It was very hit and miss. The videos I was testing already had hardcoded subtitles - it was interesting seeing the difference between 100% accurate subtitles and the auto-generated captions; probably 70% accurate... enough failure rate to make it difficult to impossible to follow the flow or understand the context.
Regional dialects may also play a part in it. The videos I was testing with were in Australian English.
