Convert Kaldi* ASpIRE Chain Time Delay Neural Network (TDNN) Model to the Intermediate Representation

You can download a pre-trained model for the ASpIRE Chain Time Delay Neural Network (TDNN) from the Kaldi* project official web-site.

Convert ASpIRE Chain TDNN Model to IR

To generate the Intermediate Representation (IR) of the model, run the Model Optimizer with the following parameters:

python3 ./mo_kaldi.py --input_model exp/chain/tdnn_7b/final.mdl --output output

The IR will have two inputs: input for data and ivector for ivectors.

Example: Run ASpIRE Chain TDNN Model with the Speech Recognition Sample

These instructions show how to run the converted model with the Speech Recognition sample. In this example, the input data contains one utterance from one speaker.

To follow the steps described below, you must first do the following: 1. Download a Kaldi repository. 2. Build it using instructions in README.md in the repository. 3. Download the model archive from Kaldi website. 4. Extract the downloaded model archive to the egs/aspire/s5 folder of the Kaldi repository.

To run the ASpIRE Chain TDNN Model with Speech Recognition sample:

  1. Prepare the model for decoding. Refer to the README.txt file from the downloaded model archive for instructions.
  2. Convert data and ivectors to .ark format. Refer to the corresponding sections below for instructions.

Prepare Data

If you have a .wav data file, you can convert it to .ark format using the following command:

<path_to_kaldi_repo>/src/featbin/compute-mfcc-feats --config=<path_to_kaldi_repo>/egs/aspire/s5/conf/mfcc_hires.conf scp:./wav.scp ark,scp:feats.ark,feats.scp

Add the feats.ark absolute path to feats.scp to avoid errors in later commands.

Prepare Ivectors

To prepare ivectors for the Speech Recognition sample, do the following:

  1. Copy the feats.scp file to the egs/aspire/s5/ directory of the built Kaldi repository and navigate there:
cp feats.scp <path_to_kaldi_repo>/egs/aspire/s5/
cd <path_to_kaldi_repo>/egs/aspire/s5/
  1. Extract ivectors from the data:
./steps/online/nnet2/extract_ivectors_online.sh --nj 1 --ivector_period <max_frame_count_in_utterance> <data folder> exp/tdnn_7b_chain_online/ivector_extractor <ivector folder>

To simplify the preparation of ivectors for the Speech Recognition sample, specify the maximum number of frames in utterances as a parameter for --ivector_period to get only one ivector per utterance.

To get the maximum number of frames in utterances, you can use the following command line:

../../../src/featbin/feat-to-len scp:feats.scp ark,t: | cut -d' ' -f 2 - | sort -rn | head -1

As a result, in <ivector folder>, you will find the ivector_online.1.ark file.

  1. Go to the <ivector folder>:
cd <ivector folder>
  1. Convert the ivector_online.1.ark file to text format using the copy-feats tool. Run the following command:
<path_to_kaldi_repo>/src/featbin/copy-feats --binary=False ark:ivector_online.1.ark ark,t:ivector_online.1.ark.txt
  1. For the Speech Recognition sample, the .ark file must contain an ivector for each frame. You must copy the ivector frame_count times. To do this, you can run the following script in the Python* command prompt:
import subprocess

subprocess.run(["<path_to_kaldi_repo>/src/featbin/feat-to-len", "scp:<path_to_kaldi_repo>/egs/aspire/s5/feats.scp", "ark,t:feats_length.txt"])

f = open("ivector_online.1.ark.txt", "r")
g = open("ivector_online_ie.ark.txt", "w")
length_file = open("feats_length.txt", "r")
for line in f:
    if "[" not in line:
        for i in range(frame_count):
            line = line.replace("]", " ")
            g.write(line)
    else:
        g.write(line)
        frame_count = int(length_file.read().split(" ")[1])
g.write("]")
f.close()
g.close()
length_file.close()
  1. Create an .ark file from .txt:
<path_to_kaldi_repo>/src/featbin/copy-feats --binary=True ark,t:ivector_online_ie.ark.txt ark:ivector_online_ie.ark

Run the Speech Recognition Sample

Run the Speech Recognition sample with the created ivector .ark file as follows:

speech_sample -i feats.ark,ivector_online_ie.ark -m final.xml -d CPU -o prediction.ark -cw_l 17 -cw_r 12

Results can be decoded as described in "Use of Sample in Kaldi* Speech Recognition Pipeline" chapter in the Speech Recognition Sample description.