Skip to content

EVA - A Framework for Evaluating Voice Agents by ServiceNow

April 29, 2026 · 30 min

Show notes

Voice AI agent evaluation — why it's fundamentally harder than text, how cascade failures derail conversations invisibly, and ServiceNow's open-source framework to establish industry evaluation standards. Featuring real audio examples showing authentication failures, leaked reasoning, and latency problems.

WHAT WE COVER 

TARA BOGAVELLI — Research Engineer, ServiceNow
Leading the open-source voice agent evaluation framework. Explains why existing benchmarks don't measure what matters and what ServiceNow is releasing to establish industry standards.

KATRINA STANKIEWICZ — Staff Machine Learning Engineer, ServiceNow
Cascade model architecture expert. Breaks down STT → LLM → TTS failure modes, named entity transcription challenges, and real audio example analysis.

GABRIELLE GAUTHIER MELANÇON — Staff Applied Research Scientist, ServiceNow
Multi-language evaluation specialist. Reveals why Large Audio Language Models lag behind, the native speaker requirement, and bot-to-bot simulation methodology. 

CHAPTERS
0:00 Introduction — The evaluation gap
1:11 ServiceNow's Open-Source Framework Announcement — Tara Bogavelli
2:43 Meet the Researchers
3:43 Voice-Specific Challenges — Tara Bogavelli
5:03 Cascade Architecture: STT → LLM → TTS — Katrina Stankiewicz
7:57 The Named Entity Problem — Katrina Stankiewicz
10:06 Evaluation Metrics: Accuracy vs Experience — Gabrielle Gauthier Melançon
11:23 Bot-to-Bot Testing at Scale — Gabrielle Gauthier Melançon 
14:30 The LALM Gap: Why Audio AI Judges Struggle — Tara Bogavelli
16:57 Real Audio Example: Flight Rebooking Gone Wrong
21:58 Breaking Down the Failures — Katrina Stankiewicz 28:30 Wrap-Up & Resources

KEY INSIGHTS

The Cascade Failure Problem: STT → LLM → TTS errors propagate invisibly Named Entity Transcription: The #1 enterprise blocker—names, confirmation codes, emails break authentication Accuracy vs Experience: Perfect task completion means nothing if users hang up due to poor experience LALM Gap: Large Audio Language Models lag behind text LLMs—human evaluators remain essential Latency Kills Conversations: Five-second pauses make users think the call dropped, breaking the experience even when tasks complete Open-Source Framework: ServiceNow releasing evaluation tools, metrics, and bot-to-bot simulation methodology for the industry.

LEARN MORE

Website: https://servicenow.github.io/eva/ GitHub:
https://github.com/servicenow/eva Blog Post:
https://huggingface.co/blog/ServiceNow-AI/eva Dataset: https://huggingface.co/datasets/ServiceNow-AI/eva

ABOUT

Hosted by Bobby Brill. ServiceNow Insights podcast explores AI research, real-world applications, and the people building the future of work. #VoiceAI #AIEvaluation #ServiceNow #MachineLearning #OpenSource #ConversationalAI #STT #TTS #LLM #VoiceAgents #AIResearch #Podcast

See omnystudio.com/listener for privacy information.

All episodes of ServiceNow Insights

Settings

Display

Language

Decides which stations you see first on the home page.

History

Recently played stations and episodes and your latest searches are only stored in this browser.

Notifications

A short notice when a podcast you follow has a new episode. Turn off single podcasts with the bell on their page.

Account

Save favorites – in the app and on the website.

Keyboard shortcuts

/
Open search
Space
Play or pause
← →
In the player: back 15 s, forward 30 s (episodes) · in a row: next tile
↑ ↓
In a row: target within the tile
Home / End
In a row: first or last tile
Esc
Close window or menu
?
This overview
Now playing

Go to station page

Play next

    Add episodes with “Play next” – they play after the current episode.