Deepslate raises €7.7M seed for speech AI

Deepslate has raised €7.7 million in seed funding in a round led by 42CAP, with Alstin Capital, SIVentures and several business angels also participating, according to reporting published on October 1 and October 9. The Berlin-based startup is building speech-to-speech AI for European languages, a category that aims to move from audio in to audio out without first routing every interaction through separate speech recognition, text and text-to-speech layers.

What does Deepslate build?

Deepslate builds a native speech-to-speech model that processes spoken language directly, rather than converting speech to text first. Tech.eu described the system as end-to-end: audio goes in and audio comes back out, with the company saying the architecture is designed to reduce delay while retaining tone, emphasis, dialect and other cues in speech.

The company’s system is built around three in-house trained components: a speech encoder, a reasoning core based on an open-weights language model post-trained for specific languages, and a speech decoder. According to reporting from Startbase, the encoder and decoder connect to the language model through trainable projectors, which means the reasoning core can be replaced without retraining the entire system.

That modular design matters for training speed and compute. Startbase reported that if the reasoning core changes, only the projectors need to be retrained, with the company expecting training runs to take days instead of months and to require less computing power than a full omni-model retrain.

Why are investors backing the company?

The round gives Deepslate capital to improve its speech-to-speech models, expand its European training data and scale go-to-market and infrastructure. Tech.eu reported that the company plans to use the funding to expand model training and its European data program, with a particular focus on German street names, personal names and dialects, while also growing sales and marketing and scaling production infrastructure in European data centres.

Deepslate’s products are already in production with insurers, contact centres and platforms, according to Tech.eu. The company offers its model through a self-service platform and API, with volume and self-hosting options for platform providers and enterprise customers. Startbase reported that usage through the self-service platform and API is expected to cost as little as two cents per minute of conversation.

The company also emphasizes deployment choices for customers that care about location and control. Startbase said Deepslate operates its models in Germany on the Telekom Cloud, while customers can deploy the model entirely within their own infrastructure and, on request, without an external connection.

How fast is the model, and how broad is its language coverage?

Deepslate says speed is a core part of the product. Tech.eu reported that the company measured a 440 millisecond response time in an independent Artificial Analysis benchmark, which it said made the system the fastest speech-to-speech model measured there as of September 2026. Startbase added that, in Deepslate’s own measurements, the system reports 250 milliseconds from the end of a user’s utterance to the first syllable of the response, excluding longer-distance network latency.

The company is also promoting quality and language coverage. Tech.eu reported that Deepslate says it achieved the best error rate in the comparison field for European languages in the CoVoST2 benchmark, while Startbase said the model reaches about 85 percent accuracy at its reported response speed in the BIG-Bench Audio benchmark. CX Network also said the company’s website lists 27 supported languages.

Deepslate is ISO 27001 certified, according to Tech.eu, and the company says data sovereignty is a prerequisite for its customers. Tech.eu quoted co-founder Paskal Paesler saying the company discloses where computing happens, its subprocessors and what remains open. The AI Insider reported that Deepslate is headquartered in Templin, Brandenburg, and that it is building speech-to-speech AI models in Europe.

Where did Deepslate come from?

The company’s starting point was an AI assistant developed in 2023 to answer questions during meetings using corporate knowledge, according to Startbase. The founders, Paskal Paesler and Jan Brachthäuser, reportedly found that about three seconds of response latency made natural conversations virtually impossible, which pushed the team toward a lower-latency speech interface.

Startbase reported that Deepslate has a 15-person team. The company’s architecture and infrastructure choices suggest it is trying to balance performance, European-language coverage and deployment control as it scales from product development into broader commercial use. With the new seed capital in hand, the next reported priorities are model training, data expansion, sales and marketing, and more production capacity across European data centres.