Summary
titu_stt_bn_fastconformer is a fastconformer based model trained on ~18K Hours MegaBNSpeech corpus.
Details on paper: https://aclanthology.org/2023.banglalp-1.16/
Using method
This model can be used for transcribing Bangla audio and also can be used as pre-trained model to fine-tuning on custom datasets using NeMo framework.
Installation
To install NeMo check NeMo documentation.
pip install -q 'nemo_toolkit[asr]'
Inferencing
Download test_bn_fastconformer.wav
# pip install -q 'nemo_toolkit[asr]'
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained("hishab/titu_stt_bn_fastconformer")
auido_file = "test_bn_fastconformer.wav"
transcriptions = asr_model.transcribe([auido_file])
print(transcriptions)
# ['আজ সরকারি ছুটির দিন দেশের সব শিক্ষা প্রতিষ্ঠান সহ সরকারি আধা সরকারি স্বায়ত্তশাসিত প্রতিষ্ঠান ও ভবনে জাতীয় পতাকা অর্ধনমিত ও কালো পতাকা উত্তোলন করা হয়েছে']
Colab Notebook for Infer: Bangla FastConformer Infer.ipynb
Training Datasets
| Channels Category | Hours |
|---|---|
| News | 17,640.00 |
| Talkshow | 688.82 |
| Vlog | 0.02 |
| Crime Show | 4.08 |
| Total | 18,332.92 |
Training Details
For training the model, the dataset we selected comprises 17.64k hours of news chan- nel content, 688.82 hours of talk shows, 0.02 hours of vlogs, and 4.08 hours of crime shows.
Evaluation
Citation
@inproceedings{nandi-etal-2023-pseudo,
title = "Pseudo-Labeling for Domain-Agnostic {B}angla Automatic Speech Recognition",
author = "Nandi, Rabindra Nath and
Menon, Mehadi and
Muntasir, Tareq and
Sarker, Sagor and
Muhtaseem, Quazi Sarwar and
Islam, Md. Tariqul and
Chowdhury, Shammur and
Alam, Firoj",
editor = "Alam, Firoj and
Kar, Sudipta and
Chowdhury, Shammur Absar and
Sadeque, Farig and
Amin, Ruhul",
booktitle = "Proceedings of the First Workshop on Bangla Language Processing (BLP-2023)",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.banglalp-1.16",
doi = "10.18653/v1/2023.banglalp-1.16",
pages = "152--162",
abstract = "One of the major challenges for developing automatic speech recognition (ASR) for low-resource languages is the limited access to labeled data with domain-specific variations. In this study, we propose a pseudo-labeling approach to develop a large-scale domain-agnostic ASR dataset. With the proposed methodology, we developed a 20k+ hours labeled Bangla speech dataset covering diverse topics, speaking styles, dialects, noisy environments, and conversational scenarios. We then exploited the developed corpus to design a conformer-based ASR system. We benchmarked the trained ASR with publicly available datasets and compared it with other available models. To investigate the efficacy, we designed and developed a human-annotated domain-agnostic test set composed of news, telephony, and conversational data among others. Our results demonstrate the efficacy of the model trained on psuedo-label data for the designed test-set along with publicly-available Bangla datasets. The experimental resources will be publicly available.https://github.com/hishab-nlp/Pseudo-Labeling-for-Domain-Agnostic-Bangla-ASR",
}