Whisper Small Fine Tuned

This model is a fine-tuned version of openai/whisper-small on the 500 samples of doctor-patient conversations. It achieves the following results on the evaluation set:

• Loss: 0.4537
• Wer: 10.6667

Intended uses & limitations

Model was trained for Educational Purposes

How to Use

You can use this fine-tuned Whisper model for transcribing short medical conversations using Hugging Face's transformers and datasets libraries.

from transformers import pipeline

# Load the fine-tuned model from Hugging Face
pipe = pipeline(
    task="automatic-speech-recognition",
    model="Bhaveen/Medical-Speech-Transcription-Whisper-Small-Fine-Tuned",
    device=0  # set to -1 for CPU, or 0 for GPU if available
)

# Run inference on an audio file
result = pipe("path_to_audio.wav")  # Replace with your actual audio file path
print(result["text"])

Requirements

Install dependencies if you haven't already:

pip install transformers datasets torchaudio

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-06
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• num_epochs: 8
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.47.1
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0