Deepslate
Category: Voice / Speech AI
Berlin-based voice AI model lab building Europe's own native speech-to-speech AI models with EU data sovereignty. Deepslate was founded in 2024. The company is led by Paskal Paesler. Based in Templin, Brandenburg, Germany. Team size: 11-50. Total funding raised: $8.8M. Latest round: Seed. Key investors include 42CAP (lead), Alstin Capital, SIVentures (prev. Smart Infrastructure Ventures), business angels (Feliks Eyser, Prof. Dr. Roland Fassauer, Jonas Suedfels, Dr. Kyung-Hun Ha, Richard Lohwasser).
- Founded
- 2024
- Headquarters
- Templin, Brandenburg, Germany
- Team size
- 11-50
- Total funding
- $8.8M
Value proposition
A single end-to-end speech-to-speech model (audio in, audio out) that delivers ultra-low latency (440ms), 27 European languages, and full EU data sovereignty via EU-hosted or self-hosted/air-gapped deployment.
Products and solutions
Opal speech-to-speech model, self-service platform & API (app.deepslate.eu), SDKs and plugins (LiveKit, Pipecat, Python, Node.js, WebSocket, WebRTC, SIP), voice cloning, enterprise self-hosting/on-premises and air-gapped deployment, Zero Data Retention option.
Unique value
Native speech-to-speech architecture (no STT→LLM→TTS chain) that preserves tone, prosody and dialect, combined with EU-only hosting and verifiable data sovereignty — a European alternative to OpenAI, Google and ElevenLabs.
Target customer
European enterprises in regulated industries — insurers, banks, contact centers, public authorities — plus voice AI platform providers and developers.
Industries served
Insurance, Banking/Financial Services, Contact Centers / Customer Service, Public Sector, Voice AI platforms, Automotive/Devices/Robotics (future)
Technology advantage
Three in-house-trained components: a speech encoder, a reasoning core based on a post-trained open-weights LLM, and a speech decoder, linked by trainable projectors. Decoupled architecture lets the base model be swapped by retraining only the projectors (days vs months). Fastest time-to-first-audio (440ms) per Artificial Analysis (Sept 2026); ~85% on Big Bench Audio; best CoVoST2 error rate for European languages (self-published methodology). ISO 27001 certified.
How they differentiate
Single native speech-to-speech model rather than a stitched STT/LLM/TTS pipeline; European-language focus and EU data sovereignty (Telekom Cloud / self-hosting / air-gapped); competitive pricing from 2 ct/min; fastest measured latency.
Main competitors
OpenAI (GPT-Realtime), Google (Gemini Live), ElevenLabs, also Gradium, Deepgram, Synthflow, Cartesia
Key partnerships
Deutsche Telekom Cloud (model hosting in Germany), Google Cloud AI Platform & Compute (startup program), LiveKit and Pipecat (integration plugins)
Notable customers
Insurers, contact centers and platform providers using the model in production (specific customers not publicly named)
Major milestones
2024: founded in Berlin, 2025: pre-seed led by SIVentures, Sept 2026: Opal ranked fastest speech-to-speech model by Artificial Analysis (440ms), Oct 2026: €7.7M seed round led by 42CAP
Growth metrics
~15-17 employees (Oct 2026); 27 languages supported; 440ms time-to-first-audio; pricing from 2 ct/min; €15 free credits (750 minutes) for new users; ISO 27001 certified
Market positioning
European sovereign voice AI model lab — positioned as the EU-hosted, GDPR-compliant alternative to US incumbents for regulated enterprises; early commercial stage with production deployments.
Geographic focus
Europe (DACH focus), competing against US (OpenAI, Google, ElevenLabs, Deepgram) and European (Gradium, Synthflow) players
Patents and IP
Not publicly disclosed; proprietary in-house speech-to-speech model architecture (encoder + reasoning core + decoder with trainable projectors).
About Paskal Paesler
Ex-Stromee (Lead of Development, led team of 10+); co-built one of Europe's largest Minecraft servers; started programming at age 11; studied at Technische Universität Berlin
Latest news about Deepslate
More Voice / Speech AI companies
Official website: https://deepslate.eu/