Skip to content
Share one workflow. DRING calls in two minutes and qualifies the need. Get a call in two minutes
Get a call in 2 minutes See Agent Factory
Research explained · Voice agents

Why is voice harder for AI?

In the τ-Voice study, the best text model solved 85% of 278 customer service tasks, while three voice models solved 31 to 51% of the same tasks on a clean phone line and 26 to 38% on a realistic one.

Same tasks, far fewer solved

The agent answered every question in the chat test. Then the same agent went on the phone line, and callers started asking for a person.

A study published in March 2026 put numbers on that gap. τ-Voice (tau-Voice) ran the same 278 customer service tasks in text and in voice. In text, the best text model in the study solved 85% of them. Three voice models, given the same tasks over a clean phone line, solved 31 to 51%. On a realistic line, with accents, background noise and a customer who interrupts, they solved 26 to 38%. The authors describe that as keeping 30 to 45% of what text can do.

For anyone comparing voice AI vs text AI, the tasks stayed fixed. The models did not, because the text score comes from a text model and the voice scores from three different audio models.

What did the study test?

τ-Voice is a voice AI benchmark written by Soham Ray, Keshav Dhandhania and Victor Barres of Sierra.ai in San Francisco, and Karthik Narasimhan of Princeton University. Sierra builds AI agents for customer service, so the study comes from inside the industry it measures. The paper is a preprint and had not been peer reviewed as of this post.

The 278 tasks come from three areas: retail, airline and telecom. Each task is a customer with a goal, such as exchanging an item or changing a booking, and a set of company rules the agent has to follow.

The study counts a task done only when the company's records end up in the right state, with the right order changed or the right booking updated. A pleasant conversation that changes nothing, or changes the wrong thing, counts as a failure. Every task success figure on this page is that measure (pass@1 in the paper, solved on one try).

In text, a text model typed with a simulated customer. In voice, three voice models listened and spoke at the same time over a simulated phone line with phone-quality audio, and the simulated customer spoke back. Each voice model takes audio in and gives audio out in one model. The voice models got the same instructions, including a rule to ask customers to spell names, emails and IDs letter by letter.

Each voice model ran in two conditions. On the clean line, the customer had a clear American accent, with no background noise and no interruptions. On the realistic line, customers spoke with different accents over background noise indoors and outdoors, sudden sounds, and dropped or muffled audio, and they interrupted, said "mm-hmm", coughed or said "hold on" to someone else.

How much did voice lose?

Task success, text and voice
AreaTextVoice, cleanVoice, realistic
All 278 tasks85%31 to 51%26 to 38%
Retail81%45 to 71%30 to 45%
Airline83%28 to 48%30 to 40%
Telecom90%20 to 58%18 to 40%

"Text" is the best text model in the study, GPT-5 with reasoning. A second text model in the study, GPT-4.1, which answers without step-by-step reasoning, solved 54% of all 278 tasks. The voice columns give the lowest and highest of the three voice models tested, Google gemini-live-2.5-flash-native-audio, OpenAI gpt-realtime-1.5 and xAI grok-voice-agent. Source: τ-Voice, Table 6.

Share of the 278 tasks solved in text and in voice. Each voice row runs from the lowest to the highest of three voice models. Source: τ-Voice, 2026.

Most of the gap had already opened on the clean line, before any noise or accent was added. There the three voice models scored 34 to 54 points below the best text model, and for most of them the realistic line took less away on top of that.

What made it worse?

On the retail tasks, the authors added noise, accents and interruptions to the clean line one at a time, then all three together.

What each difficulty cost, retail tasks
ConditionTask successChange from clean
Clean55%
Clean plus noise51%4 points lower
Clean plus accents44%10 points lower
Clean plus interruptions47%7 points lower
Realistic (all three)38%17 points lower

Average of the three voice models, retail tasks only. The percentages and the differences are printed as the paper rounds them, so a difference can sit 1 point off the subtraction. Source: τ-Voice, Table 7.

Task success on retail tasks, average of three voice models, on the clean line and with noise, accents, interruptions or all three added. Source: τ-Voice, 2026.

Accents cost the most on average, but very unevenly. One voice model lost 1 point with accents and another lost 18. Interruptions cost two of the models 14 and 11 points, while the third did slightly better with them, 4 points higher. Noise on its own cost each voice model 2 to 4 points.

The authors also scored how the voice models handled the conversation on the realistic line. The one that responded to every customer turn and yielded to every interruption (100%) correctly ignored "mm-hmm", coughs and side talk only 6% of the time, so it also answered sounds that were not meant for it. In the authors' reading, no voice model in the study was good at both finishing the task and handling the back and forth of the conversation.

Where did the calls break?

The authors read 91 failed voice conversations by hand, in two groups. Of 43 failed conversations from tasks that were solved in text but failed in clean voice, the agent caused 34 (79%) and the simulated customer 9 (21%). Of 48 failed conversations from tasks that were solved in clean voice but failed in realistic voice, the agent caused 43 (90%) and the simulated customer 5 (10%).

