Voice Data Classification Model
Budget: ₹600 – ₹1,500 INR
I need a supervised machine-learning pipeline that takes raw voice recordings as input and automatically assigns each file to the correct category. In other words, this is a pure classification problem focused on categorizing data that arrives in audio form.
Scope of work
• End-to-end data flow: ingest the recordings, extract meaningful acoustic features (MFCCs, spectrograms or any feature set you recommend) and build a robust training/validation split.
• Model development: you are free to choose your preferred stack—TensorFlow, PyTorch, scikit-learn, or an ensemble of architectures—as long as the final model hits the agreed accuracy on a held-out test set.
• Experiment tracking & reproducibility: include clear notebooks or scripts plus a brief report summarizing metrics, confusion matrix, and the hyper-parameter search process.
• Deployment readiness: package the trained model and preprocessing steps so I can load it with a single call (e.g., a saved .pt/.h5 file with a companion inference script or a lightweight REST endpoint).
Acceptance criteria
1. Minimum macro-F1 score that we will define together after a quick look at class balance.
2. End-to-end inference on a fresh audio clip takes no more than two seconds on CPU.
3. All code is clean, commented, and runs from a requirements.txt or environment.yml without modification.
Please let me know what initial dataset size you would like to see and any additional details you need before you dive in.
Scope of work
• End-to-end data flow: ingest the recordings, extract meaningful acoustic features (MFCCs, spectrograms or any feature set you recommend) and build a robust training/validation split.
• Model development: you are free to choose your preferred stack—TensorFlow, PyTorch, scikit-learn, or an ensemble of architectures—as long as the final model hits the agreed accuracy on a held-out test set.
• Experiment tracking & reproducibility: include clear notebooks or scripts plus a brief report summarizing metrics, confusion matrix, and the hyper-parameter search process.
• Deployment readiness: package the trained model and preprocessing steps so I can load it with a single call (e.g., a saved .pt/.h5 file with a companion inference script or a lightweight REST endpoint).
Acceptance criteria
1. Minimum macro-F1 score that we will define together after a quick look at class balance.
2. End-to-end inference on a fresh audio clip takes no more than two seconds on CPU.
3. All code is clean, commented, and runs from a requirements.txt or environment.yml without modification.
Please let me know what initial dataset size you would like to see and any additional details you need before you dive in.