I love having auto-generated subtitles available, but I am still skeptical that we can call them "accurate" yet. For many languages the models are quite weak, and even for English and German on YouTube I see errors all the time.
I love having auto-generated subtitles available, but I am still skeptical that we can call them "accurate" yet. For many languages the models are quite weak, and even for English and German on YouTube I see errors all the time.
Even for English (which surely has the best support) anything outside of urban American accents is still incredibly weak - you can pretty much guarantee very noticeable mistakes on even highly intelligible mainstream urban British/Irish/Aussie/Kiwi/SA accents. Stronger regional / rural accents & dialects usually render generated subtitles entirely useless.
Definitely a domain a million miles from being solved if even English is this bad.
Yeah I agree the Youtube ones are not very good, I am not sure what they have implemented behind the scenes for it. However this uses the Whisper Large model which is quite a bit better than the Youtube auto generated ones. Only catch is you need to have the video on disk and wait a bit longer lol
Youtube is very bad, but I have honestly not yet seen anyone implement this well on any platform, which indicates to me that it's not a solved problem. Very open to being proven wrong here though - do you have examples of Whisper Large in use in a large consumable sample?
It's not fully solved, there are still errors but its improved a lot the past few years especially compared to Youtube. The best approximation is the Whisper models atm as far as I have seen. There is a write up here on hugging face which talks more about the models and the benchmarks against it.
https://huggingface.co/openai/whisper-large-v3