Hacker News Reader

Comment by genewitch | Hacker News Reader

It is literally impossible to transcribe voice, especially if you whisper. There's no way to model the language, it's too large - amanzon has many computers. Your computer is like a tortoise, it can't do text-to-speech. There's no way you can get any of this to use the Web, Dav... er, Cal. Trying to do this would be like trying to torch a python with your bare metal hands.

so in conclusion: nope.

View on HN

This topic comes up periodically and I feel that there is an underlying incorrect assumption about the use case. The assumption is that these devices, mobile phones, etc, run speech-to-text and upload text to the company. Quite an undertaking! Consider the humble "Clapper", a 1980s vintage device that turned lights on or off with a simple clap of the hand. This device has a vocabulary of 1 word, a clap, yes or no. Now consider that our device or cellular telephone has a vocabulary of 32 words such as "car," "baby," "television," "insurance" and so on. Much much easier to ONLY look for those trigger words and slip 5 extra bits somewhere into a network packet. Suddenly you gets ads for a car dealership, an insurance company, children's formula, and so on.

I am going to turn off my adblocker tonight to see if a product we talked about a lot today shows up anywhere. I won't mention it here, because i haven't typed about the product at all, ever; it was only spoken and i sent 2 SMS asking two people if they had one of them i could borrow today. I won't need one again for 5 years, so there's no sense in me buying one, now.

oh also i have "FUTO Keyboard" on android, and it uses a local LLM (you can pick three "sizes" depending on your device specs and desired battery life) to transcribe my voice. It's pretty good for non-technical conversations, but it starts to trip up if one's jargon pronunciation is close to other words - "Tea Sip, I pee" sort of confounding errors.

Yeah recognising some words and having a hand crafted decision tree over an LLM can go a long way to get useful shit done.

Setting a timer, calling someone, turning on a light.

But is local CPU Speech to Text really a dead end?

there were speech synthesis boxes in the 80s, so i doubt this is intractable. at&t did a lot of work on it through 2010 (perhaps beyond, but i stopped paying attention), where the syllables are nearly perfect, assuming you can massage the source text using context clues to force the proper pronunciation. Is it "read" or "red"? well, that depends on the context.

Even in the voice-cloning TTS you have to deliberately spell things to be pronounced correctly.

so it comes down to "guessing the context" for a computer, which is something that LLMs can do fairly well, and raspberry pi can run small LLM, so who knows. Sentiment analysis is one thing, i'd probably mess with prompting an llm to "correct the grammar" and see what words change - or even "change words for synonyms that can't be misconstrued by the TTS engine, unless the tone of the phrase or passage changes dramatically, in which case, phonetically spell the word in question"

but i only think to do this because LLMs exist. if you had asked me 15 years ago how to automate "pronunciation" of english i'd have said "if IBM and AT&T and Apple can't figure it out..."

Eh, Whisper is pretty good, especially if you have some GPU to throw at it.

huh, i've never heard of openai-whisper. I certainly didn't mention it in my first sentence.

You said "especially if you whisper" :-) for the skimming reader that parses as whisper the model. I have to admit I did it at first too.

here, i'll give the key to decode my message ;-)

It is literally impossible to transcribe voice, especially if you whisper. There's no way to model the language, it's too large - amanzon has many computers. Your computer is like a tortoise, it can't do text-to-speech. There's no way you can get any of this to use the Web, Dav... er, Cal. Trying to do this would be like trying to torch a python with your bare metal hands.

that is:

  baremetal pytorch https://en.wikipedia.org/wiki/PyTorch

  WebDAV/CalDAV https://en.wikipedia.org/wiki/CalDAV 

  tortoise-tts https://github.com/neonbjb/tortoise-tts/issues

  LLM/Large Language Model https://lmstudio.ai/

  whisper https://github.com/openai/whisper/

in reverse order

You said "especially if you whisper" which I read as a level of speaking, primarily because I have experienced difficulty with Whisper while whispering :)

click parent on your comment and see my comment to your sibling, i was just obfuscating while being glib. I listed everything you need to run a voice assistant, at least software-wise.

Reply on news.ycombinator.com