Santa Clara's ServiceNow releases a free AI model that listens to speech and acts on it
Unlike the voice modes in ChatGPT or Gemini, Apriel Voice is free to download and run on a company's own computers, and it is built to carry out customer-service tasks. ServiceNow's own tests show it trailing the best paid systems overall but leading on some measures.
ServiceNow, the Santa Clara company that sells software businesses use to run help desks and office tasks, released a free AI model Oct. 29 that understands spoken requests and acts on them, the company said. It is called Apriel Voice 27B Preview.
The model listens to audio, works out what the speaker wants and can use software tools, such as looking up a booking, to get it done. It answers in text, so builders pair it with a separate program that turns the text into speech.
How it differs from ChatGPT, Gemini and Alexa
Talking to ChatGPT, Google's Gemini or Amazon's Alexa means using a finished product that runs on that company's own computers. People and businesses can use those assistants, but they cannot download the AI inside them or change how it works.
Apriel Voice is the opposite. ServiceNow released it under the MIT license, which lets anyone download, use and change it for free. That means a business could run it on its own equipment instead of sending customers' voices to another company, and could retrain it for its own work. The "27B" means it has about 27 billion parameters, the internal settings an AI learns during training. ServiceNow's setup instructions run it on four Nvidia H100 chips, powerful processors used in data centers.
It is also built for a narrower job. Instead of chatting about anything, it is designed to finish tasks in a conversation, the company said. In a demonstration on ServiceNow's blog, it helps a caller change a flight: it looks up the booking with a spoken confirmation code, searches for flights that land before 3 p.m., checks for an aisle seat and confirms the $85 fare difference before making the change.
Because it does not have its own voice, a company can swap in whatever speaking voice or speech service it likes without retraining the model, ServiceNow said. The model can also take in pictures, such as a screenshot of an error message, in the same conversation.
Is it better?
Not across the board, according to ServiceNow's own tests, which compared complete voice systems. On a test of finishing customer-service tasks by voice in airline, retail and phone-company scenarios, OpenAI's paid GPT Realtime 2.1 completed 69.8% of tasks and Google's Gemini 3.8 Live completed 84.2%, compared with 65.5% for Apriel Voice. Both paid systems also answered a little faster, in about 1.5 seconds compared with 1.9 seconds.
Apriel Voice did lead in some areas. It had the highest score of the systems tested on a measure of how natural the conversation felt, including staying brief and handling turns in the conversation, and it completed the most retail tasks, 81.6%. It also beat the other free, downloadable voice systems ServiceNow tested on most measures. Its weakest area was phone-company tasks, where it finished 45.6%.
ServiceNow did not compare it with Alexa or with Anthropic's Claude. The company calls the model a preview and says it works less well in languages other than English, with background noise or distant microphones, and on complicated multistep tasks.
It is built on Google's Gemma 4, a free AI model, and on the listening part of OpenAI's Whisper speech recognizer. People can try it in a browser demo called Apriel Live.
Source: ServiceNow, Apriel Voice 27B Preview model page; ServiceNow blog, "Introducing Apriel Voice 27B Preview," Oct. 29, 2026.