$PWD/wav_db # Default,
data_dir=$PWD/data/$mic

# make filter bank data
for dset in train dev eval; do
  steps/make_fbank.sh --nj 48 --cmd "$train_cmd" $data_dir/$dset \
    $data_dir/$dset/log $data_dir/$dset/data-fbank
  steps/compute_cmvn_stats.sh $data_dir/$dset \
    $data_dir/$dset/log $data_dir/$dset/data

  apply-cmvn --utt2spk=ark:$data_dir/$dset/utt2spk \
    scp:$data_dir/$dset/cmvn.scp scp:$data_dir/$dset/feats.scp \
    ark,scp:$data_dir/$dset/feats-cmvn.ark,$data_dir/$dset/feats-cmvn.scp

  mv $data_dir/$dset/feats-cmvn.scp $data_dir/$dset/feats.scp
done

Here apply-cmvn was for mean-variance normalization. The default setup was applied per speaker. A more common was doing mean-variance normalization for the whole corpus and then feed to the neural networks:

 compute-cmvn-stats scp:data/sdm1/train_fbank/feats.scp data/sdm1/train_fbank/cmvn_g.ark
 apply-cmvn --norm-vars=true data/sdm1/train_fbank/cmvn_g.ark scp:data/sdm1/train_fbank/feats.scp ark,scp:data/sdm1/train_fbank_gcmvn/feats.ark,data/sdm1/train_fbank_gcmvn/feats.scp

Note that kaldi always try to find features in feats.scp. So make sure the normalized features organized as Kaldi way during decoding.

Finally, you need to put the features and labels together in a file so that MXNet can find them. More specifically, for each data set (train, dev, eval), you will need to create a file like train_mxnet.feats, will the following contents:

TRANSFORM scp:feat.scp
scp:label.scp

Here the TRANSFORM is the transformation you want to apply to the features. By default we use NO_FEATURE_TRANSFORM. The scp: syntax is from Kaldi. The feat.scp is typically the file from data/sdm1/train/feats.scp, and the label.scp is converted from the force-aligned labels located in exp/sdm1/tri3a_ali. Because the force-alignments are only generated on the training data, we split the training set into 90/10 parts, and use the 1/10 hold-out as the dev set (validation set). The script run_ami.sh will automatically do the splitting and format the file for MXNet. Please set the path in that script correctly before running. The run_ami.sh script will actually run the full pipeline including training the acoustic model and decoding. So you can skip the following steps if that scripts successfully runs.

Run MXNet Acoustic Model Training

Go back to this speech demo directory in MXNet. Make a copy of default.cfg and edit necessary items like the path to the dataset you just prepared.
Run python train_lstm.py --configfile=your-config.cfg. You can do python train_lstm.py --help to see the helps. All the configuration parameters can be set in default.cfg, customized config file, and through command line (e.g. --train_batch_size=50), and the latter values overwrite the former ones.

Here are some example outputs that we got from training on the TIMIT dataset.

Example output for TIMIT:
Summary of dataset ==================
bucket of len 100 : 3 samples
bucket of len 200 : 346 samples
bucket of len 300 : 1496 samples
bucket of len 400 : 974 samples
bucket of len 500 : 420 samples
bucket of len 600 : 90 samples
bucket of len 700 : 11 samples
bucket of len 800 : 2 samples
Summary of dataset ==================
bucket of len 100 : 0 samples
bucket of len 200 : 28 samples
bucket of len 300 : 169 samples
bucket of len 400 : 107 samples
bucket of len 500 : 41 samples
bucket of len 600 : 6 samples
bucket of len 700 : 3 samples
bucket of len 800 : 0 samples
2016-04-21 20:02:40,904 Epoch[0] Train-Acc_exlude_padding=0.154763
2016-04-21 20:02:40,904 Epoch[0] Time cost=91.574
2016-04-21 20:02:44,419 Epoch[0] Validation-Acc_exlude_padding=0.353552
2016-04-21 20:04:17,290 Epoch[1] Train-Acc_exlude_padding=0.447318
2016-04-21 20:04:17,290 Epoch[1] Time cost=92.870
2016-04-21 20:04:20,738 Epoch[1] Validation-Acc_exlude_padding=0.506458
2016-04-21 20:05:53,127 Epoch[2] Train-Acc_exlude_padding=0.557543
2016-04-21 20:05:53,128 Epoch[2] Time cost=92.390
2016-04-21 20:05:56,568 Epoch[2] Validation-Acc_exlude_padding=0.548100

The final frame accuracy was around 62%.

Run decode on the trained acoustic model

Estimate senone priors by run python make_stats.py --configfile=your-config.cfg | copy-feats ark:- ark:label_mean.ark (edit necessary items like the path to the training dataset). It will generate the label counts in label_mean.ark.
Link to necessary Kaldi decode setup e.g. local/ and utils/ and Run ./run_ami.sh --model prefix model --num_epoch num.

Here are the results on TIMIT and AMI test set (using all default setup, 3 layer LSTM with projection layers):

Corpus	WER
TIMIT	18.9
AMI	51.7 (42.2)

Note that for AMI 42.2 was evaluated non-overlapped speech. Kaldi-HMM baseline was 67.2% and DNN was 57.5%.

update Feb 07

We had updated this demo on Feb 07 (kaldi c747ed5, mxnet 912a7eb). We had also added timit demo script in this folder.

To run the timit demo:

cd path/to/kaldi/egs/timit/s5/
./run.sh (setup the kaild timit demo and run it)
ln -s path/to/mxnet/example/speech-demo/* path/to/kaldi/egs/timit/s5/
set ali_src, graph_src and so on in the run_timit.sh and default_timit.cfg to the generated folder in kaldi/egs/timit/s5/exp. In the demo script, we use tri3_ali as the alignment dir
set ydim (in default_timit.cfg) to kaldi/egs/timit/s5/exp/tri3/graph/num_pdfs + 1
./run_timit.sh