A very quick repo to showcase how you can experiment with Whisper (via Transformers) and make it run for your own use-cases!
All the experiments were run on a free Google Colab T4! 🔥
Results and Colab below:
Method | Time to Transcribe |
---|---|
fp16 | 62s |
fp16 + SDPA | 60s |
fp16 + SDPA + Speculative Decoding | 37.9s |
fp16 + SDPA + Chunking + Speculative Decoding | 43.8s |
Distil-whisper + fp16 + SDPA + Chunking | 17.2s |