The latency target is 440 milliseconds. That is the number Deepslate’s engineers are chasing, the time from a user’s last spoken word to the AI’s first audio response [Deepslate, retrieved 2026]. For a Berlin startup taking on OpenAI and Google in real-time voice AI, it is the technical heartbeat of the bet. The other is geography: every inference run happens on servers in Germany.
Deepslate, founded in 2024 by Paskal Paesler and Jan Brachthäuser, is building proprietary speech-to-speech models. The architecture processes spoken language directly into spoken responses, bypassing the conventional pipeline of speech-to-text, a large language model, and text-to-speech synthesis [Startbase, October 2026]. The company claims this end-to-end approach, trained specifically on European speech patterns and cultural context, yields more natural conversations with preserved tone and dialect [Business Insider Germany / Gründerszene, October 2026]. Its seed round, a €7.7 million ($8.29 million) raise led by Munich-based 42CAP and closed in October 2026, is the fuel for that technical sprint [Tech.eu, October 2026].
The technical wedge: skipping the text middleman
Most voice AI systems are assembly lines. Audio is transcribed, the text is processed by a model like GPT-4, and a separate text-to-speech engine reads the answer back. Each step adds latency and can strip away vocal nuance. Deepslate’s system, as described in coverage of its funding, consists of a speech encoder, a reasoning core based on an open-weights language model, and a speech decoder [Tech.eu, October 2026]. The model is trained to understand the semantic meaning and emotional cadence of speech directly from audio waveforms.
The claimed advantage is twofold. First, speed. By avoiding serial processing stages, the company targets sub-500-millisecond latency, a figure crucial for natural, interruptible dialogue. Second, character. The model is designed to preserve a speaker’s tone, emphasis, and intonation, which is particularly relevant for the European dialects and languages it supports [Business Insider Germany / Gründerszene, October 2026]. The product started as an AI meeting assistant in 2023 before the founders pivoted to offer the core model as a platform [Startbase, October 2026].
A bet on European data sovereignty
The commercial pitch is not just about better audio. It is about location. Deepslate explicitly markets itself as a Germany-hosted alternative to US giants like OpenAI, Google, and ElevenLabs [Business Insider Germany / Gründerszene, October 2026]. For enterprise buyers in sectors like insurance, contact centers, and platform providers,the customers Deepslate says are already using its tech in production,data residency and GDPR compliance are non-negotiable procurement requirements [Startbase, October 2026].
The company sells access via a self-service platform and API, with pricing reported at up to €0.02 per conversation minute [Startup Researcher, October 2026]. It also offers volume discounts and self-hosting options for customers who need the model deployed within their own infrastructure. This flexibility is a direct appeal to European enterprises wary of sending sensitive customer voice data to American cloud regions.
The team from Minecraft to machine learning
The founding story is unconventional. Co-founders Paesler and Brachthäuser began programming at age 11 and first collaborated building and operating large Minecraft servers in Europe [Tech Funding News, October 2026]. That experience in running scalable, real-time online systems provided a practical foundation in distributed infrastructure. They met while programming for the game, a connection that eventually led to the formation of Deepslate [Startup Insider | Podcast on Spotify, retrieved 2026].
The team has since added go-to-market expertise. David Jankowski is listed as Founding Strategy and Growth, and Marius Kühlmann served as a GTM advisor from late 2025 to early 2026 [RocketReach, retrieved 2026][Malt, retrieved 2026]. The €7.7 million seed round, which also included Alstin Capital and SIVentures, is earmarked for model training, European infrastructure expansion, hiring, and sales [Startbase, October 2026].
| Role | Name | Note |
|---|---|---|
| Co-Founder & CEO | Paskal Paesler | Began programming at age 11; background in scalable online systems. |
| Co-Founder & CTO | Jan Brachthäuser | Began programming at age 11; technical co-founder. |
| Founding Strategy & Growth | David Jankowski | Listed in company leadership [RocketReach, retrieved 2026]. |
| GTM Advisor (Nov 2025-Jan 2026) | Marius Kühlmann | Provided early go-to-market guidance [Malt, retrieved 2026]. |
The scale-up challenge
For all its technical ambition and regulatory tailwinds, Deepslate’s path is lined with execution risks that only intensify at scale. The primary challenge is the sheer computational cost and complexity of training and serving massive, proprietary speech-to-speech models. Competing on latency with hyperscalers requires not just algorithmic cleverness but also immense capital efficiency in inference.
