Voice, Speech & Realtime AI

European speech-to-speech AI gets serious: Deepslate raises EUR 7.7M, and why consented speech data is the input that matters

Source: Tech.eu · Oct 1, 2026

Deepslate, a Berlin company founded in 2024, has raised EUR 7.7 million in a seed round led by the Munich investor 42CAP, with Alstin Capital and SIVentures also participating. The company builds speech-to-speech models for European languages — systems that process spoken input directly and respond in speech, rather than converting audio to text, generating a text reply, and reading it back. Doing away with that text round-trip lowers latency and keeps the parts of speech that text throws away: tone, emphasis, pacing and dialect. Deepslate points to sub-half-second latency across more than two dozen languages, hosted inside the EU with European compliance in mind.

The funding is a useful marker for where voice AI is heading, but the more telling detail is what the money is for. Alongside infrastructure and hiring, the company has said part of the raise will go toward expanding its European training data. That is the quiet admission at the centre of the whole voice-AI wave: the hard constraint is less the model architecture than the data it learns from. A speech-to-speech system is only as good as the real conversational audio it has heard — across accents, languages, interruptions and the messy back-and-forth of actual dialogue.

That is where the supply question gets interesting. Clean, consented, richly varied speech is not easy to come by, especially multilingual, technical speech with real conversational structure rather than scripted prompts. Interview recordings are one of the few naturally occurring sources of exactly that: real people speaking at length, being asked follow-up questions, correcting themselves, reasoning out loud — captured as audio paired with transcripts and context. Anonymized so the identifying detail is removed, that material maps directly onto what a speech-model team needs for both training and evaluation.

The takeaway: as European speech-to-speech AI attracts real capital, the competitive edge will sit with the teams that can get consented, multilingual, real-conversation speech data with documented provenance — the same shape of data an interview corpus is built from.

Key Points

• Berlin-based Deepslate raised EUR 7.7M in a seed round led by 42CAP, with Alstin Capital and SIVentures, to build European speech-to-speech models
• Its models process spoken input directly rather than transcribing to text first, which lowers latency and preserves tone, emphasis and dialect — the company cites 440ms latency across 27 languages, hosted in the EU
• Part of the raise is earmarked for expanding European training data, underscoring that speech data — not model architecture alone — is the constraint
• Speech-to-speech systems are trained and evaluated on real conversational audio; consented, anonymized interview recordings are a natural source of multilingual, technical, real-dialogue speech