Effecient Audio Transcription

Introduction

This project aims to address the gap between transcription technology at the bleeding edge and usable, performant implementations of these technologies.

Usage

Currently, your best bet is to clone the repo on a machine with Nvidia drivers and Docker. Use Docker to build and start the container. You will need considrable (50+ GB) space to build the image due to Nvidia tooling. Docker will start a local server on the instance that responds to POST requests with your file.

Benchmarks

All evals done on LibriSpeech test-clean

AWS g4dn.xlarge (Nvidia T4) 16GB/4vCPUs

Stock parakeet-ctc-0.6b:

Performance Metrics: Total Execution Time: 331.72 seconds Total Audio Duration: 19452.48 seconds Total Transcription Time: 326.83 seconds Average Transcription Time per File: 0.1247 seconds
Files Processed: 2620 Overall RTF: 0.0168 Overall WER: 0.0204

Overall RTF: 0.0168 * $.526 /hour (g4dn.xlarge) = $.0088 /hour! *granted, this assumes 100% utilization

To Do

• Worker status updates
• optimize small file loading
• Explain project
• Cost analysis of existing transcription services
• Optimize data loading
  • check if eval data exists, download otherwise
• Web-reachable server
• Optimize model for inference
  • tensorRT conversion
  • mixed precision
• Automate provisioning
• Check for local model instead of Dl

License

This project is licensed under the GNU Affero General Public License v3.0. See the LICENSE file for details.