Skip to content

Latest commit

 

History

History
45 lines (29 loc) · 1.27 KB

README.md

File metadata and controls

45 lines (29 loc) · 1.27 KB
Raw

ASR

Input

Audio file provided in .WAV format with a sample rate of 16 kHz.

example.mp4

Note that the above example is in mp4 format as GitHub does not support embedding .wav files in README files. However, audio files provided on GCP will be .wav files.

Augmentations

Training was conducted with the following audiomentations augs:

HighShelfFilter(max_gain_db=6.0, p=0.3),
LowShelfFilter(max_gain_db=6.0, p=0.3),
BandStopFilter(p=0.3),
TimeStretch(min_rate=0.9, max_rate=1.1, p=0.3)

Output

Transcription of audio file. Example: "Heading is one five zero, target is green commercial aircraft, tool to deploy is electromagnetic pulse."

Submissions

Whisper Small en non combined with niner:

  • Accuracy: 0.9934860264761505
  • Speed Score: 0.825095724074074

Whisper Small en combined with niner:

  • Accuracy: 0.9926455137633957
  • Speed Score: 0.8088188985185185

Whisper Small en combined without niner:

  • Accuracy: 0.9922252574070183
  • Speed Score: 0.8239869433333333

Parakeet RNNT 0.6b non combined:

  • Accuracy: 0.9686909014498845
  • Speed Score: 0.8335885779629629

Parakeet RNNT 0.6b combined:

  • Accuracy: 0.9892834629123766
  • Speed Score: 0.8456000411111111