PrevNext

Explore showcase

Bird Audio Classification

Ideator, Programmer · Spring 2026

Project6 / 13

Overview

At a glance

Deep learning final exploring transformer audio backbones on real field recordings. The emphasis is on reproducible training, honest evaluation, and communicating comparative behavior—not chasing leaderboard scores.

PyTorchHuggingFaceHuBERTAST✚
View GitHubView GitHub

Features

  • Side-by-side HuBERT vs AST benchmark pipeline.
  • 10-species classifier built from common U.S. birds.
  • End-to-end training/evaluation with reusable caching.
  • Comparative reporting for overall and per-class behavior.
  • Demo inference mode with playful presentation flavor.

Implementation

  • Replaced pretrained classification heads for downstream 10-class objective.
  • Built audio preprocessing and feature preparation pipeline for trainer-ready datasets.
  • Implemented HuggingFace Trainer loop with validation and iterative hyperparameter tuning.
  • Added `save_to_disk/load_from_disk` caching to skip expensive dataset rebuilds.
  • Implemented confusion matrix and class-level reporting to inspect failure patterns.
  • Measured inference runtime and computed rough efficiency metrics (params + FLOPs/MACs).
  • Added persona-style print output in inference demos for a more engaging presentation.

Challenges

  • Bugs often surfaced late into long training runs, making debug cycles expensive.
  • Optimization scope remained basic (mostly hyperparameter-level) due time and workload constraints.
  • Balancing class schedule with long training iteration cycles was difficult.
  • Not intended as publication-grade research; focused on learning and functional comparative implementation.

Reflection

  • This idea came from prior BERT assignment experience (MRPC, SQuAD, SST-2), and I wanted to apply transformer-style thinking to ecological monitoring.
  • One search showed specialized bird-audio models already exist, but I still wanted to compare base-ish approaches and learn through building.
  • It was not smooth. We had multiple “find the bug after an hour of training” moments.
  • Even if this is basic by advanced ML standards, it was still hard for me and I’m glad we got a complete pipeline running.
  • I added goofy persona narration in the demo because I think technical demos are way more memorable when they have some personality.