Midam Kim

Midam Kim is an ML engineer and linguist at ServiceNow whose work applies a linguistic perspective to enterprise voice AI. She also describes herself as a longtime researcher of speech communication “in the wild.” Her personal website identifies her as Midam Kim, PhD and carries the heading “Senior Content Analyst for Generative AI @ ServiceNow,” while her more recent public appearances identify her as an ML engineer and linguist.

A linguistic map for voice agents

In “My name is... my name is...”: A Linguistic Map for Voice Agents, Kim begins with a call she experienced. A voice agent confused the final letter of her spelled name and then pronounced the name incorrectly after she corrected it. The agent subsequently requested an account number without making clear which identifier it needed. While she slowly read the unfamiliar string, it interrupted her, failed to find the record, and asked her to repeat the same information without changing its approach. The accumulated failures led her to request a human agent.

Kim uses the example to show that a frustrating call can contain several distinct technical and linguistic failures. Recognition, pronunciation, vocabulary, timing, intent, and conversational repair may be evaluated separately, but the user experiences them as one deteriorating interaction.

• Communication as a joint activity. Kim argues that voice interaction is not merely a pipeline that receives speech and produces an answer. Human conversation requires both parties to listen, respond, coordinate turns, and continually update their understanding of the exchange. A voice agent must participate in that coordination.

• Listening and speaking across four levels. Her framework crosses two channels—listening and speaking—with four interdependent levels: sounds, words, interaction, and mental models. On the listening side, an agent must recognize speech, understand the user’s words, wait until an appropriate time to take its turn, and infer what the user intends. On the speaking side, it must pronounce speech appropriately, choose understandable words, respond at the right time, and provide the information the user actually needs.

• Interdependent components. Kim stresses that these responsibilities cannot be optimized independently. Accurate transcription does not prevent an agent from interrupting the user, and correctly storing a name’s letters does not guarantee appropriate pronunciation. Successful task completion depends on recognition, language, timing, and intent tracking remaining aligned throughout the exchange.

• Conversational repair. Acknowledging a correction is not enough if the agent’s subsequent behavior still reflects the original mistake. Likewise, asking for an identical repetition without clarifying the source of uncertainty can deepen frustration. Kim points to interactive clarification as a human repair strategy that voice agents should be able to use.

• Task-sensitive turn-taking. Someone reading an unfamiliar identifier may pause while locating or decoding its next characters. Treating every pause as the end of a turn can cause an agent to act on incomplete information. Kim therefore connects turn detection and latency with recognition, vocabulary, and the task the person is trying to complete.

• The user’s accumulating mental model. Unlike a text chat, a spoken exchange generally leaves no visible history for the user to inspect. The sounds and turns pass, but the person’s understanding, expectations, confusion, and frustration accumulate. Kim argues that voice systems should be designed around that developing mental model rather than assessed only as a series of isolated outputs.

Dynamic voice-agent orchestration

Kim translates the framework into several engineering concerns: selecting and configuring automatic speech-recognition and text-to-speech systems, applying appropriate pre- and post-processing, curating vocabulary shared by the user and agent, managing turn detection and latency, handling emotion, and retaining context across the interaction. These elements must be coordinated dynamically because conditions and user expectations can change during a call.

She connects linguistic failures with operational consequences including unsuccessful tasks, live-agent escalation, abandoned calls, and failures that users never explicitly report. In the talk, she also points to EVA Bench as an end-to-end ServiceNow benchmark for diagnosing voice agents; the recording does not establish an individual project attribution for Kim.

Kim’s longer-term questions concern adaptation and language change. People become familiar with a speaker’s accent, vocabulary, and conversational habits through exposure, and users similarly adapt to a voice agent. She asks whether systems are prepared to respond to that adaptation across encounters and whether they can remain useful as language changes over months or years.

Her ServiceNow Insights interview extends these themes to enterprise deployment, including names and other out-of-vocabulary expressions, accents, turn-taking, changing environments, and the differing needs of employees and customers. Across both appearances, Kim presents linguistics as an engineering lens for understanding why voice interactions fail and how their components can be coordinated toward successful task completion.