Speech Processing

Wav2Vec2 on SUPURB

SUPURB
- Speech processing Universal PERformance Benchmark
- Audio classification dataset with multiple tasks
  - We used the Keyword Spotting subset
Wav2Vec2 out of Facebook research
- Transformer based Method

Improvement on Accuracy score using Perforated Backpropagation + Wav2Vec2

We downloaded the best performing open source code and added our technology
- Wav2Vec2 was developed by Facebook’s research team in 2020
Graph shows validation accuracy on SUPURB competitions keyword spotting dataset
Increase in accuracy score by 0.2%
- 12% error reduction
- Average gap between top 5 papers on leaderboard for this dataset is 0.27%

Number of Perforated AI Cycles