[{"data":1,"prerenderedAt":580},["ShallowReactive",2],{"project-32409966":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":29,"sentiment":30,"stance_mix":31,"scored_comments":35,"aspects":36,"strip":47,"post_text":48,"what":49,"said":94,"maker_in_thread":208,"shape":217,"grid":221,"attempts":544,"same_week":545,"maker_shows":42,"maker_hidden":572,"reader_version":573,"card":574},"32409966","tevr-asr-tool","State-of-the-art German speech recognition in 284 lines of C++","title","Show HN: State-of-the-art German speech recognition in 284 lines of C++","https:\u002F\u002Fgithub.com\u002FDeutscheKI\u002Ftevr-asr-tool","github.com","fxtentacle",1660128341,2022,231,128,224,2570,9,"notable","ai","speech recognition","cli",[24,25,26,27,28],"german speech recognition","offline speech to text","ctc beam search","tensorflow lite","wav2vec2","mixed",-0.287,{"positive":18,"neutral":32,"mixed":33,"critical":34},20,5,23,57,{"performance":37,"pricing":39,"idea":41,"execution":44,"docs":46},{"positive":38,"critical":38},2,{"positive":40,"critical":40},0,{"positive":42,"critical":43},3,13,{"positive":45,"critical":33},1,{"positive":45,"critical":40},[45,45,33,42,33,42,42,42,33,42,42,42,33,45,40,33,38,40,40,45,45,40,40,40,33,42,42,45,45,45,45,45,45,45,33,42,45,45,42,45],null,{"problem":50,"audience":60,"pricing":48,"motivation":66,"time_to_build":48,"stack":74,"openness":91,"stage":92,"team_size":93},{"value":51,"source":52,"evidence":53},"Accurate offline German speech recognition with compact client-side code","maker",[54,57],{"id":55,"by":11,"text":56},"32411566","In my opinion, the focus here should be on the fact that this is a state of the art AI which beats Facebook's wav2vec2 by a relative 16% improvement, Scribosermo (based on NVIDIA…",{"id":58,"by":11,"text":59},"32423486","Original poster here, I just wanted to say thank you! Despite some snark - which I totally deserve, but I can't change the title anymore - I also received some very helpful…",{"value":61,"source":52,"evidence":62},"Developers and industry clients needing auditable, installable speech-to-text",[63],{"id":64,"by":11,"text":65},"32412442","For a researcher, yes. But for understanding the trick there, you need to have read and understood the CTC loss paper. For people like my industry clients, on the other hand,…",{"value":67,"source":52,"evidence":68},"Maker wanted readable, auditable speech recognition and built a new decoder from their research",[69,72,73],{"id":70,"by":11,"text":71},"32419325","Actually, this model uses attention layers which are kind of like a query\u002Fkey=>value look-up and allow the model to merge the knowledge from neighboring time-steps into the…",{"id":55,"by":11,"text":56},{"id":64,"by":11,"text":65},[75,79,85],{"value":76,"source":52,"evidence":77},"C++",[78],{"id":55,"by":11,"text":56},{"value":80,"source":52,"evidence":81},"TensorFlow Lite",[82],{"id":83,"by":11,"text":84},"32410048","And for the true \"Show HN\" experience: wget \"https:\u002F\u002Fgithub.com\u002FDeutscheKI\u002Ftevr-asr-tool\u002Freleases\u002Fdownload\u002Fv1.0.0\u002Ftevr_asr_tool-1.0.0-Linux-x86_64.zip.001\" wget…",{"value":86,"source":52,"evidence":87},"KenLM",[88],{"id":89,"by":11,"text":90},"32412935","I just used the Linux CLOC on the C++ files that I wrote. Yes, I didn't count my 3 dependencies KenLM, Wave, or TensorFlow, because those are used by pretty much all speech…","open_source","launched","solo",{"praise":95,"objections":114,"requests":152,"questions":174,"alternatives":189},[96,107],{"value":97,"source":98,"evidence":99},"Impressive compact implementation compared to DeepSpeech","commenter",[100,103],{"id":101,"by":11,"text":102},"32419351","If you compare this with Mozilla's DeepSpeech repository, you will find that correctly calling TensorFlow Lite and handling the results in only 284 lines of code is impressive,…",{"id":104,"by":105,"text":106},"32411193","lynndotpy","To