{
  "video": {
    "id": "IDNfAZVKvPE",
    "title": "\"My name is... my name is...\": A Linguistic Map for Voice Agents — Midam Kim, ServiceNow",
    "duration": 905,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "IDNfAZVKvPE",
    "title": "\"My name is... my name is...\": A Linguistic Map for Voice Agents — Midam Kim, ServiceNow",
    "one_liner": "A ServiceNow ML engineer and linguist hands you a 2×4 linguistic map — listening vs. speaking channels crossed with sounds, words, interaction and mental model — as a diagnostic grid for why your voice agent fails and where in the stack to fix it.",
    "summary": "Midam Kim opens with a real call in which a bot mishears M as N when she spells her name, TTS reads 'M I D A M' with English-centric rules and calls her 'Madam', asks for an unexplained 'account number', then cuts her off mid-digits and asks her to repeat — until she escalates to a human. She argues that human communication is a joint activity in which both parties continuously update mental models, that users expect the same of bots, and that every failure in that call lands in a specific cell of a linguistic framework: two channels (listening, speaking) × four interdependent components (sounds, words, interaction, mental model). The crucial asymmetry with chat is that voice leaves no transcript for the user: sounds, words and interaction vanish the moment they're spoken, so only the user's growing mental model persists — and that is what you must target. She offers no fix, only the framework plus ServiceNow's end-to-end Eva bench for diagnosis, and tells engineers to learn linguistics and hire linguists.",
    "key_points": [
      "The worked failure case: bot confuses M and N spelling 'm i d a m' (listening/sounds STT failure), TTS then applies English-centric reading rules and says 'Madam' (speaking/sounds), 'account number' is vocabulary not shared with the user (speaking/words), the bot talks over her while she reads 'A X 4 5 1' slowly (interaction), and then asks her to repeat the same information with no interactive clarification — proving it isn't tracking a mental model with her.",
      "The framework is a 2×4 grid: listening channel asks does the bot recognize the user's speech, understand the user's words, wait until the right timing for its turn, and understand the user's intention; speaking channel asks about pronunciation, choosing words the user can understand, speaking with the right timing, and speaking the information the user actually needs.",
      "The components are interdependent, not independent, and must be aligned — good work on sounds requires thinking at the words level, which requires accounting for interaction/turn-taking, which is required for task completion at the mental-model layer; missing any one of them makes the agent fail.",
      "Voice differs from chat because there is no visible history: 'sounds, words, interactions vanish the moment they're spoken, but the mental model proceeds and grows over the timeline' — the user's mental model is the only thing left and the real target for satisfaction.",
      "Concrete levers named: ASR model/configuration choice plus post-processing; TTS model/configuration choice plus pre-processing; carefully curated vocabulary shared between bot and user; turn detection, latency and turn-taking; emotion detection and handling; and context retention across all of these.",
      "This management has to be dynamic along the call timeline and across different kinds of people (she names kids as an example), and must hold when users are unhappy, not just when they're happy — that's the precondition for truly scalable orchestration.",
      "Business implications are linguistic implications and vice versa: attention to the framework reduces user frustration, task failures, live agent escalation, abandoned calls and silent failures.",
      "ServiceNow has built an end-to-end benchmark called Eva bench that you can use to diagnose your voice agent's status.",
      "Closing long-term questions: speakers always adapt, so users will adapt to your agent over the course of a call — is your system ready for them to use it better next time? And language always changes — is your agent ready for language change in a year, or even six months?"
    ],
    "takeaways": [
      "Use the 2×4 grid (listen/speak × sounds/words/interaction/mental model) as a diagnostic checklist: replay a failed call turn by turn and locate which cell broke, rather than blaming 'the model'.",
      "Stop treating the voice stack as a pipeline of independent components — tune ASR, TTS, vocabulary and turn-taking together, because gains in one are erased by misalignment in another.",
      "Instrument and optimize for the user's persisting mental model, not per-turn accuracy: never ask for a repeat of information already given, and use interactive clarification (the human strategy) instead of a blanket 'can you repeat that'.",
      "Curate the vocabulary the bot uses so it is actually shared with the user — terms like 'account number' that the user can't map to anything create confusion the bot then blames on the user.",
      "Handle non-English names and spelling explicitly: don't apply English-centric TTS reading rules, and treat confusable letter pairs like M/N as a known STT failure mode needing post-processing.",
      "Run ServiceNow's Eva bench end-to-end to diagnose your agent, and beyond tooling, learn linguistics and hire linguists."
    ],
    "topics": [
      "voice-agents",
      "linguistics",
      "asr",
      "tts",
      "turn-taking",
      "conversation-design",
      "benchmarks",
      "ux"
    ],
    "tools": [
      "ServiceNow",
      "Eva bench",
      "ASR/STT",
      "TTS"
    ],
    "quotes": [
      {
        "text": "Thank you for your correction. Happy to help you today, Madam.",
        "at": "01:44",
        "url": "https://www.youtube.com/watch?v=IDNfAZVKvPE&t=104s"
      },
      {
        "text": "So, sounds, words, interactions vanish the moment they're spoken, but the mental model proceeds and grows over the timeline.",
        "at": "10:00",
        "url": "https://www.youtube.com/watch?v=IDNfAZVKvPE&t=600s"
      },
      {
        "text": "We always say that voice is the most natural way of communication, but it is not actually not easy.",
        "at": "11:48",
        "url": "https://www.youtube.com/watch?v=IDNfAZVKvPE&t=708s"
      },
      {
        "text": "You can try Eva, but also you can learn linguistics and hire linguists.",
        "at": "13:21",
        "url": "https://www.youtube.com/watch?v=IDNfAZVKvPE&t=801s"
      }
    ],
    "words": 2218
  },
  "summary_url": "/#IDNfAZVKvPE",
  "transcript": {
    "html": "/transcripts/IDNfAZVKvPE.html",
    "txt": "/transcripts/IDNfAZVKvPE.txt",
    "vtt": "/transcripts/IDNfAZVKvPE.vtt"
  }
}