[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"blog-\u002Fhe\u002Fblog\u002Fvoix-temps-reel-vs-batch":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"author":10,"date":11,"tags":12,"body":16,"_type":514,"_id":515,"_source":516,"_file":517,"_stem":518,"_extension":519},"\u002Fhe\u002Fblog\u002Fvoix-temps-reel-vs-batch","blog",false,"","סוכן AI טלפוני: קול בזמן אמת או batch?","מדוע עוזר קולי דורש ארכיטקטורה שונה מהותית מ-chatbot כתוב, ומה זה מחייב לגבי זמן השהייה, barge-in ובחירת המודל.","Équipe betool","2026-04-28",[13,14,15],"קול","ארכיטקטורה","זמן-השהייה",{"type":17,"children":18,"toc":505},"root",[19,40,45,52,71,84,89,95,105,124,129,200,205,210,244,249,261,273,283,306,311,317,322,327,380,386,391,434,439,445,450,493],{"type":20,"tag":21,"props":22,"children":23},"element","p",{},[24,27,33,35],{"type":25,"value":26},"text","כשספק מציע ",{"type":20,"tag":28,"props":29,"children":30},"em",{},[31],{"type":25,"value":32},"«סוכן AI טלפוני»",{"type":25,"value":34},", שאלו אותו שאלה אחת: ",{"type":20,"tag":28,"props":36,"children":37},{},[38],{"type":25,"value":39},"«מהו זמן ההשהייה end-to-end מרגע שהמשתמש מפסיק לדבר ועד שמתחילה התגובה?»",{"type":20,"tag":21,"props":41,"children":42},{},[43],{"type":25,"value":44},"אם התשובה עולה על 1.5 שנייה — זה לא סוכן טלפוני. זה chatbot עם מודול ASR \u002F TTS מודבק לפניו.",{"type":20,"tag":46,"props":47,"children":49},"h2",{"id":48},"כלל-ה-800-ms",[50],{"type":25,"value":51},"כלל ה-800 ms",{"type":20,"tag":21,"props":53,"children":54},{},[55,57,62,64,69],{"type":25,"value":56},"אדם בטלפון סובל שתיקה לאחר שאלתו. אבל לשתיקה יש גבול: מעבר ל-800 ms, הצד השני נראה ",{"type":20,"tag":28,"props":58,"children":59},{},[60],{"type":25,"value":61},"«קפוא»",{"type":25,"value":63},". ב-1,500 ms, המשתמש אומר ",{"type":20,"tag":28,"props":65,"children":66},{},[67],{"type":25,"value":68},"«הלו?»",{"type":25,"value":70},". ב-3,000 ms הוא כבר מנתק וחושב על תקלת קו.",{"type":20,"tag":21,"props":72,"children":73},{},[74,76,82],{"type":25,"value":75},"המטרה של כל סוכן קולי רציני: ",{"type":20,"tag":77,"props":78,"children":79},"strong",{},[80],{"type":25,"value":81},"פחות מ-800 ms",{"type":25,"value":83}," בין סוף הדיבור לתחילת התגובה הנשמעת.",{"type":20,"tag":21,"props":85,"children":86},{},[87],{"type":25,"value":88},"נפרק את זה.",{"type":20,"tag":46,"props":90,"children":92},{"id":91},"תקציב-הזמן-בפירוט",[93],{"type":25,"value":94},"תקציב הזמן בפירוט",{"type":20,"tag":96,"props":97,"children":99},"pre",{"code":98},"[parole user]──▶[ASR streaming]──▶[LLM]──▶[TTS streaming]──▶[audio out]\n       ~150ms          ~80ms       ~?      ~250ms             ~50ms\n",[100],{"type":20,"tag":101,"props":102,"children":103},"code",{"__ignoreMap":7},[104],{"type":25,"value":98},{"type":20,"tag":21,"props":106,"children":107},{},[108,110,115,117,122],{"type":25,"value":109},"מתוך ~800 ms הזמינים, ",{"type":20,"tag":77,"props":111,"children":112},{},[113],{"type":25,"value":114},"כ-270 ms הולכים ל-ASR + TTS + העברה",{"type":25,"value":116},". נשארים ",{"type":20,"tag":77,"props":118,"children":119},{},[120],{"type":25,"value":121},"530 ms ל-LLM",{"type":25,"value":123}," — כולל prefill, thinking וייצור הטוקן