A Berlin startup claims it has built the fastest speech-to-speech AI model on the planet, and it just locked down €7.7 million in seed funding to prove it. Deepslate, founded in 2024, pulled in the investment with Munich’s 42CAP leading the round alongside Alstin Capital, SIVentures, and a group of business angels.
The company’s speed claims come with receipts. Independent benchmarking firm Artificial Analysis measured Deepslate’s time to first audio at 0.44 seconds, putting it just behind Krafton’s Raon. Deepslate’s own figures suggest an even snappier 250 milliseconds from the moment a user stops speaking to the model’s first audible response.
Co-founders Pascal Paesler and Jan Brachthäuser both started writing code at age 11 and later operated two of Europe’s largest Minecraft servers. Paesler eventually led a ten-person team at Stromee, where his interest in AI deepened. The pair spent two years exploring projection models before shifting Deepslate’s entire architecture toward speech-native processing.
Most voice AI systems, including ElevenLabs, run a three-stage pipeline: speech-to-text, language modeling, then text-to-speech. Paesler contends that approach destroys prosody, timbre, and emotional nuance. Deepslate instead uses a speech encoder, an open-weights reasoning core post-trained for specific languages, and a speech decoder. Trainable projectors connect these components, which means swapping in a better base model takes days rather than months. The company has not revealed which base model it uses.
Deepslate’s Opal model scores about 85% on Big Bench Audio, a reasoning benchmark delivered as speech. The company also claims the best CoVoST2 error rate, though that figure relies on Deepslate’s own published methodology.
European sovereignty anchors the pitch. All inference runs in Deutsche Telekom’s German cloud, and customers can deploy on their own infrastructure, including air-gapped environments. ISO 27001 certification opens doors with banks, insurers, and public agencies that cannot use US cloud providers.
Seventeen employees now power the startup. Competitors like OpenAI, Google, and Grok dwarf Deepslate in resources. Yet investors Andreas Schenk of Alstin Capital and Julian von Fischer of 42CAP both point to the team’s ability to scale latency-critical systems for millions of users. Whether that edge survives as tech giants accelerate their own real-time voice releases remains the central question.
