Deepgram is optimizing its Nova-3 speech-to-text model for the Hexagon NPU in Qualcomm’s Snapdragon X platform. This allows enterprise speech recognition to run locally on the device, without a cloud connection. Developers and manufacturers can now build real-time voice applications.
Much of voice AI has traditionally relied on a cloud-based architecture. Each interaction requires audio to leave the device, travel to the cloud, be processed elsewhere, and return. This process introduces latency and raises privacy concerns. Deepgram aims to break away from this model by performing speech recognition on the device itself.
The collaboration focuses on the Snapdragon X Series, the platform Qualcomm is using to target AI PCs. By optimizing Nova-3 for the Qualcomm Hexagon NPU, Deepgram promises faster, more reliable voice experiences that work even where connectivity isn’t guaranteed, such as in a car, on a factory floor, or with an XR headset.
The first Snapdragon X chips achieved around 45 TOPS, well above the 40 TOPS that Microsoft sets as the minimum requirement for a Copilot+ PC. The newer Snapdragon X2 generation doubles that to approximately 80 TOPS.
That computing power enables running more complex ASR models locally rather than just lightweight edge variants. Qualcomm has long cited real-time transcription, translation, and voice isolation as core tasks for the NPU. Deepgram is now building on that with a model specifically aimed at enterprise use.
What Nova-3 can do
According to Deepgram, Nova-3 is the first voice AI model with real-time multilingual transcription. The company claims a word error rate of 6.89 percent on production audio, which represents a 24.7 percent lower error margin than the second-best competitor. Users can also customize the vocabulary themselves without retraining the model.
Qualcomm sees opportunities for a wide range of industries in this combination. “The on-device AI capabilities of Snapdragon processors combined with Deepgram’s enterprise-class voice AI uniquely bring those requirements together in a way that opens new possibilities for real-time voice experiences across the device ecosystem,” says Upendra Kulkarni, Vice President of Product Management at Qualcomm. According to both companies, the initiative is intended to pave the way for voice technology in the automotive, mobile, industrial edge, IoT, and wearables sectors.
Tip: IBM integrates Deepgram voice AI into watsonx Orchestrate