הראשון.",{"type":20,"tag":21,"props":125,"children":126},{},[127],{"type":25,"value":128},"השלכות מעשיות:",{"type":20,"tag":130,"props":131,"children":132},"ul",{},[133,151,167,183],{"type":20,"tag":134,"props":135,"children":136},"li",{},[137,142,144,149],{"type":20,"tag":77,"props":138,"children":139},{},[140],{"type":25,"value":141},"מודל reasoning גדול, ב-non-streaming",{"type":25,"value":143},": כמה שניות לפני המילה הנשמעת הראשונה. ",{"type":20,"tag":77,"props":145,"children":146},{},[147],{"type":25,"value":148},"מחוץ למשחק",{"type":25,"value":150},".",{"type":20,"tag":134,"props":152,"children":153},{},[154,159,161,166],{"type":20,"tag":77,"props":155,"children":156},{},[157],{"type":25,"value":158},"מודל high-end ב-streaming",{"type":25,"value":160},": טוקן ראשון מהיר ביום טוב, הרבה פחות ביום רע. ",{"type":20,"tag":77,"props":162,"children":163},{},[164],{"type":25,"value":165},"קביל אבל מתוח",{"type":25,"value":150},{"type":20,"tag":134,"props":168,"children":169},{},[170,175,177,182],{"type":20,"tag":77,"props":171,"children":172},{},[173],{"type":25,"value":174},"מודל מהיר ב-streaming",{"type":25,"value":176}," (הסדרות ה«קלות» של Claude, GPT ו-Gemini): ",{"type":20,"tag":77,"props":178,"children":179},{},[180],{"type":25,"value":181},"נוח",{"type":25,"value":150},{"type":20,"tag":134,"props":184,"children":185},{},[186,191,193,198],{"type":20,"tag":77,"props":187,"children":188},{},[189],{"type":25,"value":190},"מודל open-weight קטן המוגש על ה-GPU שלכם (vLLM)",{"type":25,"value":192},": ",{"type":20,"tag":77,"props":194,"children":195},{},[196],{"type":25,"value":197},"מצוין",{"type":25,"value":199}," — ופרטי.",{"type":20,"tag":21,"props":201,"children":202},{},[203],{"type":25,"value":204},"מדדו את ה-TTFT על התעבורה שלכם: הוא משתנה לפי הספק, האזור והשעה.",{"type":20,"tag":21,"props":206,"children":207},{},[208],{"type":25,"value":209},"ארכיטקטורת קול רצינית מחייבת לכן:",{"type":20,"tag":211,"props":212,"children":213},"ol",{},[214,224,234],{"type":20,"tag":134,"props":215,"children":216},{},[217,222],{"type":20,"tag":77,"props":218,"children":219},{},[220],{"type":25,"value":221},"הכל ב-streaming",{"type":25,"value":223}," — ASR, LLM, TTS. אף חוליה אחת ב-batch.",{"type":20,"tag":134,"props":225,"children":226},{},[227,232],{"type":20,"tag":77,"props":228,"children":229},{},[230],{"type":25,"value":231},"מודל LLM נבחר לפי TTFT",{"type":25,"value":233}," (time to first token), לא לפי ציון MMLU.",{"type":20,"tag":134,"props":235,"children":236},{},[237,242],{"type":20,"tag":77,"props":238,"children":239},{},[240],{"type":25,"value":241},"TTS מתחיל לדבר מהטוקן הראשון שהתקבל",{"type":25,"value":243},", לא עם סיום התגובה המלאה.",{"type":20,"tag":46,"props":245,"children":247},{"id":246},"ה-barge-in",[248],{"type":25,"value":246},{"type":20,"tag":21,"props":250,"children":251},{},[252,254,259],{"type":25,"value":253},"בערוץ כתוב, המשתמש ממתין שהבוט יסיים. בטלפון — ",{"type":20,"tag":77,"props":255,"children":256},{},[257],{"type":25,"value":258},"הוא מפריע",{"type":25,"value":260},". שלוש מתוך חמש פעמים בשיחות טבעיות.",{"type":20,"tag":21,"props":262,"children":263},{},[264,266,271],{"type":25,"value":265},"סוכן קולי שאינו יודע להיות מופרע נשמע ",{"type":20,"tag":28,"props":267,"children":268},{},[269],{"type":25,"value":270},"«רובוטי