This seems to happen across multiple Mma versions...

Eh, now it works again. Something really funky is going on, or I have forgotten exact semantics of this stuff.

Not sure if this is a good question, but does anyone know how to change a list {1,2,3} to a set of replacement rules {a->1, b->2, c->3}?

mathematica.stackexchange.com/questions/3070/… Nevermind, I found the question!

Bit surprised at the performance of SpeechRecognize. I have some basically ideal audio and I get practically garbage results. Sadly there's no obvious way to "seed" the output given a script or expected words or phrases or something to get more usable output

@CarlLange I just did a handful of tests via SpeechRecognize[AudioCapture[]] and it was pretty dead-on - could you share any of those audio files?

@kirkus Yeah, if you like! They're irish-accented, which is probably what the problem is. I can share them with you later on if you're interested. I built a video-to-subtitles thing to make it quicker to edit videos:

@CarlLange I really wonder how sensitive those things are to accent. As far as I can tell, Google assistant, for instance, is pretty bad understanding English accents of my social circle despite they're quite understandable in my opinion (and not the dreaded "rally English"...)

(Normally you need to listen to the audio to remember what it's saying, and it makes the editing process quite slow, so subtitles that even just give the gist really speed things up)

Not a great screenshot, but hopefully you get the idea :)

@CarlLange sure, even just one of them that performs poorly would be interesting to test.

@kirkus we.tl/t-hdtaD5gKST

I did very little pre-processing and so on (just normalising and amplifying) before running SpeechRecognize so I'd be all ears for some ideas that might get slightly better results. But at the end of the day I'd like to be able to inform the function about phrases like "Kerry Way" so it doesn't give me "carryme" or similar

well, I certainly see what you mean

I suspect it might also be because of the female voice, I would take a totally unfounded guess that the training data might underrepresent women

have you tried running it through any other speech-to-text engines, just to compare?

the audio sounds pretty clean to me

Not really, I also tried the DeepSpeech NN from the NN repo, but I'm pretty sure SpeechRecognize is just using that under the hood anyways. I didn't get better results with it. Mostly I was just procrastinating from actually editing the video

good STT is pretty tough, I had a moderate level of exposure to it at a hearing aid company I worked for

@kirkus Yeah, I found that SpeechRecognize was quite sensitive to volume, which is why I added the gain and so on. But I would have expected better performance than what I got.

this was 10 years ago, and they actually thought they were going to get real-time audio language translation on the fly on a hearing aid

I just wonder how e.g. Android or iOS would do interpreting that

@kirkus Ah, that sounds like one of those self-driving car ideas that sounds just plausible enough to get funding for... But I'm sure it was an interesting experience. I had thought that NNs were pretty new to STT and had hoped to see a bit of a leap in accuracy (and maybe it would if I actually did some real work with the DeepSpeech network's decoder).

I think Azure has some pretty good STT, but I don't have the time at the minute to really try it out azure.microsoft.com/en-us/services/cognitive-services/…

(Since I should be editing video 😅)

Recent Google blog post on the topic: ai.googleblog.com/2021/01/…