What went wrong in 91 failed calls
Agent's errorFailed in clean voiceFailed only in realistic voice
Wrong reasoning or a skipped step1316
Misheard what the customer said1016
Said it did something it did not do66
Stopped responding14
Ran out of time41
Agent's errors, total34 of 4343 of 48

The paper's names for the rows are logical, transcription, hallucination, VAD/unresponsive and timeout. Source: τ-Voice, Table 9.

In both groups, the identity check was the main place calls broke. Agents got names and emails wrong even when the customer spelled them letter by letter, as the instructions told the agent to ask. With the customer not identified, nothing else in the task could happen.

Some failures were quieter. In one conversation the agent told the customer "I've updated your shipping address" and made no change in the system. Requests with several parts caused trouble too, with agents completing one part and forgetting the rest.

Hearing was not the whole story. In the realistic group, wrong reasoning and mishearing were equally common, 16 of 48 failures each, and the authors conclude that the gap between voice and text is not purely a speech recognition problem. How to test the hearing part on your own calls is in our post on choosing speech-to-text.

What does it not show?

The study ran in English only. The customers were simulated, with synthetic voices, and the authors say the accent results should be read as indicative. They also note that real callers are less patient than the simulated ones.

It measured whether the task got done and how the turns were taken. It did not measure how the agent sounded, whether customers were satisfied, or tasks that were partly completed.

The text and voice scores come from different models. The best text model can reason for as long as it needs before it answers, while a voice model has to answer in real time, so part of the gap is the model and not the phone line.

The three voice models are of one kind, audio in and audio out, in the versions released between December 2025 and February 2026. Agents built from separate speech recognition, reasoning and speech parts were not tested, so these numbers say nothing either way about them, and the authors list them as future work. The agents also worked from a general set of instructions and the benchmark's tools, and an agent built and tested for one company's calls is a different setup that the study does not measure in either direction.

DRING did not run this study and was not tested in it. Models change fast, and the numbers describe the versions that were tested.

What should you test first?

Run your test as a phone call. In the study, the same tasks scored lower over phone-quality audio than in text, and lower again once accents, noise and interruptions were added. A text chat or a laptop microphone skips both, and your callers are in the column "Voice, realistic".

Start with the identity check, since that was the main place the failed calls in the study broke. Take ten names, emails and order numbers of the kind your callers have, read them out the way callers spell them, and count how many the agent gets right.

After every test call, open the system and look at the record. If the agent said it changed an address, the new address should be there, which is what went wrong in the shipping address case in the study.

Next, make the calls harder in the ways that cost the voice models task success in the study. Use the accents your callers actually have. Have a tester interrupt, cough and say "hold on" to a colleague halfway through, and note whether the agent answers the cough or the side talk. Then give the agent one request with three parts and check that all three are done.

When a vendor quotes a task success number for its voice agent, ask for it counted the way the study counted it, by what changed in the system, on phone audio. For a full plan, our guide to testing a voice AI agent before production covers the test set and the release gates.

Can you quote this page?

Yes. The paper is published under Creative Commons Attribution 4.0 (CC BY 4.0), and you can reuse this summary on the same terms, with credit to the paper first. Cite the study like this.

Ray, S., Dhandhania, K., Barres, V., Narasimhan, K. (2026). τ-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains. arXiv:2603.13686.

The benchmark code is public on GitHub. If you quote a number from this page, please credit the paper and add this line.

Summary: DRING, dring.ai/blog-voice-ai-vs-text-ai

The five findings, written so you can quote them as they stand:

  • On 278 customer service tasks, the best text model solved 85%, and the three voice models solved 31 to 51% on a clean line (τ-Voice, 2026).
  • On a realistic line, with accents, noise and interruptions, the three voice models solved 26 to 38%, which the authors describe as keeping 30 to 45% of what text can do (τ-Voice, 2026).
  • On the study's retail tasks, accents cost the three voice models 10 points of task success on average, interruptions 7 and background noise 4 (τ-Voice, 2026).
  • In 91 failed voice calls the authors read by hand, 79% and 90% of the failures in the two groups were the agent's own errors, and the identity check was the main place calls broke, with names and emails misheard even when spelled out (τ-Voice, 2026).
  • No voice model in the study was good at both finishing the task and handling the conversation, and the one that responded to 100% of customer turns correctly ignored coughs and side talk only 6% of the time (τ-Voice, 2026).

Where does DRING come in?

The study measured general voice models working from a rulebook. A DRING agent is built from separate parts, with selected speech-to-text, text-to-speech and reasoning models, testing and monitoring in the shared foundation it runs on, and that kind of agent was not in the study.

Every DRING agent runs 1,000 to 10,000 simulated conversations, built for your company, before the first real call. They are generated from your own workflow, documents and call recordings, and an agent goes to production only after it clears that set, as our quality page describes. The tests in "What should you test first?" are the ones to ask any provider for, DRING included.

Also read: Voice AI or IVR: Which Should You Choose? · What Is an AI Voice Agent? How DRING Runs One

Start with your hardest call

Leave your number and DRING calls you in two minutes. Tell it the call your team dreads most.