Total Samples: 3350
Training Samples: 3182
Validation Samples: 168
Regions: 20 different Bengali-speaking regions (Dhaka, Chittagong, Sylhet, Rajshahi, Khulna, etc.)

Training Hyperparameters

Epochs: 1
Batch Size: 8
Gradient Accumulation: 2
Effective Batch Size: 16
Learning Rate: 0.0001
Optimizer: OptimizerNames.ADAMW_TORCH
LoRA Rank: 32
LoRA Alpha: 32
Target Modules: q_proj, v_proj, k_proj, o_proj, encoder_attn layers, fc1, fc2

Training Results

Training Time: 41.20 minutes
Final Training Loss: 0.3332
Speed: 1.29 samples/second
GPU: Tesla T4

Usage

Installation

pip install transformers librosa soundfile torch

Basic Usage

python

import torch
import librosa
import numpy as np
from transformers import WhisperForConditionalGeneration, WhisperProcessor

# Load model and processor
model = WhisperForConditionalGeneration.from_pretrained(
    "seyam2023/bangla-whisper-large-v3",
    torch_dtype=torch.float16,
    device_map="auto"
)
processor = WhisperProcessor.from_pretrained("seyam2023/bangla-whisper-large-v3")

# Load audio file
audio, _ = librosa.load("your_audio.wav", sr=16000, mono=True)
audio = audio / (np.max(np.abs(audio)) + 1e-8)

# Process and transcribe
inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(model.device, dtype=torch.float16)

with torch.no_grad():
    pred_ids = model.generate(
        input_features,
        max_new_tokens=128,
        num_beams=1,
        do_sample=False
    )

transcription = processor.batch_decode(pred_ids, skip_special_tokens=True)[0]
print(transcription)

Batch Processing

python

# Process multiple audio files
audio_files = ["file1.wav", "file2.wav", "file3.wav"]

for audio_file in audio_files:
    audio, _ = librosa.load(audio_file, sr=16000, mono=True)
    audio = audio / (np.max(np.abs(audio)) + 1e-8)
    
    inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")
    input_features = inputs.input_features.to(model.device, dtype=torch.float16)
    
    with torch.no_grad():
        pred_ids = model.generate(input_features, max_new_tokens=128)
    
    text = processor.batch_decode(pred_ids, skip_special_tokens=True)[0]
    print(f"{audio_file}: {text}")

Performance

This model has been trained on diverse Bengali audio data from multiple regional dialects, making it robust for:

Standard Bengali (Dhaka dialect)
Regional variations (Chittagong, Sylhet, Noakhali, etc.)
Various audio quality conditions
Different speaking styles and speeds

Limitations

Optimized for Bengali language only
Performance may vary with:
- Heavy background noise
- Very low-quality audio recordings
- Non-native Bengali speakers
- Mixed language speech (code-switching)

Training Infrastructure

Hardware: Tesla T4
Framework: PyTorch, Transformers, Unsloth
Precision: FP16 (Mixed Precision Training)

Citation

If you use this model, please cite:

bibtex

@misc{bangla-whisper-large-v3,
  author = {Touhidul Alam Seyam and Md Abtahee Kabir and Noore Tamanna Orny},
  title = {Bangla Whisper Large V3: Bengali Speech Recognition},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/seyam2023/bangla-whisper-large-v3}}
}

Team

Team Huntrix

Touhidul Alam Seyam
Md Abtahee Kabir
Noore Tamanna Orny

Acknowledgments

Base model: OpenAI Whisper Large V3
Training optimization: Unsloth
Dataset: Custom Bengali regional audio dataset

License

Apache 2.0

Contact

For questions or feedback, please open an issue in the model repository or contact Team Huntrix.

seyam2023/huntrix-whisper-large-v3

Bangla Whisper Large V3 - Bengali Speech Recognition Model

Model Description

Training Details

Training Data