Pomegra
Berlin-based Deepslate raised a €7.7M seed round led by 42CAP to build European speech-to-speech AI models that run in EU data centres or fully on-premises.
Key Takeaways
- Deepslate closed a €7.7M (about $8.8M) seed round led by Munich-based 42CAP; valuation is undisclosed.
- Alstin Capital, existing investor SIVentures and business angels also took part in the round.
- The model skips the text step and, per a September 2026 benchmark, returns first audio in 440 milliseconds.
Lead
Deepslate, a Berlin startup founded in 2024, announced a €7.7M seed round on 1 October 2026. 42CAP led the round, with Alstin Capital, existing backer SIVentures and several business angels participating. The company builds speech-to-speech models that process spoken audio directly and answer in audio, and it hosts them inside the EU. Customers can also run them on their own hardware, including air-gapped environments.
What Does Deepslate Actually Build?
Deepslate builds a single end-to-end voice model, called Opal, in place of the usual three-stage chain of speech-to-text, a language model and text-to-speech. The system has three components trained in-house: a speech encoder, a reasoning core based on an open-weights language model, and a speech decoder. The company says this keeps tone, emphasis, dialect and pronunciation that get stripped out when speech is flattened into text.
The pitch targets European speech in particular, with emphasis on German street names, personal names and regional dialects, where text-first pipelines tend to fail. Distribution runs through a self-service platform, an API, volume plans and self-hosting. Insurers, contact centres and platform providers already use it, though Deepslate has not named customers.
How Fast Is It, and Who Verified That?
In a September 2026 test by Artificial Analysis, Deepslate recorded a response latency of 440 milliseconds, the fastest among the speech-to-speech models compared. The company also cites the best error rate for European languages on the CoVoST2 benchmark and roughly 85% on the Big Bench Audio reasoning test. Those are the company's own readings of third-party benchmarks, and the full list of models in the comparison has not been published alongside them.
CEO Paskal Paesler and CTO Jan Brachthäuser, who previously ran large European Minecraft servers, lead the company. Paesler frames data sovereignty as a prerequisite and not an extra: "Either it can be verified or it is worthless." Deepslate holds ISO 27001 certification and runs on German cloud infrastructure from Deutsche Telekom.
Why Does On-Premises Deployment Matter?
On-premises deployment matters because insurers, banks and public bodies in the EU often cannot send recorded customer calls to a US-hosted model. Voice data can carry biometric and health information, so the compliance burden under European data rules is heavier than for plain text. A model that can sit inside a customer's own data centre removes that objection before procurement starts.
The big US labs offer speech-to-speech systems too, from OpenAI, Google and xAI. They are mostly delivered as hosted APIs, which is the gap Deepslate is aiming at. Self-hosting also moves the cost of GPU capacity onto the customer, and the startup has not said what that means for its margins or pricing.
How Crowded Is the Voice AI Market?
The field is crowded and well funded. ElevenLabs carries a valuation reported at $11B, and the French startup Gradium raised a €100M seed round. Deepgram and Synthflow also sell into contact centres. Against those figures, a €7.7M seed is modest, and Deepslate cannot match rivals on spend for training compute or global sales.
The round also implies something about the company's earlier funding. SIVentures backed Deepslate before this round, but the size of that earlier cheque has not been disclosed, so the step-up in valuation cannot be measured. 42CAP leading a seed this size signals conviction in a regional, compliance-led niche over a bid for global consumer voice.
What Will the Money Fund?
Deepslate plans to improve its speech-to-speech models, expand its European training data, grow its sales and marketing teams, and scale infrastructure across European data centres. The company says demand is outpacing its sales capacity, a claim without public figures behind it. Reported team size varies between 15 and 17 people, so hiring will reshape the company quickly.
Outlook
Deepslate has a defined niche, European languages with verifiable data residency, and a latency result that, if it holds up, puts it level with much larger rivals. The test of the seed round is whether named enterprise customers and published revenue follow within the next two funding cycles. The larger risk is that US providers add EU hosting and dialect coverage faster than a 15-to-17-person team can build a lead.
