I had an interesting AI experience this morning. I went to a physio therapist and he recommended I go see my family doctor about something related to my pain. He used a lotta big words. I asked if I could write them down, but realised I couldn’t spell them or type fast enough. So I quickly switched to the chatgpt app, started a new chat and pressed the voice recording function. I handed the phone to him, he told my phone what was up and what I should tell the doctor, and a moment later Chatgpt had turned the recording into words, and also summarised the problem into language I could understand and study later.
The quality of the voice dictation is wonderful compared to what we had 1 year ago, and extraordinary compared to what we had 5 years ago.
Dunno about all dictation apps, but I’ve used the built in dictation a lot and I wouldnt have asked him to use it, because it’s not very good by comparison.
I think all of the other dictation apps I’ve used - the new whisper ones - use the same OpenAI dictation backend. The same one ChatGPT uses. Dunno for sure though.
If I had my Mac Studio running parallels with me, I could have tried using dragon naturally speaking, but i think the iPhone and Mac’s built in dictation is better than that. The phone is definitely easier to carry with me.
Ok. Next time you see a doctor ask them to write it down (they do this anyway), and provide it to you (which they also do). Then you don’t need to pay an AI company to sum it up for you.
In my experience at least, all of the new AI-powered voice recognition apps I have used have been orders of magnitude better than the traditional dictation apps available up until now.
This. When it gets into medical terminology, in particular anything that is Latin derivative, traditional dictation stuff falls over.
I also understand that some doctors don’t do after visit summaries. It depends on the size of the practice, how their systems work, etc.
That said, as a reasonably intelligent person, I consider it a professional failing if my doctor can’t explain something clearly enough to me that I can re-articulate it to somebody else.
It’s cool to know which tools do the job most reliably – but on some level, it bothers me that the job has to be done at all.
That may be true, but there’s no way I’m going to feed my or my family’s private medical information into an AI chatbot. I wouldn’t even use traditional dictation software for that kind of thing unless I’m convinced it’s never going to leave my device.
The new AI transcriptions are so much better than everything else, it’s not even close. They recognize mispronounced words. They capitalize proper nouns like book and movie titles based on the context of the sentence. I have no doubt that the AI transcription you got was leaps and bounds better than the normal transcription on the iPhone. The only drawback to them is that you have to break yourself of all the bad habits you adopted in speaking to that idiot Siri for the past five years.
FWIW, there are (at present) several AI models at work in this scenario: the transcription of what you/the therapist says is done with openAI’s Whisper model. The produced text is then read by ChatGPT (just as if you typed it) and based on that, it produces a summary or whatever you ask it to do.
So when you notice the quality of the transcription (probably as compared to the iOS dictation function), that is Whisper for you, not ChatGPT (strictly speaking). Whisper, btw. is free to use locally on your Mac (independently of ChatGPT), eg. to transcribe recordings of interviews, meetings, whatever, which is totally amazing.
That said: as soon as OpenAI roles out the audio part of GPT-4o, Whisper will no longer be part of the pipeline when we talk to ChatGPT. It will then understand the audio signal directly, i.e. without transcribing it into text, which is absolutely mind boggling. The main effect is obviously speed. That’s how it manages to respons within 300 ms. I can’t wait to try it out.
I wish that any of the AI chat bots could produce OCR scans of marked registers from UK polling stations after an election. These registers are marked by hand some carefully with ruled lines through an entry to show that someone has voted some a careless tick that could ambiguously apply to two different people. Such registers are then scanned and made available to candidates and their electoral agents for subsequent analysis.
Because thes registers contain personal information (names, address, dates of birth for those achieving majority in that year, plus whether they voted or not) the pages cannot be uploaded to any of the bots because the companies keep all uploads.
In a one-off one page experiment Claude produced something approximating the actual data but it was not perfect and one a few entries produced utter garbage.
