Deepslate’s bet is that European voice AI can stay inside the bloc without losing speed, and the Berlin company now has €77M to prove it.
Conventional stacks pass speech through three stages: audio becomes text, a language model composes a reply, and the reply is turned back into audio. Deepslate throws out that relay, running its own speech-to-speech models that take audio in and hand audio back, holding onto tone, emphasis and dialect while shaving latency.
Three in-house-trained pieces make up the system: an encoder for speech, a reasoning core that rests on an open-weights language model Deepslate post-trains language by language, and a decoder that speaks the answer. Swapping the model underneath is possible without rebuilding everything around it.
The company points to numbers. In an independent Artificial Analysis benchmark its model posted a 440-millisecond response time, which Deepslate calls the quickest speech-to-speech system measured as of September 2026, and it claims the best error rate among European languages in CoVoST2. It also holds ISO 27001 certification.
Revenue is already coming in, with insurers, contact centers and platforms running the model in production, bought through a self-service platform and API or self-hosted at higher volume. Sovereignty is the selling point: Deepslate argues its customers treat it as a prerequisite rather than a bonus, and that it counts for nothing unless it can be verified. The new money goes to model training and European data work, particularly German street names, personal names and dialects, plus sales hiring and added capacity in EU data centers.
