heidihealth.com.au

Part II

I suppose it's not unusual for LLMs to sometimes take 1 step forward, and 2 steps back.I use a homebrew AHK app with open-source plugins to do my own day-to-day STT on the desktop. After OpenAI rolled out "gpt-4o-transcribe" and "gpt-4o-mini-transcribe", I went to test-drive these "SOTA" models - with high hopes.And they were instantly worse than the 2-year-old Whisper-1 in some important ways to me.Basically, on the "zigzagging path of progress" LLM takes, I am willing to cut Heidi some slack. I even wonder how much control Heidi actually has over the model(s) they are using.Below is taken from Heidi's own "Advanced Guide on Template":"General AI instructions can sometimes work better at the start or end of a template. If an instruction isn’t performing as expected, try moving it. There’s no fixed rule, so experimentation is key."Based on stuff like that, how much do we think Heidi even knows what goes on in Heidi's mind? ¯\_(ツ)_/¯Anyhow, a few quick peer-to-peer tips, FWIW:1. Use "paragraph", and not "new paragraph", as the trigger word to deliver double line breaks. Just tell Heidi to "MUST double-space" or "MUST add a blank line" upon hearing "paragraph". That's because Heidi will reliably transcribe "paragraph" but not "new paragraph" - reliably, even if latter was added as a phrase in prompt.FWIW, neither could gpt-4o-transcribe, nor whisper-v1 (but paradoxically more reliably than gpt-4o). Yes, Dragon could give you "new paragraph" 1 million times out of a million. But gpt-4o-transcribe only gets it 6-7 times out of 10, for me anyway.But AI gets the singleton word "paragraph" 99%. Therefore I use that for trigger instead.2. For snippet insertion, use nouns Heidi will reliably discern, such as "snippet", "template", or "fragment", to lead the trigger phrases for your "AutoText" (as in Replacement). Initially I was using "Insert" as the lead word, since I do that in DMO. It sucked because Heidi interpreted "insert" as an action.3. For putting line breaks into a snippet, use "<br><br>" (it may or may not show up correctly here, depending on how your browser parses). Anyway, that's down to browser HTML parsing, and not likely a Heidi feature per se.4. For the few times I used Ambient Listening virtually, Heidi did just fine with Google Voice as a soft phone while I ran both from separate tabs in the same Chrome browser. FYI Heidi does NOT recommend this approach in their published "Telehealth Guide".I didn't bother with any desktop app just yet. But I commiserate with the previous reviewer if the desktop app is indeed having ++ teething problems. Unlike LLMs doing their "own thing" to some extent, getting an app to function as intended is a prerogative that Heidi exercises control over.Now onto my own grievances:1. Heidi monumentally dropped the ball on its phone app - there are no two ways of putting it. An audio-based phone app doesn't process phone audio? Can I market a piece of exercise equipment that one can't exercise on? As it stands, it is a mere "memo recorder" that uploads to the Heidi server (and glitchily at that).Is telehealth not done over the telephone, which everyone nowadays understands to imply cellular devices primarily?And if there is one scenario where I do need "Ambient Listening" more than any other, it is over the mobile device. These are the times when I am often on the go where I have no means other than Heidi to record something wholesomely.BTW, above is the chief reason I docked 1 star for my review. That ambient listening didn't wow consistently is something on which I can exercise my share of "forbearance". That a so-called phone app doesn't transcribe phone conversation is neither comprehensible nor excusable!2. Heidi should work on improving the reliability of "custom vocabulary". I accept that this one might be a tall order, especially since the OpenAI "SOTA" models are hit-and-miss, even with prompting. But here is hope.3. A glitch I noticed very recently is that my custom templates tended to be AWOL when I freshly log into a browser session. Reloading the browser doesn't bring them back. Often after the first transcription they then showed up.4. A piece of "low-hanging fruit" Heidi can pluck quickly is to implement a feature whereby I can have a set of "global AI instructions" that apply across the board to a slew of templates. I am tired of having to copy and paste to a bunch of others when I make changes to one.5. Minor: Add a toggle for those of us who don't fancy AI to put anything at all into any section that bears the title "Assessment".6. Relatively minor: Declutter the Settings menu. But dedicate a more prominent button to "Regenerate output" (which does have a keyboard shortcut, granted).Would appreciate if you can fix what's in your control, Heidi!

Date of experience: Aug 03, 2025