Deepslate raises $8.8M seed to build a speech-to-speech AI model in Europe
Tech Funding News Sofia Chesnokova ● Covered by 5 sources
Deepslate raised €7.7M to build a speech-to-speech model in Berlin. It says its setup is faster than rivals and keeps data on German servers.
Based on reporting by Tech Funding News, Sofia Chesnokova — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Deepslate has raised €7.7 million in seed funding, with Munich-based 42CAP leading the round. Alstin Capital, SIVentures and several business angels also joined in. The Berlin startup is building its own speech-to-speech model and runs it on servers in Germany.
That matters because Deepslate is pushing against the usual voice AI pipeline. Founder Paskal Paesler says most of the market, including ElevenLabs, still breaks speech into three steps: speech-to-text, language modelling and text-to-speech. Deepslate’s pitch is that this chain throws away voice details along the way — the sentiment, the way words are pronounced, the prosody, the timbre. So the company went after a single model that can hear and speak directly.
The technical setup is a mix of a speech encoder, a reasoning core based on an open-weights language model that has been post-trained for specific languages, and a speech decoder. Trainable projectors connect the parts. If a better base model comes along, Deepslate says it only has to retrain the projectors, cutting retraining from months to days. The base model itself has not been disclosed.
On speed, Deepslate says it takes 250 milliseconds from the end of a user’s sentence to produce a reply. Artificial Analysis measured a time to first audio of 0.44 seconds, putting it just behind Krafton’s Raon. The company also says Opal, its model, scores about 85% on Big Bench Audio, and claims the best CoVoST2 error rate based on its own methodology. Deepslate says insurers, contact centres and other platforms are already using the model in production, though it does not name them.
The sovereignty angle is doing a lot of work here. Deepslate runs on Deutsche Telekom’s cloud in Germany and can also be deployed on customer infrastructure, including air-gapped environments. It has ISO 27001 certification, which the company says matters for insurers, banks and public authorities that can’t use US-based cloud services. And in Europe, that kind of hosting story may be the product as much as the model itself.
My take — AI-written commentary, not fact-checked reporting
This is the right bet: Europe does not need more voice AI wrappers pretending to be infrastructure. It needs models that can run locally, pass procurement checks, and avoid the usual cloud hostage situation. The funny part is that “sovereign AI” sounds like a slogan until a bank or insurer shows up with a compliance team and a cheque.
Read more about this at: Tech Funding News