Voice recognition | Hacker News Reader

Voice recognition

imperialviolet.org

47 pointsingve10 comments

For those who are wondering:

This article is about (automatic) speech recogniton, also known by its acronym ASR or as speech-to-text (STT).

It has nothing to do with automatic speaker identification, which in turn is also different from speaker diarization. The latter refers to the process of automatically identifying from a group of copresent speakers who is the one speaking at any point in time without caring about the actual identity of that person.

I so much hate it when people use "voice recognition" when they mean "speech recognition". Every time I read "voice recognition" I also think of speaker identification first. I think the names are quite self explanatory, so I have a hard time understanding why people use the terms the "wrong" way.

I tried something like that GPT trick and it is pretty hard to get it to not over-correct and over-index the vocabulary words.

For instance if I tell it that "OneTake AI" (my company) or "OneTake Chat" (one of our features) belong to the vocabulary words, it will twist correctly-transcribed sentences to make it appear in there.

> Here is the biggest opportunity in artificial intelligence today if you're using Chat in your app

(which is already grammatically correct!)

would suddenly become

> Here is the biggest opportunity in OneTake AI today if you're using OneTake Chat in your app

This sycophantism of the LLM causes more trouble than it fixes. It's even worse if the vocabulary provided includes names, because then the model will randomly twist some sentences (mostly the benefit-driven ones, e.g. "so you can get faster results today") to add "... thanks to EXPERT_NAME!"

It is also very difficult to get the model to edit structured text (like a JSON of the transcription) without breaking the schema or the IDs.

I am still iterating to find a better method, and I think the "Functions" API can also help.

If this can be valuable to anybody else, here is an improved version of OP's prompt that will perform better than the one in the original article (it fixes every mistake plus punctuation and capitalization), with chatgpt-3.5-turbo-16k at a temperature of 0.7.

I find that using the wording in the OP's article seems to create fewer sycophantic corrections than my previous approach, so that's a win. However the structure of my prompt seems to work better. So here's a combined result.

System prompt:

```

  The following English transcription is the output of a poor speech recognition system. Please correct its errors and reply with the corrected version without any preamble.

  Some of the uncommon words that the speech recognition system probably misspelled are: 
  ["U2F", "CTAP", "FIDO", "CTAP2", "JSON", "FIDO2", "Web Authn", "bearer token", "hash", "phishing", "RP ID", "passkey", "caBLE"]
  If those words are present and misspelled, please fix their spelling.

  Please fix the spelling, grammar, any inconsistencies and transcription errors in each sentence.

```

Note: to improve performance, I also recommend:

- providing the model with a before-and-after example, by using the USER and ASSISTANT roles to roleplay the whole thing (not putting the example in SYSTEM!)

- providing the source text as a CSV or JSON, and asking the model to provide JSON as the output.

- prefixing the Assistant response with an Assistant reminder of the instruction, like this:

ASSISTANT

```

  Return a corrected version with correct English grammar, punctuation and spelling.

  Correction (as JSON):

```

It just so happens I'm in the middle of scraping numeric data out of a proprietary application using dictation. It took a bit of experimentation to get good results, but it's faster than any other scraping method I could find for this specific use case. I'm using the dictation tool in the Drafts app, and used ChatGPT to write a Drafts Action script to clean up and format the data after dictation.

Lately I've been using Google Recorder, and then when I have text I _actually_ want transcribed I take the audio and put it through Whisper.

From there it would probably be helpful to send it through GPT, but I haven't gotten that far. I think GPT is more likely able to fix formatting than text at that point. But overcorrection here is an issue, the formatting isn't obvious, and I haven't explored the prompting in depth.

For actual correction I think it might be superior to get GPT to highlight issues and provide alternatives (using some made-up markup language), but not actually rewrite.

To get even _more_ fancy, I can imagine giving GPT the low-quality Google transcription, having it identify terms and subject, and then using that as a prompt to Whisper.

I'm building a tool for transcriptions where your brand, product or any other technical jargon or heck even your own name does not need a fine-tuned model all the time. Both as a free native Mac app & a SaaS tool for those who need to process in bulk. Hit me up at nebo@minusgreed.com to check it out, will launch as FortuneSpeech.com once out of beta.

Thanks to this post, I kind of have another idea for a 'corrections to transcript' feature that Llama2-7B even on CPU can help with.

See sebastiennight's parent comment above - I think they meant to reply to you.

Curious to see your response.

Actually I wasn't, but if this can be useful to GP I shared my actual prompt in a follow-up comment.

I would be interested to see if LLama2 can perform a similar task.

You can also try Whisper, or try google or amazon's ASR and give it hints/dictionary of common words.

Reply on news.ycombinator.com