Reson Collector
Dataset Recording & Audio Labeling Pipeline for Acoustic ML
“How do you build an acoustic dataset collection pipeline from scratch?”
Case Study Index (5 Sections)↓
Research Problem & Motivation
Machine learning models for signal processing depend heavily on clean, consistently labeled datasets. Reson Collector automates the recording workflow, enforcing consistent distance, duration, and sampling parameters across hundreds of audio gesture trials.
Manually recording, naming, and organizing audio clips for machine learning model training was slow and error-prone. I needed a tool that ensured every training sample had identical recording conditions and metadata.
Signal & Data Pipeline
Audio Prompt -> SoundDevice Recording Buffer -> Clipping/Silence Threshold Check -> WAV Encoder -> JSON Manifest Generator.
Engineering Decisions
Writing audio buffers to disk during active recording caused minor frame drops on lower-spec hardware.
Buffered raw audio frames in memory NumPy arrays and flushed to disk asynchronously after trial completion.
Higher temporary RAM usage during long recording sessions.
Eliminated audio frame drops across all dataset recording sessions.
Challenges & Solutions
Detecting Clipped or Corrupted Trials
Issue: Users occasionally performed gestures outside the microphone range, producing useless silent audio samples.
Decision/Solution: Added instant post-trial amplitude threshold checks to automatically reject silent or clipped recordings before saving.
Future Research Directions
Key Takeaways
High-quality dataset collection pipelines save vast amounts of debugging time during ML model training.
Enforcing metadata structure at capture time prevents dataset inconsistency downstream.
Related Engineering Projects
Reson
Researched and built a desktop application harness exploring ultrasonic sound reflection analysis for touchless gesture detection.
FocusX
Built a zero-dependency minimal Pomodoro timer using vanilla JavaScript with system epoch time synchronization to overcome browser tab throttling.