לגמרי»",{"type":25,"value":272}," — הוא ממשיך לגלגל את המשפט בזמן שהמשתמש מנסה להפריע. זה בלתי-נסבל ומסגיר מיד שמדברים עם מכונה גרועה.",{"type":20,"tag":21,"props":274,"children":275},{},[276,281],{"type":20,"tag":77,"props":277,"children":278},{},[279],{"type":25,"value":280},"barge-in",{"type":25,"value":282}," טכני מחייב:",{"type":20,"tag":130,"props":284,"children":285},{},[286,291,296,301],{"type":20,"tag":134,"props":287,"children":288},{},[289],{"type":25,"value":290},"TTS מפסיק מיד כאשר ASR מזהה אות קולי מהמשתמש.",{"type":20,"tag":134,"props":292,"children":293},{},[294],{"type":25,"value":295},"LLM זונח את ההשלמה הנוכחית ללא היסוס.",{"type":20,"tag":134,"props":297,"children":298},{},[299],{"type":25,"value":300},"ה-buffer האודיו היוצא מנוקה מיידית (אחרת שומעים הברה-חצי שיורית).",{"type":20,"tag":134,"props":302,"children":303},{},[304],{"type":25,"value":305},"היסטוריית השיחה מציינת שהתגובה לא נשמעה במלואה — כדי שלא לחזור עליה בטעות.",{"type":20,"tag":21,"props":307,"children":308},{},[309],{"type":25,"value":310},"פרט טכני אחד שמשנה את כל החוויה.",{"type":20,"tag":46,"props":312,"children":314},{"id":313},"ניהול-אי-ודאות",[315],{"type":25,"value":316},"ניהול אי-ודאות",{"type":20,"tag":21,"props":318,"children":319},{},[320],{"type":25,"value":321},"בטלפון, ASR טועה לא מעט. בעוד שחלון צ'אט תמיד מקבל את הטקסט המדויק שהמשתמש הקליד, הסוכן הקולי מקבל תמלול שכיח שיהיה בלתי-מושלם.",{"type":20,"tag":21,"props":323,"children":324},{},[325],{"type":25,"value":326},"אסטרטגיות שעובדות:",{"type":20,"tag":130,"props":328,"children":329},{},[330,340,364],{"type":20,"tag":134,"props":331,"children":332},{},[333,338],{"type":20,"tag":77,"props":334,"children":335},{},[336],{"type":25,"value":337},"בקשת אישור",{"type":25,"value":339}," על רכיבים בעלי השלכות משמעותיות (מספר תיק, סכום, שם משפחה יוצא דופן). לא באופן שיטתי — אחרת זה מייגע.",{"type":20,"tag":134,"props":341,"children":342},{},[343,348,350,355,357,362],{"type":20,"tag":77,"props":344,"children":345},{},[346],{"type":25,"value":347},"אל תסמכו עיוורונית על ה-LLM",{"type":25,"value":349}," כשה-ASR אמביגואלי. אם המשפט שהתקבל הוא ",{"type":20,"tag":28,"props":351,"children":352},{},[353],{"type":25,"value":354},"«אני רוצה לבטל את החוזה שלי»",{"type":25,"value":356}," עם ציון ביטחון נמוך, עדיף לשאול ",{"type":20,"tag":28,"props":358,"children":359},{},[360],{"type":25,"value":361},"«התכוונת לבטל את החוזה?»",{"type":25,"value":363}," מאשר להפעיל את הליך הביטול.",{"type":20,"tag":134,"props":365,"children":366},{},[367,372,374,379],{"type":20,"tag":77,"props":368,"children":369},{},[370],{"type":25,"value":371},"גזרו סטיות בשורש",{"type":25,"value":373}," — סוכן שחורג מהסקריפט חייב להיות מוחזר על ידי צומת condition או מפעיל, לא על ידי פרומפט שאמור ",{"type":20,"tag":28,"props":375,"children":376},{},[377],{"type":25,"value":378},"«להספיק»",{"type":25,"value":150},{"type":20,"tag":46,"props":381,"children":383},{"id":382},"הטלפון-הוא-גם-stack",[384],{"type":25,"value":385},"הטלפון הוא גם stack",{"type":20,"tag":21,"props":387,"children":388},{},[389],{"type":25,"value":390},"מעבר ל-AI, סוכן טלפוני