The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize

Author: AI Engineer (Arize) · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: The Transcript Looked Fine. The Call Wasn’t. — Debugging Voice Agents, Arize · Channel: AI Engineer (Arize) · 17:17 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI

Summary

The talk explains why voice agents are difficult to debug: transcripts can look correct while audio reveals latency, interruptions, misheard words, and failed actions. It recommends observing the complete interaction—including audio, traces, session turns, tool calls, and audio-specific metrics—instead of relying on text logs alone. The speaker describes open, provider-agnostic tracing conventions and audio-focused evaluations for issues such as sentiment, latency, interruptions, and task success. The broader improvement loop is to observe, evaluate, fix, and validate agent behavior, with the possibility of automated investigation and fixes.

Key points

Voice agents have audio-specific failure modes, including latency, turn-taking problems, transcription drift, and interruptions.
A transcript may hide important failures; the refund example involved dead air, the agent speaking over the caller, and the wrong order number being processed.
Debugging should bring audio, transcripts, traces, session turns, tool calls, and metrics together in one view.
Useful voice metrics include time to first audio, latency, audio token cost, and interruption events.
Open Inference semantic conventions map audio events to an OpenTelemetry-compliant schema and support querying across model providers.
Replaying a full session helps identify who spoke when, where interruptions occurred, and which tools the agent called.
Audio-specific evaluations can assess sentiment and tone, latency, interruptions, accuracy, transcription drift, and task success.
Putting evaluation results directly on relevant spans supports filtering, investigation, and follow-up workflows.
The recommended improvement loop is observe, evaluate, improve, and verify; the speaker also describes agent experiments that replay failed traces against fixes.

Chapters

Start	Chapter	Summary
00:01	Why voice agents are hard to debug	Fuad introduces the talk and explains how the rapid growth and fragility of voice agents make their failures difficult to see in text logs.
02:37	When the transcript hides the failure	A refund call demonstrates how latency, interruption, and a misheard order number can be obscured by an apparently reasonable transcript.
04:13	Tracing and replaying voice sessions	The speaker outlines a unified view of audio, transcripts, traces, metrics, and tool calls, backed by open, provider-agnostic conventions.
09:58	Evaluating audio-specific behavior	The talk covers audio-based evaluation of tone, latency, interruptions, transcription accuracy, and task success, with scores attached to spans.
13:38	A continuous improvement loop	The speaker describes observing, evaluating, fixing, and validating voice-agent behavior, including experiments that replay failed traces against proposed fixes.
15:44	Getting started and closing remarks	Fuad encourages listeners to instrument and evaluate their voice agents, points to resources, and shares a free Voice Pro offer.

Quotes

the log lies. — 01:36

you need to see the conversation beyond just the transcript. — 03:41

you’re not just reading transcripts, you’re listening to audio — 07:52

The loop really follows observe, evaluate, and improve. — 13:38

Concepts and entities

voice agents, Arize AI, Open Inference, OpenTelemetry, Gemini Live, Grafana, Claude, SRE agent, agent experiments

Full transcript (original language)

00:01 [music] >> Folks, thanks for coming in out. I know it’s lunch time, so I know a lot of you folks are trying to eat as well. Appreciate you making the time and for everyone listening online, thanks for tuning in as well. Uh my name’s Fuad. I’m a product manager here at Arize. Uh I’ve been leading up kind of the experimentation side of the house and one thing that I’ve been really interested in is voice agents. Uh

00:33 quick show of hands here, maybe online too. Who here is building a voice agent? Anyone? Got Okay, we got a couple. Uh I think it’s a really cool space. I’ll be talking a little bit about voice traces, session views of voices, voice evals, and some of the stuff that’s new in Arize. But I do want to kind of mention this isn’t a product pitch. This is, you know, really on why voice agents are invisible and how to see them. You can use Arize to do that, you can use another platform, but what I want you

01:05 guys to take away from this is some of the failure modes and how you can actually start to see that. So, we all know voice agents are growing really, really fast. Have a few tweets up uh kind of proving this. GPT real-time just dropped. That was literally May 7th. Uh I have a screenshot from a tweet an hour ago from the LiveChat CEO talking about some of this. Uh and Vercel on June 29th, which is only a couple days ago, um on voice agents in Vercel. So, as you can tell,

