A customer wants to exchange a puzzle, change a flight or fix a phone plan. Give the request to an AI through text and it has one job: understand the instructions and leave the account in the correct state. Put the same sort of work inside a live phone call and several new jobs appear. The AI has to hear names and order numbers, notice when the customer has finished speaking, avoid treating every “mm-hmm” as a new request, use the right tools and begin answering before the silence feels awkward.
A benchmark published in March 2026 tried to measure what happens when those extra jobs arrive. Across 278 retail, airline and telecom tasks, a reasoning text model completed 85 percent. Three audio-native voice agents completed between 31 and 51 percent under the cleanest speech condition. When the calls included varied accents, background noise, interruptions and other ordinary features of conversation, the range fell to 26–38 percent.
It is tempting to conclude that giving an AI a voice makes it stupid, although the experiment cannot establish that because it compared different model families working under different constraints. It shows a large gap between the best text result and the voice systems tested, followed by a smaller but still meaningful loss when the audio became more realistic.
What the agents were asked to do
The benchmark, called τ-Voice, extends an existing text-agent test into live, two-way speech. “Full duplex” is the technical term for a conversation in which both sides can speak and listen at the same time, so the agent has to handle interruptions and short acknowledgements while continuing to pursue the task.
The 278 tasks covered customer service rather than trivia. There were 114 retail tasks involving returns, exchanges and order changes, 50 airline tasks involving bookings, cancellations and seat changes, and 114 telecom tasks involving billing, authentication and service changes. A single conversation could contain several requests or a customer changing their mind.
Success was checked against the final state of the simulated business system. If the customer asked for an address change, a polite promise was insufficient. The address in the database had to change correctly. The benchmark reports pass@1, meaning the proportion of tasks completed on the first attempt.
The text baselines were GPT-5.2 with medium reasoning and GPT-4.1 without a reasoning mode. They completed 85 and 54 percent respectively. The voice systems were OpenAI’s gpt-realtime-1.5, Google’s gemini-live-2.5-flash-native-audio and xAI’s grok-voice-agent.
| System | Task completion |
|---|---|
| GPT-5.2 text, medium reasoning | 85% |
| GPT-4.1 text, non-reasoning | 54% |
| Voice agents, clean calls | 31–51% |
| Voice agents, realistic calls | 26–38% |
That second row matters. The best voice result under clean conditions was 51 percent, only three points behind the non-reasoning text baseline. The spectacular 85-versus-31 comparison mixes two differences: modality and reasoning capability. Anyone using the benchmark to claim that voice alone cuts performance by more than half would be claiming more than the experiment measured.
A call creates failure points that text does not have
The clean condition was generous. It used clear American-accented synthetic speech with no background noise or interruptions, although the audio still passed through telephone-quality encoding. The realistic condition added other accents, indoor and outdoor noise, dropped audio frames, muffling, coughs, interruptions, backchannels and remarks such as “hold on” that were not directed at the agent.
The researchers then added noise, accents and turn-taking separately on the retail tasks. Across providers, accents caused the largest average drop, followed by turn-taking and noise. The size of each effect differed sharply by provider, so this is not evidence that every accent or every model suffers equally.
Authentication exposed the practical problem because names, email addresses and order identifiers often have to be captured exactly before any later action is permitted. The system can understand the customer’s intention perfectly and still fail when one letter in an email address enters the tool call incorrectly. Asking the person to spell the name still leaves the individual letters travelling through the audio channel.
Timing creates a different class of error because a person saying “mm-hmm” may only be signalling that they are listening. One agent in the benchmark responded to almost every such sound, another frequently interrupted the customer and a third failed to respond to nearly a third of user turns. A text message arrives with a visible boundary, while speech arrives continuously and forces the agent to infer when an utterance is finished and whether it requires action.
Then the system has to keep the conversation and the business record aligned. In one reviewed failure, an agent said it had updated a shipping address without making the required tool call. Other agents completed one part of a multi-step request and forgot the rest. This is the same distinction we examined when people preferred Gemini’s voice agent even though other models completed more tasks: the sentence a person hears and the state left in the backend are two separate records.
Voice is operating under a clock
A text reasoning model can remain silent while it works because the user expects some delay after sending a difficult prompt and the interface can display a loading state. A phone conversation supplies less cover because long pauses feel like a broken connection, starting too quickly risks interrupting the caller and waiting too long makes the agent feel unresponsive.
The systems in τ-Voice therefore had to solve the task while also managing the performance of a live conversation. The fastest provider began responding in about 0.90 seconds in the realistic condition. Yet the providers traded weaknesses rather than producing one clear winner: the fastest was poor at ignoring sounds that required no answer, while another was more selective but interrupted far more often.
The engineering choice therefore includes more than model intelligence. A common production design first converts speech to text, passes the transcript to a text LLM and then converts its answer back to speech, which provides a transcript and separate stages to inspect. An audio-native model can listen and speak directly, preserving information such as tone and reducing handoffs, but its reasoning, turn-taking and spoken response must coexist in a real-time system.
The delay created by those stages is explored separately in where the delay comes from after you stop talking to a voice agent. The important point here is that speed and task completion are coupled. A model that spends longer thinking may answer better and still make a terrible live caller.
What this comparison still cannot tell us
τ-Voice is closer to a real transaction than a test that asks spoken trivia questions, but it leaves several questions open.
The voices came from text-to-speech personas rather than recorded human callers, and the study was English-only. It measured complete success rather than partial progress and did not score how natural or pleasant the generated voices sounded. The researchers also compared audio-native systems with text baselines without including a speech-to-text, text LLM and text-to-speech cascade built around the same central model. Their limitations section proposes that baseline as future work because it would help separate the cost of the voice modality from the cost of choosing a different architecture.
A clean follow-up therefore needs matched conditions in which one set of tasks goes directly to a text model, the same central model runs behind a spoken cascade and an audio-native system receives the same work. The comparison should preserve the transcript, tool calls, final database state, response timing and interruptions so that a wrong name in the spoken cascade can be traced to either recognition or the LLM’s handling of a correct transcript.
Until that comparison exists, “voice AI is dumber” is a useful suspicion and a bad conclusion. The current evidence says something narrower: the voice agents tested had to perform several jobs that the text baselines did not, and none of them performed all of those jobs reliably.
For a real deployment, run the same booking or account change repeatedly through text and voice, keep the recording and then open the database to check what actually happened.
Sources
- τ-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains, March 2026.
- Pipeline vs. realtime: which is the better voice-agent architecture?, LiveKit, April 2026.
- Benchmarking LLMs for Voice Agent Use Cases, Daily, 2026.