Deepslate Raises €7.7M to Keep Speech-to-Speech AI Inside Europe, Claiming the Fastest Voice Model Measured

Deepslate, a Berlin-based voice AI company, raised €7.7 million (about $8.8 million) in seed funding, announced Oct. 1. Munich-based technology investor 42CAP led the round, with participation from Alstin Capital, existing backer SIVentures and several business angels. The company plans to put the capital into model training, a broader European data programme, sales and marketing, and production infrastructure in European data centres.

The technical bet is architectural. Conventional voice assistants run a three-stage pipeline — speech recognition, a language model, then text-to-speech — which adds latency and strips away information such as tone, emphasis and dialect. Deepslate instead uses an end-to-end model that takes audio in and returns audio directly. Its system has three in-house trained parts: a speech encoder, a reasoning core built on an open-weights language model that Deepslate post-trains for specific languages, and a speech decoder that generates the spoken response. Because the language model is a replaceable component, the company can swap in a newer model without retraining the entire system.

On latency, Deepslate says its model responds about 250 milliseconds after the user stops speaking. Artificial Analysis, the independent benchmarking firm, measured time to first audio at 0.44 seconds — a third-party figure the company cites to claim the fastest speech-to-speech model measured by that benchmark as of September 2026. Deepslate also says its technology recorded the best error rate in the comparison field for European languages on the CoVoST2 benchmark. Both performance claims originate with or rely on the vendor's own framing and have not been reproduced publicly at scale.

Data residency is the commercial argument. The models run on servers in Germany, the company says its technology is hosted entirely within the European Union, and it holds ISO 27001 certification. "For our customers, data sovereignty is not a nice-to-have, it is a prerequisite. And either it can be verified or it is worthless," the company said, adding that it discloses where computing happens and who its subprocessors are. Deepslate says the technology is already in production with insurers, contact centres and platforms, and it sells access through a self-service platform and API, with volume licensing and self-hosting options for platform providers and large enterprises.

The funding is being directed in part at the unglamorous problem of European speech data. Deepslate says it will expand its training data programme with a focus on German street names, personal names and dialects — the categories where generic voice models tend to fail — while working to push latency lower and improve voice quality. It will also grow its sales and marketing teams and scale production infrastructure to handle rising call volumes.

Co-founders Paskal Paesler, the chief executive, and Jan Brachthaeuser, the chief technology officer, are positioning the company against much larger rivals, including ElevenLabs — which this week completed a $300 million employee share sale — and the native live-speech features shipped by OpenAI and Google. In that comparison a €7.7 million seed round is small, which makes the differentiated constraint, rather than model quality alone, the thing to watch.

What it means: speech is one of the few AI categories where latency and data residency are both purchasing criteria, because buyers are often regulated — insurers, banks, public services and contact centres handling personal data. That gives European vendors a defensible niche even without frontier-scale compute. The risk is that the niche closes quickly if OpenAI, Google and Anthropic keep improving native speech models and offer EU hosting, in which case Deepslate's advantage narrows to benchmark leadership on European dialects and named-entity accuracy — claims that will need third-party verification to keep winning contracts.