From my experiments generating subtitles from TV/movie audiotrack, 75% (worst case) to 95% (best case). If you model it as a standard distribution, somewhere ar

From my experiments generating subtitles from TV/movie audiotrack, 75% (worst case) to 95% (best case). If you model it as a standard distribution, somewhere around 85-90% accuracy. Most services provide much better accuracy for stuff like calls or conferences with proper microphones and minimal background noise than for things like TV shows and movies. If the input audio is noisy, I would do some noise filtering before piping it into conversion.