Fine-Tune Whisper for Italian: Model Training Only (Datasets Provided)

We are seeking a machine learning expert experienced in speech recognition to fine-tune OpenAI’s Whisper model for Italian. We will provide all necessary Italian audio datasets—your role is to train and optimize the model so it delivers high accuracy on Italian transcriptions. Note: This project is strictly for model training. You do not need to handle front-end, WebAssembly integration, or deployment. We only require the trained model and associated training artifacts. Key Tasks & Responsibilities 1. Model Training & Optimization • Use our provided Italian audio datasets (hours of curated content). • Develop or adapt training scripts (e.g., in PyTorch) for Whisper fine-tuning. • Propose optimal hyperparameters, data augmentation methods, and training schedules. 2. Performance Evaluation • Measure accuracy using standard metrics (e.g., WER, CER). • Provide before/after comparisons if relevant. • Suggest further optimizations if the results fall below a certain threshold. 3. Deliverables • Fine-Tuned Model Weights (compatible with Whisper/OpenAI framework). • Training Pipeline Code (scripts or notebooks) to reproduce your results. • Evaluation Report detailing metrics and any key observations about performance. What We Provide • All necessary Italian audio datasets (fully prepared and licensed). • Any domain-specific details or guidelines on usage, if needed. • Budget and guidance on cloud GPU usage (if applicable). What You Bring • Proven experience fine-tuning or training ASR (Automatic Speech Recognition) models. • Familiarity with Whisper or similar transformer-based speech models. • Ability to communicate training progress, challenges, and solutions clearly. • (Optional) Knowledge of Italian or prior projects in non-English ASR is a plus—but not mandatory. Timeline & Budget • Timeline: Please propose a realistic start-to-finish duration (including model evaluation). • Budget: Open to fixed price or hourly. Indicate your preference and rates.