Deepslate raises €7.7M seed for European speech-to-speech AI

Berlin-based Deepslate has raised a €7.7M seed round led by Munich’s 42CAP, joined by Alstin Capital, returning investor SIVentures and several business angels, to train Opal, its speech-to-speech AI model, on European languages and European infrastructure.

Announced on 1 October 2026, this is the week’s clearest signal that Europe intends to compete at the model layer of voice AI, not just the application layer above it. Below, I lay out the round, the architecture behind it, and why a seed-stage Berlin lab selling sovereignty by the minute deserves a close listen.

A Munich cheque for a Berlin model lab

42CAP, the Munich-based seed fund, leads the €7.7M round. Alstin Capital joins, SIVentures returns from the previous round, and several business angels complete the table. The amount was announced in euros; the $8.8M figure in circulation is a conversion, nothing more.

Per the announcement, the money goes to expanding model training and the company’s European data programmes – German street names, personal names and dialects are named explicitly – alongside cutting latency further, improving voice quality, building out sales and marketing, and scaling production infrastructure across European data centres.

Founded by Paskal Paesler (CEO) and Jan Brachthäuser (CTO), Deepslate already sells into insurers, contact centres and platform businesses – the rooms where a mangled surname or a transcript in the wrong cloud costs real money.

One model where everyone else runs three

Most voice agents are a relay race: speech-to-text, then a language model, then text-to-speech, with every baton pass adding delay and error. Deepslate‘s Opal is a single end-to-end system – a speech encoder, a reasoning core built on an open-weights language model, and a speech decoder, all three trained in-house. Audio in, audio out, no text in between.

The published numbers back the design: a 440 millisecond response time, ranked the fastest of any speech-to-speech model in the Artificial Analysis benchmark (September 2026), an 85% reasoning score on Big Bench Audio, and the best error rate for European languages on the CoVoST2 benchmark, per the company’s figures of October 2026. Release pricing is €0.02 per conversation minute, which the company sets against the €0.06-0.15 it says comparable US models charge.

The distribution is the European tell: Opal runs on Deutsche Telekom’s German cloud or fully on a customer’s own premises, ISO 27001 certificate included. “Data sovereignty is not a nice-to-have, it is a prerequisite,” said Paesler in the announcement. For the insurers and contact centres on the customer list, that sentence is the pitch.

A 29% curve and a lead investor on a streak

The market Opal feeds is steep on every reading: voice AI overall is sized at $11.71bn in 2026 and projected to reach $32.47bn by 2030, a 29% CAGR (The Business Research Company, September 2026). One layer up, voice AI agents compound even faster – from $2.32bn in 2025 to a projected $27.45bn by 2032, a 42% CAGR (MarketsandMarkets, 2026). The agents need models, and a European agent increasingly wants a European one.

Our own records frame the cheque. Ten days ago, Lyon’s Reecall raised €3.5M to sell white-label voice AI to software vendors – the application layer of exactly this stack. The same day, Kontext took a $4M seed, also led by 42CAP, to police what AI agents do at runtime. Add Osavul’s Series A of 1 October, where 42CAP sits as an existing backer, and the Munich fund has appeared on these pages three times in ten days. That is not deal flow; that is a thesis.

Sovereignty grows a price tag

So what does this round signal? That European data sovereignty has stopped being a compliance slogan and started being a line on the spec sheet, with a price attached. Deepslate does not ask customers to pay a premium for the European model; at €0.02 a minute it undercuts the US alternatives it benchmarks against. What we are seeing across the recent rounds is the voice stack separating into layers – models, agents, guardrails – and European seed funds backing each layer in the same weeks, largely with European money.

The good news is the constraint is doing the work. Having to run on EU infrastructure, in European languages, at contact-centre prices pushed a small Berlin team towards the end-to-end architecture the larger labs are only now converging on. Watch whether the big US voice platforms answer with EU-hosted offers, and whether Opal holds its benchmark lead once they do. Every round in the stack is on our fundraising data page.

Europe keeps hearing it cannot train competitive models. A Berlin lab just raised €7.7M to disagree – in German, dialects included. Let’s listen closely.