Model Overview
In order to prepare and experiment with the model, it's necessary to install NVIDIA NeMo Toolkit [1].
We advise installing it once you've installed the most recent version of PyTorch.
This model have been trained on NVIDIA GeForce RTX 2070:
Python 3.7.15
NumPy 1.21.6
PyTorch 1.21.1
NVIDIA NeMo 1.7.0
pip install nemo_toolkit['all']
Model Usage:
The model is accessible within the NeMo toolkit [1] and can serve as a pre-trained checkpoint for either making inferences or for fine-tuning on a different dataset.
How to Import
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.EncDecCTCModel.restore_from(restore_path="stt_kz_quartznet15x5.nemo")
How to Transcribe Single Audio File
We can get a sample audio to test the model:
wget https://asr-kz-example.s3.us-west-2.amazonaws.com/sample_kz.wav
Then this line of code is to transcribe the single audio:
asr_model.transcribe(['sample_kz.wav'])
How to Transcribe Multiple Audio Files
python transcribe_speech.py model_path=stt_kz_quartznet15x5.nemo audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"
If you have a manifest file about your audio files:
python transcribe_speech.py model_path=stt_kz_quartznet15x5.nemo dataset_manifest=manifest.json
Input and Output
This model can take input from mono-channel audio .WAV files with a sample rate of 16,000 KHz.
Then, this model gives you the spoken words in a text format for a given audio sample.
Model Architecture
QuartzNet 15x5 [2] is a Jasper-like network that uses separable convolutions and larger filter sizes. It has comparable accuracy to Jasper while having much fewer parameters. This particular model has 15 blocks each repeated 5 times.
Training and Dataset
The model was finetuned to Kazakh speech based on the pre-trained English Model for over several epochs.
Kazakh Speech Corpus 2 (KSC2) [3] is the first industrial-scale open-source Kazakh speech corpus.
In total, KSC2 contains around 1.2k hours of high-quality transcribed data comprising over 600k utterances.
Performance
The model achieved:
Average WER: 15.53%
through the applying of Greedy Decoding.
Limitations
Because the GPU has limited power, we used a lightweight model architecture for fine-tuning.
In general, this makes it faster for inference but might show less overall performance.
In addition, if the speech includes technical terms or dialect words the model hasn't learned, it may not work as well.
Demonstration
For quicker inference, you can test the model on Hugging Face Space here: NeMo_STT_KZ_Quartznet15x5
References
[2] QuartzNet 15x5
