[{"data":1,"prerenderedAt":152},["ShallowReactive",2],{"project-41224623":3},{"id":4,"name":5,"line":5,"line_source":6,"title":7,"url":8,"domain":9,"maker":10,"posted_at":11,"year":12,"points":13,"comments":14,"rank_in_year":15,"shown_in_year":16,"top":17,"tier":18,"category":19,"subcategory":20,"format":21,"keywords":22,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":32,"aspects":33,"strip":35,"post_text":37,"what":38,"said":63,"maker_in_thread":92,"shape":101,"grid":104,"attempts":112,"same_week":113,"maker_shows":144,"maker_hidden":145,"reader_version":146,"card":147},"41224623","LLM Aided Transcription Improvement","title","Show HN: LLM Aided Transcription Improvement","https:\u002F\u002Fgithub.com\u002FDicklesworthstone\u002Fllm_aided_transcription_improvement","github.com","eigenvalue",1723472122,2024,11,4,2489,3515,71,"shown","ai","transcript cleanup","cli",[23,24,25,26,27],"transcription","whisper","llm","speech to text","markdown","quiet",null,{"positive":31,"neutral":32,"mixed":31,"critical":31},0,2,{"execution":34},{"positive":31,"critical":31},[36,31,36,31],3,"I recently submitted another project for using LLMs to correct errors and improve formatting of OCRed documents which was well received. The low cost and high quality\u002Fspeed of the latest \"value tier\" models from OpenAI and Anthropic have made it possible to get compelling results at a very reasonable price in that application.\n\nIt occured to me that the same approach taken there (namely, splitting documents into chunks and sending each chunk through a chain of LLM prompts that each take the output of the previous prompt and apply an additional layer of processing) could be easily applied to a ",{"problem":39,"audience":45,"pricing":29,"motivation":49,"time_to_build":29,"stack":53,"openness":62,"stage":62,"team_size":62},{"value":40,"source":41,"evidence":42},"Raw spoken-word transcripts are rambling and not polished enough to read as written content","maker",[43],{"id":44},"post",{"value":46,"source":41,"evidence":47},"YouTubers and people with recorded content who want written versions",[48],{"id":44},{"value":50,"source":41,"evidence":51},"Extend the author's OCR correction approach to spoken-word transcripts",[52],{"id":44},[54,58],{"value":55,"source":41,"evidence":56},"Whisper (transcription output input)",[57],{"id":44},{"value":59,"source":41,"evidence":60},"OpenAI and Anthropic value-tier LLMs",[61],{"id":44},"unknown",{"praise":64,"objections":73,"requests":78,"questions":86,"alternatives":91},[65],{"value":66,"source":67,"evidence":68},"Useful for turning voice memos and rambling recordings into readable text","commenter",[69],{"id":70,"by":71,"text":72},"41231314","gavmor","I record long, rambling voice memos in noisy environments which Whisper struggles to parse. Perhaps this can rescue me from the tedium of hand-stitching the fragmented results…",[74],{"value":75,"source":67,"evidence":76},"Garbage in, garbage out: poor Whisper output on noisy audio limits results",[77],{"id":70,"by":71,"text":72},[79],{"value":80,"source":67,"evidence":81},"Explain how chunk size is chosen and its effect on quality",[82],{"id":83,"by":84,"text":85},"41233728","ramonverse","i'm curious about the chunk splitting approach you mentioned. how do you determine the optimal chunk size for processing? seems like there could be a tradeoff between context…",[87],{"value":88,"source":67,"evidence":89},"How is the optimal chunk size determined and tested",[90],{"id":83,"by":84,"text":85},[],{"comments":32,"answered":93,"key":94},1,[95,98],{"id":96,"text":97},"41231384","Give it a try and let me know how well it works for you! You might improve results by slightly tweaking the first stage of the prompting to mention that these are your own voice…",{"id":99,"text":100},"41234058","I just tried messing around with the chunk size until the results looked best to my eye. One important thing to note is that each chunk includes a portion of the previous chunk…",{"commenters":36,"max_depth":32,"maker_replies":32,"first_comment_mins":102,"lifespan_hours":103},679,20,[105,107,109,111],[70,31,36,106,71,31],"project\u002Fneutral\u002Fnone",[96,31,31,108,10,93],"maker\u002Fneutral\u002Fdocs",[83,93,36,110,84,31],"project\u002Fneutral\u002Fexecution",[99,93,31,110,10,93],[],[114,119,124,129,134,139],{"id":115,"name":116,"points":117,"category":118},"41247023","ytch",2741,"consumer_social",{"id":120,"name":121,"points":122,"category":123},"41224689","PGlite",509,"data",{"id":125,"name":126,"points":127,"category":128},"41203306","LLM-aided OCR",479,"productivity",{"id":130,"name":131,"points":132,"category":133},"41217109","Grandma's Word",454,"games",{"id":135,"name":136,"points":137,"category":138},"41203475","Thunder Compute",344,"infrastructure",{"id":140,"name":141,"points":142,"category":143},"41213711","flutter_rust_bridge",301,"languages",15,false,"v2",{"id":148,"name":5,"line":5,"ai":145,"cat":19,"year":12,"points":13,"comments":14,"rec":28,"s":29,"top":17,"kw":149,"maker":10,"domain":9,"tier":18,"attempts":31,"strip":150,"url":8,"mix":151,"praise":66,"objection":75,"sub":20},41224623,[23,24,25,26,27],"3030",[31,32,31,31],1794036659196]