דורש:",{"type":20,"tag":130,"props":392,"children":393},{},[394,404,414,424],{"type":20,"tag":134,"props":395,"children":396},{},[397,402],{"type":20,"tag":77,"props":398,"children":399},{},[400],{"type":25,"value":401},"trunk SIP",{"type":25,"value":403}," אצל ספק (Twilio, Voxbone, OVH, Sewan…).",{"type":20,"tag":134,"props":405,"children":406},{},[407,412],{"type":20,"tag":77,"props":408,"children":409},{},[410],{"type":25,"value":411},"media gateway",{"type":25,"value":413}," (LiveKit-SIP, FreeSWITCH, Asterisk) שמדברת SIP בצד הספק ו-WebRTC בצד הפלטפורמה.",{"type":20,"tag":134,"props":415,"children":416},{},[417,422],{"type":20,"tag":77,"props":418,"children":419},{},[420],{"type":25,"value":421},"מתמלל",{"type":25,"value":423}," עמיד לרעש, מבטאים וחפיפות קוליות.",{"type":20,"tag":134,"props":425,"children":426},{},[427,432],{"type":20,"tag":77,"props":428,"children":429},{},[430],{"type":25,"value":431},"מסנתז",{"type":25,"value":433}," שמבטא כהלכה שמות פרטיים, מספרים ותאריכים.",{"type":20,"tag":21,"props":435,"children":436},{},[437],{"type":25,"value":438},"אף אחד מהרכיבים הללו אינו פשוט. כולם מקיימים אינטראקציה ביניהם.",{"type":20,"tag":46,"props":440,"children":442},{"id":441},"מה-betool-מטפלת-בו",[443],{"type":25,"value":444},"מה betool מטפלת בו",{"type":20,"tag":21,"props":446,"children":447},{},[448],{"type":25,"value":449},"ארכיטקטורת הקול של betool — LiveKit + LiveKit-SIP + worker ייעודי + LLM ב-streaming אל סינתזת הדיבור — מכסה את כל הצרכים האלה כברירת מחדל. אתם נשארים אדוני:",{"type":20,"tag":130,"props":451,"children":452},{},[453,464,482],{"type":20,"tag":134,"props":454,"children":455},{},[456,458,462],{"type":25,"value":457},"ה",{"type":20,"tag":77,"props":459,"children":460},{},[461],{"type":25,"value":401},{"type":25,"value":463}," (הספק שלכם, המספרים שלכם).",{"type":20,"tag":134,"props":465,"children":466},{},[467,468,473,475,480],{"type":25,"value":457},{"type":20,"tag":77,"props":469,"children":470},{},[471],{"type":25,"value":472},"מודלים",{"type":25,"value":474}," (מודלים מנוהלים, BYOK או מודלים פרטיים) וה",{"type":20,"tag":77,"props":476,"children":477},{},[478],{"type":25,"value":479},"קולות",{"type":25,"value":481}," (תמלול Deepgram או faster-whisper; סינתזה ElevenLabs, Cartesia, OpenAI או Deepgram Aura).",{"type":20,"tag":134,"props":483,"children":484},{},[485,486,491],{"type":25,"value":457},{"type":20,"tag":77,"props":487,"children":488},{},[489],{"type":25,"value":490},"פרמטרים",{"type":25,"value":492}," (רגישות ה-barge-in, זיהוי סוף תור, משפטי המתנה, העברה לנציג אנושי).",{"type":20,"tag":21,"props":494,"children":495},{},[496,498,503],{"type":25,"value":497},"את ה-plumbing של זמן האמת אתם לא מגדירים — זה עבודת הפלטפורמה. אתם מעצבים את ",{"type":20,"tag":77,"props":499,"children":500},{},[501],{"type":25,"value":502},"השיחה",{"type":25,"value":504},", שזה הנושא העסקי האמיתי.",{"title":7,"searchDepth":506,"depth":506,"links":507},2,[508,509,510,511,512,513],{"id":48,"depth":506,"text":51},{"id":91,"depth":506,"text":94},{"id":246,"depth":506,"text":246},{"id":313,"depth":506,"text":316},{"id":382,"depth":506,"text":385},{"id":441,"depth":506,"text":444},"markdown","content:he:blog:2.voix-temps-reel-vs-batch.md","content","he\u002Fblog\u002F2.voix-temps-reel-vs-batch.md","he\u002Fblog\u002F2.voix-temps-reel-vs-batch","md"]