whisper-medium-hi32

README

🗣️ Whisper-Medium-hi32: Fine-Tuned English ASR (Diverse Accents)

This repository hosts Whisper-Medium-hi32, a fine-tuned version of OpenAI's Whisper-Medium, developed by Marwan Kasem for automatic speech recognition (ASR) on clean, conversational English across diverse global accents.

---

📌 Overview

Whisper-Medium-hi32 is optimized for:

• 📞 Conversational English
• 🌍 Diverse Accents (UK, Irish, American, etc.)
• 🧪 Real-world scenarios (non-studio recordings)

This model is trained to better handle the variability of everyday speech, making it ideal for call center analytics, transcription services, and research in sociolinguistics or dialectal speech.

---

📊 Evaluation Results

These results were obtained on a clean, diverse English evaluation set featuring a wide range of accents.

---

🧠 Model Details

• Base Model: openai/whisper-medium
• Fine-Tuner: Marwan Kasem
• Frameworks: PyTorch + HuggingFace Transformers + PEFT

---

🚀 Intended Use

This ASR model is ideal for:

• ✅ Transcribing global English in interviews, meetings, podcasts
• ✅ Enhancing low-resource dialect ASR pipelines
• ✅ Evaluating robustness to accentual variation

---

⚠️ Limitations

• ⚠️ Trained on a subset of English (not multilingual)
• ⚠️ May underperform on noisy or overlapping speech
• ⚠️ Dataset details are currently under documentation

---

🏋️ Training Configuration

🧪 Training Results

---

🧰 Environment

• PEFT: v0.14.0
• Transformers: v4.47.0
• PyTorch: v2.5.1+cu121
• Datasets: v3.3.1
• Tokenizers: v0.21.0

---

💡 Usage Example (Python)

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torchaudio

processor = WhisperProcessor.from_pretrained("MarwanKasem/whisper-medium-hi32")
model = WhisperForConditionalGeneration.from_pretrained("MarwanKasem/whisper-medium-hi32")

# Load and preprocess audio
speech_array, sampling_rate = torchaudio.load("sample.wav")
inputs = processor(speech_array[0], sampling_rate=sampling_rate, return_tensors="pt")

# Run inference
generated_ids = model.generate(inputs.input_features)
transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(transcription[0])