benchmark performance (Broadcast News)

[Veldhoen]

benchmark with the same setup for fair comparison:

• [ ] faster-whisper
• [ ] Whisper-X
• [ ] Huggingface transformers-implementation
• [ ] original OpenAI (baseline)

[greenw0lf]

Parameters to experiment with that apply to all implementations:

• float16 vs. float32 (whether there's a huge WER difference between the two, if not it is preferred to use float16 as it is faster and consumes less GPU memory)
• large-v2 vs. large-v3

[greenw0lf]

Labelled data benchmarked, unlabelled data still to go

[greenw0lf]

Small mistake: loading the diarization model for each file instead of just once for WhisperX (alignment model needs to be loaded per file as it uses language info)

[greenw0lf]

Re-benchmarking faster-whisper (both labelled and unlabelled) because time to transcribe should be measured from the point when model.transcribe is run until the output of the function is saved to a JSON file.

Also re-benchmarking WhisperX for the reason mentioned in the previous comment (expected less time spent per file)