🚨 BRDialect 🚨

BanglaTalk: Towards Real-Time Speech Assistance for Bengali Regional Dialects

📝 Paper, 🖥️ Github

BRDialect - ASR system is trained on ten regional dialects of Bangladesh using the Ben10 dataset from Bengali.AI.

Load the BRDialect ASR System

Prerequisite

!pip install -U transformers
!pip install https://github.com/kpu/kenlm/archive/master.zip
!pip install pyctcdecode

Log in to HuggingFace

from huggingface_hub import login
login("TOKEN")

Load base model and BRDialect

## BRDialect
from huggingface_hub import hf_hub_download

kenlm_model_path = hf_hub_download(repo_id="Jakir057/BRDialect", filename="BRDialect/5gram_kenlm.arpa")
state_dict_path = hf_hub_download(repo_id="Jakir057/BRDialect", filename="BRDialect/wav2vec2_bangla_regional_dialect.pth")

from transformers import AutoProcessor, AutoModelForCTC, Wav2Vec2ProcessorWithLM
import torch
import numpy as np
import pyctcdecode
import librosa

base_model_id = "ai4bharat/indicwav2vec_v1_bengali"
processor = AutoProcessor.from_pretrained(base_model_id)
model = AutoModelForCTC.from_pretrained(base_model_id)
model.load_state_dict(torch.load(state_dict_path)["model"])

vocab_dict = processor.tokenizer.get_vocab()
sorted_vocab_dict = {k: v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])}
decoder = pyctcdecode.build_ctcdecoder(
    list(sorted_vocab_dict.keys()),
    str(kenlm_model_path)
)
processor_with_lm = Wav2Vec2ProcessorWithLM(
    feature_extractor=processor.feature_extractor,
    tokenizer=processor.tokenizer,
    decoder=decoder
)
model.freeze_feature_encoder()
model.eval()

Transcription Generation

sampling_rate = 16000
path = "AUDIO_PATH"
frame, sr = librosa.load(path, sr=sampling_rate, mono=True)

inputs = processor(
    frame,
    sampling_rate=sampling_rate,
    return_tensors="pt",
    padding=False
)

with torch.no_grad():
    logits = model(inputs.input_values.to("cpu")).logits

np_logits = logits.squeeze(0).cpu().numpy()
result = processor_with_lm.decode(np_logits, beam_width=256)
text = result.text
print(f"Transcription={text}")

Citation

@article{hasan2025banglatalk,
  title={BanglaTalk: Towards Real-Time Speech Assistance for Bengali Regional Dialects},
  author={Hasan, Jakir and Dipta, Shubhashis Roy},
  journal={arXiv preprint arXiv:2510.06188},
  year={2025}
}

@inproceedings{javed2022towards,
  title={Towards building asr systems for the next billion users},
  author={Javed, Tahir and Doddapaneni, Sumanth and Raman, Abhigyan and Bhogale, Kaushal Santosh and Ramesh, Gowtham and Kunchukuttan, Anoop and Kumar, Pratyush and Khapra, Mitesh M},
  booktitle={Proceedings of the aaai conference on artificial intelligence},
  volume={36},
  number={10},
  pages={10813--10821},
  year={2022}
}