those criticizing the title, how would you improve it? Dependencies do the heavy lifting , but that's true even for a hello-world. It seems obvious enough that I wouldn't see…",{"value":108,"source":98,"evidence":109},"Readable source code",[110],{"id":111,"by":112,"text":113},"32412897","einpoklum","I think the author should probably use some C++ \u003Calgorithm>'s in lines 241 and later, avoiding raw loops and making the code clearer. Also, yeah, it's not 284 lines of C++ by any…",[115,134,141],{"value":116,"source":98,"evidence":117},"284 line count ignores large dependencies and model weights",[118,122,126,130],{"id":119,"by":120,"text":121},"32410952","postalrat","And 27 includes plus a huge model. \"284 lines of C++\" is something you could fit on a small microcontroller. This isn't 284 lines of C++.",{"id":123,"by":124,"text":125},"32416056","woah","284 lines of C++ and how many trained weights? Touting the \"low number of lines\" on a neural network project seems kind of silly, since the logic is encoded in the weights. Kind…",{"id":127,"by":128,"text":129},"32416133","codeflo","I don’t want to take anything away from the achievement, because this looks very useful, but the headline misled me a bit. I’m note sure whether it’s my affinity for code golf,…",{"id":131,"by":132,"text":133},"32410984","mrweasel","I don't know. It's not that it isn't impressive and it is very illustrative of how much you can build, utilizing the tools that are available to all of us. It's just at what…",{"value":135,"source":98,"evidence":136},"Possible LM overtraining inflates results",[137],{"id":138,"by":139,"text":140},"32413636","nshm","This model seems strongly overtrained on CV test set. Usually improvement from LM rescoring is just 10% relative. In the paper https:\u002F\u002Farxiv.org\u002Fpdf\u002F2206.12693.pdf the…",{"value":142,"source":98,"evidence":143},"Title misleading",[144,148],{"id":145,"by":146,"text":147},"32418973","pkaye","Includes a standard C++ library called tensorflow...",{"id":149,"by":150,"text":151},"32410981","cgeier","I was quite interested in this, but after seeing that it includes three big other repositories, my interest was immediately gone. I'm not saying this isn't an interesting…",[153,160,167],{"value":154,"source":98,"evidence":155},"Streaming or real-time recognition support",[156],{"id":157,"by":158,"text":159},"32412670","a2128","This is very impressive work. Could this model be capable of doing streamed input for live speech recognition?",{"value":161,"source":98,"evidence":162},"Accuracy on accents and dialects",[163],{"id":164,"by":165,"text":166},"32410694","BurningPenguin","How well does it do with accent or slight dialect influences?",{"value":168,"source":98,"evidence":169},"Accuracy on tongue-twister-like sounds",[170],{"id":171,"by":172,"text":173},"32414129","maratc","I'd like to know its accuracy on the sound of Rhabarberbarbara[0] or Nackenzacken[1]. [0]https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=l3_tRPRt9x8 [1]https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=Mngm8k4_qAc",[175,182],{"value":176,"source":98,"evidence":177},"How does it compare to BPE tokenization",[178],{"id":179,"by":180,"text":181},"32412544","gok","The \"TEVR\" token encoding idea is interesting. Any chance you compared it with other subword tokenization techniques like BPE?",{"value":183,"source":98,"evidence":184},"Can it be statically linked for embedded use",[185],{"id":186,"by":187,"text":188},"32411645","fwsgonzo","Is it possible to get speech recognition into a statically linked executable? ldd myprogram should say \"not a dynamic executable\" Are there any libraries for those of us in…",[190,194,200,204],{"value":191,"source":98,"evidence":192},"Mozilla DeepSpeech",[193],{"id":55,"by":11,"text":56},{"value":195,"source":98,"evidence":196},"Vosk",[197],{"id":198,"by":11,"text":199},"32414122","Also