This directory contains scripts and tools for training speech recognition models using the Kaldi toolkit.

1. Overview
2. Directory Structure
3. Installation
4. Training Process
   • Data Preparation
   • Dictionary Preparation
   • MFCC Feature Extraction
   • Acoustic Model Training
   • TDNN Chain Model Training
   • Decoding
5. Results
6. Contributing

This repository provides tools for training custom speech recognition models using Kaldi. It supports acoustic model training, language model creation, and decoding pipelines.

.
├── cmd.sh                         # Command configuration for training and decoding
├── conf/
│   ├── mfcc.conf                  # Configuration for MFCC feature extraction
│   └── online_cmvn.conf           # Online Cepstral Mean Variance Normalization (currently empty)
├── local/
│   ├── chain/
│   │   ├── run_ivector_common.sh  # Script for i-vector extraction during chain model training
│   │   └── run_tdnn.sh            # Script for training a TDNN model
│   ├── data_prep.sh               # Data preparation script for creating Kaldi data directories
│   ├── download_and_untar.sh      # Script for downloading and extracting datasets
│   ├── download_lm.sh             # Downloads language models
│   ├── prepare_dict.sh            # Prepares the pronunciation dictionary
│   └── score.sh                   # Scoring script for evaluation
├── path.sh                        # Script for setting Kaldi paths
├── RESULTS                        # Script for printing the best WER results
├── RESULTS.txt                    # Contains WER results from decoding
├── run.sh                         # Main script for the entire training pipeline
├── steps -> ../../wsj/s5/steps/   # Link to Kaldi’s WSJ steps for acoustic model training
└── utils -> ../../wsj/s5/utils/   # Link to Kaldi’s utility scripts

• cmd.sh: Defines commands for running training and decoding tasks.
• path.sh: Sets up paths for Kaldi binaries and scripts.
• run.sh: Main entry point for the training pipeline, running tasks in stages.
• RESULTS: Displays Word Error Rate (WER) for the trained models.

• Kaldi: Kaldi toolkit must be installed and configured.
• Required tools: ffmpeg, sox, sctk for data preparation and scoring.

1. Clone the Vosk API repository.
2. Install Kaldi and ensure the KALDI_ROOT is correctly set in path.sh.
3. Set environment variables using cmd.sh and path.sh.

Run the data preparation stage in run.sh:

bash run.sh --stage 0 --stop_stage 0

This stage downloads and prepares the LibriSpeech dataset.

Prepare the pronunciation dictionary with:

bash run.sh --stage 1 --stop_stage 1

This step generates the necessary files for Kaldi's prepare_lang.sh script.

Run the MFCC extraction process:

bash run.sh --stage 2 --stop_stage 2

This step extracts Mel-frequency cepstral coefficients (MFCC) features and computes Cepstral Mean Variance Normalization (CMVN).

Train monophone, LDA+MLLT, and SAT models:

bash run.sh --stage 3 --stop_stage 3

This stage trains GMM-based models and aligns the data for TDNN training.

Train a Time-Delay Neural Network (TDNN) chain model:

bash run.sh --stage 4 --stop_stage 4

The chain model uses i-vectors for speaker adaptation.

After training, decode the test data:

bash run.sh --stage 5 --stop_stage 5

This step decodes using the trained model and evaluates the Word Error Rate (WER).

WER can be evaluated by running:

bash RESULTS

Example of RESULTS.txt:

%WER 14.10 [ 2839 / 20138, 214 ins, 487 del, 2138 sub ] exp/chain/tdnn/decode_test/wer_11_0.0
%WER 12.67 [ 2552 / 20138, 215 ins, 406 del, 1931 sub ] exp/chain/tdnn/decode_test_rescore/wer_11_0.0