01:36 uh this space is exploding fast and that is going to start to matter because as the space grows really fast, it is also one of the hardest to debug categories. So, first off, it’s exploding. Second off, it’s pretty fragile. There’s a lot of failure modes you can’t really see. Um and uh things like latency, turn-taking, and transcription drift become really, really important. Uh, and the log lies. If you’re just looking at transcripts, I’ll play a little audio file for you guys in a second. Uh, it doesn’t really

02:07 capture what’s actually going on under the hood. And just to give you guys some ideas of use cases, um, there’s a lot of them from customer support, I think, which are the more boring ones, down through drive-thru ordering. If you’ve been to White Castle, they have a a bot called Linda now. Uh, Bojangles has uh, Bo Linda as well. There’s McDonald’s is doing a lot of these as well, which I think is pretty cool. And personally, I’d be pretty pissed off if I was at McDonald’s at 2:00 a.m. and I got 600 cheeseburgers. So, I think I’ll show you

02:37 a little bit about the failure modes and why it actually matters. So, uh, this is an example where the transcript looks fine, but the call wasn’t. And if you read the text logs, you know, user says, “I need a refund for order 14.” Agent says, “Sure. Started a refund for order 40.” User says, “No, 14.” And it looks like the agent corrected itself. But, this wasn’t actually the case, and I’m going to I’m going to play the call. It’s AI-generated. Let’s see if it works. >> I need a refund for order 14.

03:10 >> Sure. I’ve started a refund for order 40. >> No, 14. >> Your refund is confirmed. >> So, if you notice in the transcript, if you could hear the audio there, uh, it wasn’t, you know, the the the confirm- confirmation on the on the 40 wasn’t actually done by the agent. It kind of interrupted. Uh, and there was a big, uh, kind of latency spike, uh, after the user asked for the initial refund. Um, so, we what actually happened was that two 2.4 seconds of dead air before the agent

03:41 actually responded. The agent talking over the caller. 14 misheard, and obviously, the flat robotic tone is is not a a look, either. Um, so, these are all audio specific issues. If you just read this from an LM output, you would have no idea anything is going on. So, what you need to do is you need to see the conversation beyond just the transcript. Um, so that means the audio, the transcript, and the trace all in one place. The session as a whole, so multiple turns from the user back and forth in

04:13 the same in the same trace view. And then metrics on those, so what was the latency like? What was the time to first audio like? Were there any interruptions? Sentiment analysis, et cetera. And these are some of the features we’ve just released at the beginning of this month in Arize AIX. Uh, and really the point of this is the same workflows that you’re running now to help your coding agents, uh, which are, you know, pretty complex processes now with harnesses, evals, uh, handoffs, and subagents, tool calls, you want to start to bring some of those best

04:44 practices into your voice agents as well. So, first off, you got to trace it. You got to see each individual span. Um, you got to see the live audio and play it inline from your trace. Really matters to have all this data in the same place so you can start to debug things. Uh, you want to see inputs and outputs. You want to see the transcript, obviously, but you also want to see, you know, specific metrics like I mentioned, like time to first token, uh, audio token cost, interruption events. And then finally, you want to see linkage between these. You don’t want to just see these as a list. You

05:15 actually want to see some of the linkage between things, right? So, Arize gives you that one trace view where you can actually play the audio inline, view the transcript, view latency inline, see the conversation turn between the user, uh, and the assistant, uh, and see your evals inline as well. And this gives you a really, really powerful tool to debug. Under the hood, what this looks like is we’ve actually mapped, uh, the semantic conventions for audio onto open

05:46 inference. So, if you guys aren’t familiar, Arize is first to market with open inference in 2023. It is the first set of semantic conventions that is completely open source that maps kind of what is happening with GenAI to an OTel-compliant set of conventions. And I say this to say it’s open source, it’s free, and it maps pretty much any of the major foundational models or model providers. You guys can go rip this and send it to a Grafana back end. Completely free to use. Not pitching the product here, just explaining some of

06:18 the real engineering work it takes to map things like web socket events to an easily understandable set of semantic conventions. So, you can start see things like the session life cycle, the audio input, the transcript, conversation items, the output, and then the response life cycle and token counts as well. And so, yeah, real engineering work went behind this. It’s all open source. We’ve mapped all the foundational models for you. What this allows you to do is you can now be provider agnostic and have one

