This repository contains a simple recipe for training a hybrid DNN-HMM (Deep Neural Network - Hidden Markov Model) speech recognition model using Kaldi.
Follow the Kaldi installation guide to set up Kaldi. After installing Kaldi, replace run.sh with the one available in egs/librispeech recipe.
Before using this recipe, you need to prepare two metadata files: dataset/train.txt and dataset/test.txt. Each line in these files should contain a tab-separated list of metadata for a specific training example:
file_id utt_id spk_id start end path/to/file.wav transcriptionIt is assumed that each training file may contain multiple utterances. Therefore, start and end refer to the start and end times of each utterance, in seconds, respectively.
A lexicon is a dictionary that maps each word in the dataset to its corresponding phonetic transcription, detailing how each word is pronounced using smaller linguistic units called phonemes. Each line of dataset/lexicon.txt maps a word from your dataset to a space-separated list of its phonetic representation. The file lexicon.txt.example contains an example subset of a Persian lexicon.
word1 p1 p2 p3
word2 p5 p1 p2 p4
...A language model in .arpa format is required, which can be created using the KenLM toolkit.
The following steps are included in the training pipeline:
- Dataset preparation (stage 1-2)
- MFCC feature extraction (stage 3)
- Training a context-independent (monophone) HMM (stage 4)
- Aligning monophone states to training samples (stage 5)
- Training a context-dependent (triphone) HMM (stage 5)
- Extracting mel features (stage 6)
- Aligning triphone states to training samples (stage 7)
- Training a hybrid DNN-HMM model (stage 7)
Run certain stages of the recipe:
./run.sh 1 3 # stage=1 stop_stage=3The models were trained on a Persian dataset containing approximately 40 hours of speech data, and the trained language model was a bigram.
| Model | WER (%) |
|---|---|
| Context-independent HMM | 56.44 |
| Context-dependent HMM | 33.30 |
| Hybrid DNN-HMM | 26.56 |