Fatimaa75/whisper-base-bahraini

Whisper-Base Fine-tuned for Bahraini Arabic

Fine-tuned version of openai/whisper-base on the Bahraini Speech Dataset for Bahraini Arabic dialect transcription with Arabic-English code-switching support.

Developed as part of the Nota AI Meeting Assistant senior project at the University of Bahrain.

Model Details

Property	Value
Base model	openai/whisper-base
Parameters	72.6M
Language	Bahraini Arabic + English code-switching
Task	Automatic Speech Recognition
Format	Quantized ONNX (INT8)
Model size	74MB

Training Details

Hyperparameter	Value
Learning rate	3e-5
Batch size	32
Max steps	8,000
Warmup steps	500
Precision	bf16
GPU	NVIDIA A100 (40GB)
Training time	~2.5 hours

Dataset

• Name: Hishambarakat/Bahraini_Speech_Dataset
• Size: 69,224 training clips (~42 hours after filtering)
• Filtering: Removed clips shorter than 1.5 seconds
• Augmentation: Gaussian noise, time stretching, pitch shifting

Preprocessing

• Audio resampled from 24kHz to 16kHz
• Arabic diacritics removed
• Bahraini dialect spelling preserved (no MSA normalization)
• English words preserved as-is

Evaluation Results

Model	WER
whisper-base (no fine-tuning)	~88%
This model (V1, 6000 steps)	58.3%
This model (V2, 8000 steps + filtering + augmentation)	54.4%

Note: WER is measured against model-generated test labels, not human annotations. True WER against human ground truth is estimated to be 5-10 points lower.

Usage with Transformers.js (Browser)

import { pipeline } from '@huggingface/transformers';

const transcriber = await pipeline(
  'automatic-speech-recognition',
  'Fatimaa75/whisper-base-bahraini'
);

const result = await transcriber(audioData, {
  language: 'arabic',
  task: 'transcribe',
});

console.log(result.text);

Usage with Python

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

processor = WhisperProcessor.from_pretrained("Fatimaa75/whisper-base-bahraini")
model = WhisperForConditionalGeneration.from_pretrained("Fatimaa75/whisper-base-bahraini")

# Process audio (must be 16kHz)
inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    predicted_ids = model.generate(
        inputs.input_features,
        language="arabic",
        task="transcribe",
        no_repeat_ngram_size=3,
    )

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

Limitations

• Model capacity: whisper-base (74M params) has a WER ceiling for this dialect — whisper-small is recommended for production
• Short clips: Clips under 1.5 seconds remain unreliable
• Heavy dialect: Non-standard Bahraini words cause failures
• Code-switching: English word preservation is inconsistent
• Test labels: Evaluation uses model-generated labels, not human annotations

Intended Use

Designed for integration into Nota, a privacy-first AI meeting assistant. Best suited for:

• Meeting transcription (post-processing)
• Transcript search and summarization
• Bahraini Arabic speech with mixed English terminology

Not recommended for:

• Real-time live captioning where accuracy is critical
• Single-word or very short utterance recognition

Citation

If you use this model, please cite:

• University of Bahrain - ITCS 499 Senior Project 2025-2026
• Nota AI Meeting Assistant
• Fine-tuned Whisper-base for Bahraini Arabic ASR