06:49 queryable schema. So, those semantic conventions normalize sort of these disparate model providers, whether it’s OpenAI, real-time, or Google with Gemini Live into a set of schema that you can query. And if you’ve ever tried to debug at 2:00 a.m. off of a customer call, you know how important it is to be able to query across thousands, millions, even billions of traces that are coming in. You know how important it is to actually get to the part that you need to see, to filter that, to sort by

07:20 that. All that becomes incredibly important. And, you know, as new models drop, you also want to be able to swap in new providers and make sure that your systems are still functioning. So, no bespoke parsing per vendor. It’s one auto instrumenter. It captures every single thing that I just mentioned. And then finally, you want to be able to replay the whole conversation. So I play I played the audio for you guys in the beginning of this, but the session ID is essentially what stitches every turn into the timeline, the full call with

07:52 every websocket event that’s submitted with some tags on whether it’s the user or the assistant. You can actually play the audio in the UI directly from the trace. No external tooling. No exports. Just play it straight from the trace. That allows you to see who spoke when. When the agent paused, where it got cut off, where the user got cut off, what tools the agent actually went and called when it did this. And so this really starts the paradigm shift, which is you’re not just reading transcripts, you’re listening to audio, you’re really

08:23 understanding the failure cases from there with the trace open right next to it. And um I’ll just show you kind of quick quick sneak preview of what this looks like in Arise. As you can see here, this is an example of like you know an agent that can call to get a specific weather event. You can actually see when the agent kicks off that tool call in line with the rest of the audio. And this allows you to debug failures of that tool call while you’re looking at the session audio. So you’re no longer looking at disparate events from distant

08:53 systems, but you’re actually looking at the entire trace tree for everything that happened in that session. So that real session, few issues that we saw, right? Turn one, “Hi, need help with the order, a refund for order 14.” We saw you know kind of that time to first audio of 2.4 seconds of dead air, which really impacts the user experience. We saw in turn three that the agent actually spoke over the caller’s interaction, which is not a good look for any assistant, especially one that’s trying to be helpful. And

09:25 then ultimately if you look at the tool call, you’ll be able to also understand that the refund was processed for order 40 and not 14. Um which is a huge huge mistake and actually seeing that live in the tool call is really really important. And every single one of these things would be invisible in a text log. But with the session view you can actually scrub the audio and find those failure surfaces within the tools, within what the agent is doing as well. The next thing I want to talk about real quick is evals that are specific to

09:58 audio. So I kind of mentioned a few of them. I think tonality and sentiment analysis is really important. You want to be able to classify user emotion based on the actual audio. Transcript can’t do this for you. It’s very very easy to look at a transcript and see that it it, you know, looks fine, but tonality really matters. Positive, neutral, negative, not just a guess. Latency is really really important in audio and it’s important in a few use cases. Models are getting better. There’s, you know, jet bar architectures, etc. But time to first

10:30 audio SLAs are still really important for customer service use cases. And you want to be able to flag things that are breaking that. You also want to flag things like I mentioned like interruptions, inaccuracy, transcription drift, and the task success. Um And you want to be able to run these evals directly against the audio with evals that are based on audio models. So, you know, LLM as a judge evals on a transcript are only so good. But actually evaling the underlying audio is really important. And so, um we have kind of the tooling available

11:02 in Arise to be able to do that and we’re supporting all the major audio platforms as well as they come out with new models. So again, evals specific to audio really matter. They’re not just transcription based evals. And as you start to scale up your LLM as a judge evals or even move to agent as a judge that are doing like complex tasks like, for example, doing our back checks to making sure the agent actually access the system it was supposed to have access to. Being able to run these agent as a judge workflows in Arise, starts to become

11:33 really important and we’re kind of seeing the industry move away from, you know, just an LM as a judge analyzing a transcript or even analyzing audio, but actually accessing tools, accessing third-party systems, and hydrating some of that evaluation information inside of it. So, really easy to get started with evals in Arise through skills. It is literally as simple as, “Hey Claude, create an audio eval for me for sentiment analysis using GPT audio and catch frustrated user tone on those

12:04 audio spans.” And Arise will kind of take care of all the rest for you. The CLI under the hood is able to query spans, do some trace analysis, filter for the spans that are audio specific, actually create that eval, deploy it, and put it against traces and spans in your actual system. And this is part of why Arise exists. Yes, of course, you can do this yourself, but when you’re trying to query, you know, billions of spans with sub-second query latency, it becomes really hard and also deploying these live evals as they come in. So,