- The benchmark gap. The company claims it beats OpenAI and Google in voice AI benchmarks, but these are self-reported and not yet independently verified at scale [Startup Insider | Podcast on Spotify, retrieved 2026]. Winning enterprise deals requires third-party validation and case studies with named, referenceable customers, which are not yet public.
- The capital race. The €7.7 million seed is substantial, but it is a fraction of the resources available to its stated competitors. Continuous model improvement and infrastructure scaling are capital-intensive games. The next funding round will be a critical test of investor belief in the unit economics of a Europe-first AI model.
- The feature frontier. While focusing on core speech-to-speech, the company must also build out the enterprise-grade tooling,detailed analytics, robust security audits, integration libraries,that large buyers expect. A social media manager role advertised in Berlin suggests a push to build brand awareness, but the heavier lift is in sales engineering [Arbeitnow, retrieved 2026].
The technical breakdown is straightforward: the architecture trades the flexibility of a modular, best-of-breed text layer for the potential latency and authenticity gains of a unified audio model. It is a bet on specialization. The sober assessment is that the wheels could come off if the model’s performance or cost-per-inference does not consistently outperform the assembled pipeline of larger rivals, or if the European data sovereignty argument proves insufficient to overcome a significant feature gap.
The next twelve months
The immediate roadmap is clear from the funding announcement. Deepslate will deploy its capital into more model training, expanding its German infrastructure footprint, and building its commercial team. The key milestone to watch will be the announcement of its first major, publicly named enterprise customer. Another will be an update on supported languages,currently cited as 27,and demonstrable latency improvements under load [Deepslate, retrieved 2026].
The company has placed a precise technical bet in a geopolitically charged market segment. Its success hinges on executing that bet with the discipline of engineers who understand that 440 milliseconds is not just a metric, but the entire experience.
Sources
- [Startbase, October 2026] Deepslate erhält 7,7 Millionen Euro für europäische Sprach-KI | https://www.startbase.de/news/deepslate-erhaelt-77-millionen-euro-fuer-europaeische-sprach-ki/
- [Business Insider Germany / Gründerszene, October 2026] Maschmeyer investiert in Berliner Sprach-KI-Startup - seht hier das Pitchdeck | https://www.businessinsider.de/gruenderszene/business/maschmeyer-investiert-in-berliner-sprach-ki-startup-seht-hier-das-pitchdeck/
- [Tech.eu, October 2026] Deepslate secures €7.7M for its European voice AI platform | https://tech.eu/2026/10/01/deepslate-secures-eur77m-for-its-european-voice-ai-platform/
- [Tech Funding News, October 2026] Deepslate raises $8.8M seed to build a speech-to-speech AI model in Europe | https://techfundingnews.com/deepslate-raises-8-8m-seed-to-build-a-speech-to-speech-ai-model-in-europe/
- [Startup Researcher, October 2026] Deepslate Raises €7.7 Million Seed Round to Build Europe’s Voice AI | https://www.startupresearcher.com/news/deepslate-raises-7-7-million-seed-round-to-build-europe-s-voice-ai
- [Deepslate, retrieved 2026] Deepslate | https://deepslate.eu/
- [Startup Insider | Podcast on Spotify, retrieved 2026] Von Minecraft zum KI-Durchbruch: Wie Deepslate OpenAI in Benchmarks schlägt - Paskal Paesler | https://open.spotify.com/episode/0eTe826ILl4QF0JSVREtDW
- [RocketReach, retrieved 2026] David Jankowski - Deepslate | LinkedIn | https://www.linkedin.com/in/david-jankowski-0bb88a1bb/
- [Malt, retrieved 2026] Marius Kühlmann - Various Companies | LinkedIn | https://www.linkedin.com/in/marius-k%C3%BChlmann-1a2b3c/
- [Arbeitnow, retrieved 2026] Social Media Manager - Berlin | https://www.arbeitnow.com/jobs/companies/deepslate/social-media-manager-berlin-228354