{"id":33141,"date":"2024-04-11T21:49:30","date_gmt":"2024-04-12T01:49:30","guid":{"rendered":"https:\/\/ckms.ca\/?p=33141"},"modified":"2024-04-11T21:49:30","modified_gmt":"2024-04-12T01:49:30","slug":"project-document-my-aws-experience-trying-to-create-a-simple-voice-to-text-service","status":"publish","type":"post","link":"https:\/\/blog.ckms.ca\/?p=33141","title":{"rendered":"project – document my aws experience trying to create a simple voice-to-text service"},"content":{"rendered":"\n

I want to have a simnple web service that can take my digital voice recored (hereafter dvr) and transcribe to text, send transcription (after my approvall?) to a llm like chatgpt or gemini or whavs, returning a summary, a summary half as long\/twice as condensed, and optionally the soundtrack files, and optionally in other format\/styles, return by emaik, or ideally direct to my preferred local storage (or why not just cloud? or hwy not pboth<\/p>\n\n\n\n

chatgpt to summarizze; look at filename or other metadata for instucitno, for instance, if the filename or metad has “letter-to-editor”, then the results are returned as a leteer to the editor, or “letter to mom”, even having it guess by looking at content and past content, earning each detinees style of conversation.<\/p>\n\n\n\n

currrntly i do the following manually<\/p>\n\n\n\n

  1. Copy file from my dvr to my PC<\/li>
  2. upload audio fiiles to AWS bucket<\/li>
  3. on aws, create a transcription job, manually filing the details;
    1. I can do this at a rate of 10 ramblings in 15 minutes if nbot distracted; but tedious<\/li>
    2. maybe name the files by the first 10 weords?
      1. summerizer steps below via llm probs do a bestter job<\/li>
      2. maybe have it pause with suggestin? job transcribe is done, then make it wait until confirmation of nameine; this is important be cause so much can be done just with filenme system, not thing external, ex. maybe better use of properties llike os\/2 metadata<\/li>
      3. <\/li><\/ol><\/li><\/ol><\/li>
      4. on competion (no mre than a minute so far), download the 3 files
        1. aws bucket explorer doesn’t let multiselected files to be downloads, so one at a time ffs<\/li>
        2. aws returns the transcription in .json format, so need to extract the text<\/li>
        3. should be trivial with python, or aws tools?<\/li><\/ol><\/li>
        4. send the transcripts to me by email
          1. email easy<\/li>
          2. ideally files would be stored locally or in google drive or othjer cloud service<\/li>
          3. better: mkae a post to blog or notion or wheves?<\/li><\/ol><\/li>
          4. <\/li><\/ol>\n\n\n\n

            2024-04-09 – do 20 transcriptions manual, 2 x ~15 minutes each 10<\/p>\n\n\n\n

            2024-04-10 – did 10 more manual transcription jobs:<\/p>\n\n\n\n

            \"\"<\/a><\/figure>\n\n\n\n

            with 79 files in my output bucket:<\/p>\n\n\n\n

            \"\"<\/a><\/figure>\n\n\n\n

            2024-04-11 – further work on documenting how I do this manually using AWS Free Tier<\/p>\n\n\n\n

            extracting text of the transcription from a .json file is trivial; Gemini+ did it:<\/p>\n\n\n\n

            \"\"<\/a><\/figure>\n\n\n\n

            as can a .json file viewer:<\/p>\n\n\n\n

            \"\"<\/a><\/figure>\n\n\n\n

            2024-04-11 – i’ve bumped into my limit for tranjscriptions; 3600 seconds ( 60 mins), whcih might be the time of the audio playing time of the ~40 I did last few days?<\/p>\n\n\n\n