AI video to text conversion quality

I study Korean via fairy tales from youtube. The video to reader conversion used to be really accurate. Not perfect, but pretty darn good. Lately it has been terrible. Even multiple mistakes in one sentence. Has anybody else experienced this?

2 Likes

I experiencing the same problem with the LingQ importer. Also, I raised an issue about this problem. check this out as a makeshift.

1 Like

I tried doing that but couldn’t get it to work. Even with your prompt, it still doesn’t work. ChatGPT is changing the words a lot. Any ideas?

1 Like

I’ve tried several LLMs including GPT, Gemini, and Claude. Of them, Gemini-flash was the best option. The following is my impression about other models:

  • gpt-4o-mini: better than LingQ importer, but sometimes omits some words and sentences.
  • claude-haiku: more expensive than other mini models.

Of course expensive model will do its job better, but a transcript is typically quite long, so we have to compromise expense and quality.

Since I’m using LLM through API, It might be differ on the website.

2 Likes

Gemini works much better. Thanks a bunch

1 Like

당신의 한국어 공부를 응원합니다. :slight_smile:

2 Likes

Please ignore this. I’m just using it to include the instructions:

Your task is to transform the following text into complete sentences while preserving every original word and phrase.

Please follow these steps to complete the task:

  1. Convert the input text into sentences according to these guidelines:
    a. Preserve every original word : Do not add, remove, or change any words from the original text.
    b. Create proper sentences with appropriate punctuation.
    c. Form logical paragraphs by grouping related sentences. Start a new paragraph when there’s a shift in topic or speaker.

  2. Separate at natural paragraph changes

  3. Separate at character changes (when a different character starts speaking)
    d. Apply formatting: Use no line breaks within paragraphs and separate paragraphs with a blank line.

  4. Remove any timestamps, “add a note,” and “jump to” if they happen to be included in the original text.

  5. After completing the conversion, review your work to ensure:

  • Every original word is preserved.

Remember, the key is to improve readability and structure while faithfully preserving every original word and phrase. Pay extra attention to ensure no words or phrases are skipped in the process. What you only can add to the original text are punctuation, new lines, and separating paragraphs by character and natural breaks.

1 Like