Hacker News Reader

Comment by nojs | Hacker News Reader

> The service costs money. It is not automagic speech to text, but rather there is often (usually?) an actual human typing it in.

I would wager whisper is more accurate than said human, and definitely cheaper.

View on HN

Deaf here. In some contexts yes. In many, nope. Whisper can't talk to the other person. AI transcription systems on the phone are very awkward for both parties. Slow. Long pauses. Asking for clarification is a pain since the turn taking is so slow. The error rate is also way too high in some contexts (some heavily accented speakers for example).

Should’ve asked them for them to send you some of that supposed savings from using AI for taking their bet.

Interesting, thanks for sharing. I didn’t realise it was such an active task. Sounds like it’s much more complex than just transcription.

Transcription services are still horrible if you have (a) some kind of accent, (b) non-standard terms, or (c) noise. They can be better than nothing if you want some meeting notes and no one is taking them, but they're definitely not something you want to rely on for an important conversation.

A human in a conversation of this type has the major advantage that, when they're not sure they've understood a word, they can ask the other party to repeat it.

There are (still) actually good transcriptionists who will deliver a meticulous transcript with all your weirdly abbreviated jargon thoroughly researched (which can be a feat of truly impressive "detective" work at times), multiple speakers discerned, expert punctuation choices that fit the tone and meaning, etc. But it costs way more than most people can/will pay. I worked at a "real" (non-sweatshop) transcription outfit ca. 2012. Judging by my pay rate of $75 per audio hour at the time, I would guess that the really good operations these days must be charging close to $200 per audio hour. A quick google found one anecdatum corroborating this guesstimate: http://www.transprof.com/prices-faq

"you mean to tell me AI can't just replace a human?"

Whisper transcription is extremely good. For most real-time transcription tasks I do think it performs as well or better than the average human at this point. In this case it sounds like the task is more active and involves actually thinking and talking to the other party, which makes it much harder. But for simply transcribing phone calls, the state of the art is very very good.

Reply on news.ycombinator.com