12:36 you’re scoring them while users are actually interacting with your platform is really important. And then the final thing I’ll mention is you also want these evals on the spans themselves because then you start to be able to do really complex querying and filtering across your data. So, you’re not just looking at a failure and looking up in an external system. You actually have the eval scores in line with the exact audio that you’re looking at. And that’s why some of this is really powerful in Arise. Being able to see those eval outputs directly on the spans

13:07 that are relevant helps you kind of get a full understanding of what is going wrong with the user interaction. And then you can kick off workflows from that. So, evals and monitors attached to those evals can kick off investigation agents that automatically start to dive into, you know, what tool calls and misrepresenting, Um, where is my agent failing? Is it capturing, you know, maybe a product surface area that I didn’t code it for? Is there a new edge case I need to code for in my agent? Maybe I need to expose

13:38 a new tool. All those become um, possible once you have the results directly on the spans. And so, I’ll I’ll just kind of flag and and end with this thought. The same loop that you’re doing to improve some of these tools like coding agents, we now need to do for voice. And the loop really follows observe, evaluate, and improve. And then from that loop, you can start to get some really really cool continuous improvement loops within the Arize platform or within another

14:09 platform if you choose to use it as well. Um, and the idea is once you observe and you have the data available to you, then you evaluate and you have the scores in line with your observations. Then you can start to kick off fixes, whether they’re fixes that you make or whether they’re fixes that an agent is automatically making on behalf of you. And then you can start to verify and validate those fixes. So, we just released agent experiments as well in Arize. Agent experiments, you could think of it kind of like Postman with tracing. So, imagine a world where you have audio

14:42 streaming in from your users, it’s automatically traced, you have live evaluations in the platform that are automatically scoring those interactions. And then you have SRE agents, so not an actual engineer that you hire, but an SRE agent that’s looking through the traces, understanding the failure modes, hypothesizing a fix, and then actually standing up a fix for you in dev, replaying those failed traces for you against that fixed endpoint, tracing that to validate that the latency actually decreases. And you wake

15:12 up in the morning, you look at a PR, the PR has a description of a problem. Hey, I found out why the latency was high. Turns out we were not truncating our tool results, and so we were passing in way too much information to the agent. Context got overloaded. I introduced the truncation algorithm. I shipped it. I actually stood it up in dev. I tested against the traces that failed. Here’s the report card. Here’s how latency decreased. And you wake up in the morning and all of this is ready for you, and you just click approve or deny on the PR. That is the future that I

15:44 think we want to see at Arize, where self-healing software and continuous improvement are realities for everyone who’s building agents. And I think making that possible for voice is really, really important. So, start seeing your agents. Add the Open Inference Instrumenter. Open source, free to use, OTEL compliant. You can send it to any back end. Um so, use Arize, don’t use Arize, start tracing your audio. The second part of this is actually starting to do that analysis, and we would recommend Arize to do that. Sign up for a free account. Uh I’ll have a

16:16 code up at the end of my presentation for you guys to sign up for a a free year of pro. Um but start checking time to first audio. Start checking your interruptions. And start evaluating um how these interactions are actually going and whether they’re accomplishing your use cases. Whether it’s a tone or latency eval, um try using the Arize skills to actually create that eval. Uh we have some docs set up for you if you want to follow along. Bunch of examples and cookbooks ready. Uh and bring voice into the same rigor as the rest of your

16:46 stack. You wouldn’t be okay with this happening at the coding agent level. You wouldn’t be okay with it happening uh for an agent that’s able to execute stocks on Robinhood. You shouldn’t be okay with it, uh though the stakes might seem a little lower, for a McDonald’s agent that’s ordering your cheeseburger, either. Um and so I’ll end with that. Please, please keep building. Please connect with me also on Twitter. I still call it Twitter, not X. Uh on Twitter or LinkedIn. Would love to kind of talk and hear about some of the use cases you guys are building towards so we can add

17:17 more echos to our library and make sure this is useful for you. And if you found this talk useful and you want to start investigating your voice agents, there is a free code up on the screen. It gives you a year of Voice Pro for free. It’s a i w f 2026 and you can get started and start tracing your voice agents. So thank you very much. My name is Ali. I’ll be available for questions for a couple minutes after but really appreciate you guys listening. Cheers.