What Sierra's test measured and what it found
On October 21, Sierra, a US customer service agent company, published τ-i18n, an extension of its τ-bench test suite. It evaluates voice agents in five languages besides English: Spanish, Brazilian Portuguese, Hindi, Korean and Mandarin. It made 2,700 calls with three voice agents (GPT Live 1, Gemini 3.8 Live and Grok Voice Think Fast 2.0) on the same airline, retail and telecom tasks in each language.
The main findings were:
- Spanish holds up well on tasks: the task completion rate in Spanish was only 2.4 points below English; in Portuguese, the gap was 6.4 points.
- Conversation is the weak point: in more than 80% of calls, GPT Live 1 and Gemini 3.8 Live failed to respond to a customer turn, answered slowly or did not yield the floor. Grok Voice Think Fast 2.0 had this problem in 16% of calls.
- They react to what they shouldn't: depending on the agent, between 26% and 60% of calls had more than three responses to filler words like “uh-huh” or to voices not addressed to the agent.
- No one wins at everything: GPT Live 1 led in task completion (79.9%) and language quality; Grok led in interaction, but with low language quality. In Sierra's words, the agent that responds fastest is not necessarily the one that speaks most naturally.
Sierra released the code so others can add languages and regional variants.
This week's launches: Mistral and ElevenLabs
On October 20, French company Mistral AI introduced Voxtral 3, a family of models to listen, transcribe and speak, with a kit for building voice agents based on LiveKit. According to Mistral, real-time transcription has latency below 500 milliseconds, each agent turn completes in under one second, Spanish is among the supported languages and phone calls connect via SIP. The voice agent service costs from 0.04 US dollars per minute and can be used from servers in Europe or installed on the customer's infrastructure.
That same day, ElevenLabs launched Eleven Realtime, its first speech-to-speech models for agents, in two versions: Mini, for low-cost conversations, and Max, for complex tasks such as support or lead qualification. The company reports an average time to first audio of 0.57 seconds for Max and 0.62 seconds for Mini, support for more than 90 languages and compatibility with OpenAI's realtime interface. All these figures come from the vendor.
Why latency decides whether a call works
In a conversation between people, the pauses between turns are very short. When a voice agent is slow to respond, the customer repeats the question, talks over the agent or hangs up. That's why the industry competes in milliseconds: each step (listening, thinking and speaking) adds delay, and on top of that comes the audio's trip across the phone network and the internet.
Sierra's test shows that speed is not enough. A good agent must also know when the customer has finished speaking, ignore noise or comments directed at someone else and allow itself to be interrupted. It's the same reason operators like Orange and Elisa are starting to run voice AI inside their networks, closer to the user.
What it means for companies in Colombia and Latin America
The news is good for anyone who wants to automate calls in Spanish: there are more options, lower per-minute prices and task performance close to English. But before hiring a voice agent, it's worth asking the provider these questions:
- How long does it take to respond on a real call? Ask them to measure the delay from when the customer finishes speaking until they hear the response, over the phone network, not in a web demo.
- How does it handle interruptions and filler words? Try interrupting it, saying “uh-huh” or talking to someone next to you.
- Has it been tested with Colombian Spanish? The tests use general Spanish; accents, proper names, ID numbers and local addresses must be validated with real calls.
- Where are the voice and transcripts processed and stored? This is personal data protected by data protection law.
- How does the call get transferred to a person? There must always be a way out to a human agent, with the context of what has already been discussed.
At 19 80 TIC Cloud we connect AI voice agents to your numbers and your cloud PBX or call center, so you can test them with your own customers and escalate to a human agent when needed.