in Table 6, you see that Facebook's wav2vec 2.0 XLS-R went from 12.06% without LM to 4.38% with 5-gram LM. In comparison to that, I found TEVR going from 10.10% to 3.64%…",{"value":201,"source":98,"evidence":202},"wav2vec2 XLS-R",[203],{"id":198,"by":11,"text":199},{"value":205,"source":98,"evidence":206},"Kaldi",[207],{"id":55,"by":11,"text":56},{"comments":209,"answered":210,"key":211},42,17,[212,213,214],{"id":83,"text":84},{"id":198,"text":199},{"id":215,"text":216},"32426820","I now had time and did some testing and the CER is already pretty much excellent for TEVR even without the language model, so it appears to me that what the LM does is mostly to…",{"commenters":218,"max_depth":219,"maker_replies":209,"first_comment_mins":219,"lifespan_hours":220},60,12,77,[222,224,226,230,232,235,237,240,241,245,248,252,254,257,259,263,265,267,270,272,275,278,280,284,287,289,291,294,295,297,299,301,303,305,307,310,312,314,317,320,324,328,330,331,334,336,338,341,344,346,350,353,357,358,361,363,365,368,370,372,376,379,383,387,388,391,394,397,399,401,403,405,407,409,411,413,415,417,419,420,423,425,427,430,432,435,437,439,441,443,446,450,452,455,457,459,462,463,465,468,471,473,475,478,480,482,484,486,487,490,492,494,495,498,501,503,507,509,512,513,516,518,521,524,528,532,536,540],[83,40,40,223,11,45],"project\u002Fneutral\u002Fdocs",[138,45,45,225,139,40],"project\u002Fcritical\u002Fperformance",[227,45,40,228,229,40],"32413993","commenter\u002Fneutral\u002Fnone","rob74",[231,45,45,225,139,40],"32414114",[233,45,40,228,234,40],"32419622","Evidlo",[198,45,40,236,11,45],"project\u002Fpositive\u002Fperformance",[238,45,33,239,139,40],"32414916","maker\u002Fpositive\u002Fperformance",[215,45,40,236,11,45],[242,38,42,243,244,40],"32411038","project\u002Fneutral\u002Fpricing","kevin_b_er",[246,38,40,247,11,45],"32411826","project\u002Fneutral\u002Fidea",[249,38,42,250,251,40],"32419098","project\u002Fneutral\u002Fnone","mciancia",[253,38,40,250,11,45],"32419169",[255,38,42,250,256,40],"32428733","gogusrl",[258,38,40,250,11,45],"32429233",[260,42,33,261,262,40],"32410931","project\u002Fpositive\u002Fidea","captainmuon",[264,42,40,247,11,45],"32412099",[157,266,42,250,158,40],4,[268,266,40,269,11,45],"32412797","project\u002Fneutral\u002Fperformance",[271,266,42,250,158,40],"32413539",[273,266,40,274,11,45],"32418736","other\u002Fneutral\u002Fnone",[276,266,42,243,277,40],"32418095","phkahler",[279,266,40,243,11,45],"32418633",[281,266,42,282,283,40],"32419186","maker\u002Fneutral\u002Fnone","Birch-san",[285,266,40,286,11,45],"32419234","commenter\u002Fpositive\u002Fnone",[58,33,40,288,11,45],"other\u002Fpositive\u002Fnone",[171,290,42,269,172,40],6,[292,290,33,236,293,40],"32416714","marshray",[70,290,40,269,11,45],[296,290,40,228,293,40],"32427598",[179,298,42,247,180,40],7,[300,298,40,247,11,45],"32413027",[302,298,42,247,180,40],"32413589",[164,304,42,269,165,40],8,[306,304,40,236,11,45],"32411970",[308,304,42,269,309,40],"32412482","junkerm",[311,304,40,223,11,45],"32412623",[186,18,42,313,187,40],"project\u002Fneutral\u002Fexecution",[315,18,40,316,11,45],"32411753","project\u002Fpositive\u002Fexecution",[318,18,33,319,187,40],"32417727","project\u002Fpositive\u002Fnone",[321,322,42,250,323,40],"32419526",10,"amelius",[325,326,40,274,327,40],"32410873",11,"tjungblut",[119,219,45,329,120,40],"project\u002Fcritical\u002Fnone",[55,219,40,250,11,45],[332,219,40,228,333,40],"32411976","kylebgorman",[335,219,40,228,11,45],"32412149",[337,219,40,228,120,40],"32412366",[339,219,40,228,340,40],"32412818","dr_dshiv",[342,219,40,228,343,40],"32413905","georgia_peach",[345,219,40,274,11,45],"32414429",[347,219,33,348,349,40],"32414466","maker\u002Fpositive\u002Fnone","dataflow",[351,219,40,274,352,40],"32414616","xdfgh1112",[354,219,33,355,356,40],"32420915","project\u002Fpositive\u002Fdocs","jpetso",[123,43,45,329,124,40],[359,43,45