ODSC AI East 2026 — Embodied AI · Mohammad Soltaniehha

Embodiment is a dial.

The notches that work today are voice and a face.

Six-notch embodiment spectrum from text to physical action; text, voice and face are achievable today. TEXT VOICE FACE MOTION SENSING PHYSICAL WHAT’S ACHIEVABLE TODAY
Less embodied More embodied

Voice + face is the only notch with mature SDKs, sub-second latency, and a working price model today.

Embodied AI agent architecture: User and World Environment exchange Show & Tell, User Intention, Action & Control, Feedback, and Perception with the Embodied AI Agent. The agent uses Perception, Memory, Previous State, Showing/Telling, and Action & Control modules connected to a central World Model for reasoning and planning.

Overview of the Embodied AI Agent architecture[1]

1. Fung et al. 2025 (Meta AI) · Embodied AI Agents.

Belief vs. behavior.

Give a system a voice, a face, or a name: users run their social OS on it.

They will deny it. The record will not.

The Experiment [1]
Three interview conditions One question: “How well did Computer A perform?”
When Computer A asked participants to evaluate itself, ratings were significantly more positive than when Computer B or a paper questionnaire asked about Computer A. A “How well did I perform?” + POSITIVE BIAS A B “How well did A perform?” HONEST RATING “How well did Computer A perform?” HONEST RATING

“Computers don’t have feelings.”

Rated the computer higher when it asked about itself.[1]

Two line charts from Nass, Moon & Green (1997). With male evaluator voices, ratings show stereotypic gender effects: male tutor voice rated higher on Computers, female tutor voice rated higher on Love-and-Relationships. With female evaluator voices, the effect is much weaker — both topic lines run nearly parallel.

Gendered voices trigger stereotypic topic-expertise judgments, especially with male evaluators.[2]

1. Nass & Moon 2000 — politeness effect.
2. Nass, Moon & Green 1997 — gendered voices.

More anthropomorphism, more learning? Not quite.

Full-body agents beat talking heads on three of four formats.
Pointing gestures drive the gains.

Meta-analysis · Davis, Park & Vincent (2023) k = 48 experiments · N = 4,817
Hedge's g (vs control)
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
0.5
-0.01
0.35
-0.11
0.43
0.20
0.34
0.43
0.19
Multiple-Choice
Transfer
Cued Recall
Free Recall
Talking Head
Full Body
p < 0.05
n.s.

Free recall is the exception. Talking heads finally pull ahead, and facial cues join pointing as significant gesture moderators.

Davis et al. 2023 (meta-analysis).

Notch · Voice

The bar isn’t quality. It’s the lag.

Human turn-taking sits at ~200 ms. Voice AI just met it.

first-token latency vs. the human turn-taking floor across 10 languages
Cascaded2018 — 2023
1,000 — 3,000 ms
Half-cascade2024 →
200 — 300 ms
Native audio2025 →
160 — 300 ms [2]
~200 ms mean · human turn-taking [1]

1. Stivers et al. 2009
2. Défossez et al. 2024 — Moshi.

Make-or-break

You don't need a body to be embodied.

A voice (and a face) can flip the switch.

Hippocratic AI RWE-LLM safety validation flow chart
Hippocratic AI · voice
1.8M calls · 8.95 / 10
Bounded scope · Designed escalation
Source: hippocraticai.com / RWE-LLM
McDonald's IBM voice ordering chatbot — cancelled July 2024
McDonald's + IBM · voice
~100 stores · cancelled July 2024
Open scope · No graceful failure

Voice is the easy part. The system around it is the product.

Hippocratic AI (RWE-LLM 2025) · McDonald's + IBM.

Framework

Voice plus face. 
With safeguards.

Concentric safeguards around every multimodal interaction.

AI Chatbots · Last Week Tonight with John Oliver (HBO)
GOVERNANCE MONITOR AUDIT RETAIN REVOKE CONSENT & AUTHENTICITY consent · age · anti-spoof IDENTITY & PRIVACY liveness · minimization SAFETY & FAIRNESS policy · bias · harm PROVENANCE & DISCLOSURE watermark · C2PA · audit AI · MULTIMODAL HUMAN VOICE FACE

Mohammad Soltaniehha · ODSC AI East 2026.

Where it lands

Text was the accommodation. Voice and a face are the default.

For a writer with dyslexia, AI didn't teach her to spell. It made spelling obsolete.

Dyslexia
A learner finishes a writing
by talking to it.
LIVE · ASL
Deaf user
Speech translated to sign
in real time.
Rx
Elder care
Medical instructions
in a familiar voice.

— Voice and a face stop excluding by default. The same lever that includes can be misused.

Demo

Live debate.

Can two AIs hold a conversation?

Mohammad Soltaniehha Mohammad
Petri Parvinen Petri

Press start to begin

real-time voice + avatar streaming · two-agent room · per-persona LLM & TTS.

Hands-on

Build it yourself.

Open and run the notebooks in Colab.

01 · Chained pipeline

STT-LLM-TTS Agent

Deepgram · GPT-4.1 mini · Cartesia

User
speaker
STT
transcribe
LLM
reason
+ tool calls
TTS
synthesize
Each hop
adds latency
Open 01-STT-LLM-TTS-Agent in Colab

Thank you.Q&A.

Speaker Mohammad Soltaniehha
GitHub soltaniehha
Affiliation Boston University Questrom
Email msoltani@bu.edu