,329,360,40],"32416252","munk-a",[362,43,40,228,11,45],"32418691",[364,43,40,247,11,45],"32417086",[366,43,40,274,367,40],"32418005","joe__f",[369,43,40,274,11,45],"32418246",[371,43,40,274,367,40],"32418285",[373,374,40,274,375,40],"32410899",14,"dahfizz",[377,374,40,274,378,40],"32412614","petercooper",[380,381,40,228,382,40],"32411181",15,"henrydark",[384,381,45,385,386,40],"32419032","project\u002Fcritical\u002Fidea","prewett",[101,381,40,316,11,45],[127,389,42,390,128,40],16,"project\u002Fmixed\u002Fnone",[392,389,45,385,393,40],"32416788","thaumasiotes",[395,389,40,228,396,40],"32417597","eurasiantiger",[398,389,40,228,393,40],"32417637",[400,389,40,228,396,40],"32417659",[402,389,40,228,11,45],"32419257",[404,389,40,228,393,40],"32420296",[406,389,40,247,11,45],"32423029",[408,389,40,274,396,40],"32423319",[410,389,40,228,393,40],"32430673",[412,389,40,228,396,40],"32434987",[414,389,40,228,393,40],"32451083",[416,389,40,228,128,40],"32419689",[418,389,40,247,11,45],"32417118",[104,210,33,261,105,40],[421,210,42,250,422,40],"32411432","itcrowd",[424,210,33,316,105,40],"32411732",[426,210,40,313,11,45],"32411880",[428,210,45,385,429,40],"32412814","UncleEntity",[431,210,40,269,11,45],"32412903",[433,210,42,434,429,40],"32417173","project\u002Fmixed\u002Fidea",[436,210,40,228,105,40],"32413802",[438,210,40,274,429,40],"32417108",[440,210,45,385,229,40],"32414039",[442,210,40,247,11,45],"32419047",[444,210,40,228,445,40],"32414495","glouwbug",[447,448,45,385,449,40],"32410684",18,"OrangeMonkey",[451,448,45,385,132,40],"32410750",[453,448,40,228,454,40],"32410842","hardware2win",[456,448,45,385,375,40],"32410891",[458,448,40,236,11,45],"32412403",[460,448,45,385,461,40],"32413270","agileAlligator",[131,448,45,385,132,40],[464,448,40,247,11,45],"32411619",[466,448,42,247,467,40],"32411959","andsoitis",[469,448,40,228,470,40],"32414658","Someone",[145,472,45,385,146,40],19,[474,472,40,247,11,45],"32419030",[476,472,45,385,477,40],"32419101","numpad",[479,472,40,247,11,45],"32419147",[481,472,42,434,477,40],"32419249",[483,472,42,223,234,40],"32419682",[485,472,40,247,11,45],"32423041",[149,32,45,385,150,40],[488,32,42,247,489,40],"32411012","eternalban",[491,32,40,261,11,45],"32411931",[493,32,33,261,150,40],"32412277",[64,32,40,316,11,45],[496,32,42,269,497,40],"32414250","tialaramex",[499,32,40,274,500,40],"32415012","yorwba",[502,32,40,269,11,45],"32423060",[504,505,45,385,506,40],"32417214",21,"olliej",[508,505,40,236,11,45],"32418662",[111,510,42,511,112,40],22,"project\u002Fmixed\u002Fexecution",[89,510,40,313,11,45],[514,510,45,515,112,40],"32413704","project\u002Fcritical\u002Fexecution",[517,510,40,228,11,45],"32413918",[519,34,42,390,520,40],"32410969","vintermann",[522,34,40,523,11,45],"32411513","project\u002Fneutral\u002Fdesign",[525,526,42,250,527,40],"32410794",24,"nottorp",[529,530,45,515,531,40],"32411707",25,"ale42",[533,534,45,515,535,40],"32415679",26,"nice_byte",[537,538,45,515,539,40],"32412830",27,"snickerbockers",[541,542,45,515,543,40],"32418260",28,"CamperBob2",[],[546,551,556,560,564,568],{"id":547,"name":548,"points":549,"category":550},"32376154","Figure",417,"games",{"id":552,"name":553,"points":554,"category":555},"32348461","Penumbra",267,"dev_tools",{"id":557,"name":558,"points":559,"category":555},"32344671","Superblocks",193,{"id":561,"name":562,"points":563,"category":555},"32352291","RocketRide",176,{"id":565,"name":566,"points":567,"category":555},"32353479","PDF-Diff",169,{"id":569,"name":570,"points":571,"category":550},"32342820","Killer Crossword",117,false,"v2",{"id":575,"name":5,"line":6,"ai":572,"cat":20,"year":13,"points":14,"comments":15,"rec":29,"s":576,"top":18,"kw":577,"maker":11,"domain":10,"tier":19,"attempts":40,"strip":578,"url":9,"mix":579,"praise":97,"objection":116,"sub":21},32409966,-0.29,[24,25,26,27,28],"1153533353335105200110005331111111531131",[18,32,33,34],1793991050019]