The Pulse
Deepslate Raises €7.7 Million to Build European Speech AI
Berlin-based Deepslate has raised €7.7 million in seed funding led by 42CAP to develop its speech-to-speech AI models. The company plans to expand European-language training data, improve its models and scale infrastructure in European data
AI.info Team · Written with AI ·
Deepslate’s speech model Opal starts generating audio in 440 milliseconds, according to Artificial Analysis’s benchmark. On October 1, the Berlin-based company announced a €7.7 million seed round to build out that speech-to-speech technology, with Munich investor 42CAP leading the financing.
Alstin Capital, existing investor SIVentures and several business angels also joined the round, Deepslate said. The company plans to spend the capital on European training data, model development, infrastructure and hiring in sales, marketing and engineering. The announcement puts a specific European focus on an area where customer-service systems often rely on a chain of models hosted across different services.
One model takes speech in and sends speech back
Many voice systems first turn speech into text, pass that text to a language model, then convert the answer back into audio. Deepslate says its system handles spoken input and output directly, aiming to avoid the delays and loss of vocal cues that can come with those separate steps. Those cues include emphasis, dialect and the way a phrase is delivered.
Tech.eu describes Opal as three connected components: a speech encoder, a reasoning core based on an open-weights language model adapted for specific languages, and a speech decoder. The company says it can replace the underlying language model without retraining the full system. That design gives Deepslate a way to update the reasoning layer while continuing to develop the speech components for European languages.
European data and hosting are part of the pitch
The funding will support additional training data, with the company highlighting German street names, personal names and dialects. Deepslate also plans to expand production infrastructure in European data centres and improve latency and voice quality. For customers handling calls, errors in names and local pronunciation can be as noticeable as a slow response.
Deepslate says insurers, contact centres and platforms already use its technology in production, but has not named those customers publicly in the funding announcement. It offers a self-service platform and API, while enterprise and platform customers can seek larger deployments or self-hosting options. The company says its technology is hosted within the EU and is ISO 27001 certified.
“For our customers, data sovereignty is not a nice-to-have, it is a prerequisite. And either it can be verified or it is worthless. That is why we disclose where the computing happens, who our subprocessors are and what is still open,”
Paskal Paesler, co-founder of Deepslate, as quoted by Tech.eu
What the 440-millisecond result measures
Artificial Analysis lists Opal’s time to first audio at 0.44 seconds, the lowest figure on its speech-to-speech comparison page. That is the time until the model begins producing audio, not a measure of the full time needed to complete a conversation. Deepslate’s co-founder David Jankowski also cited the 440-millisecond result in the funding announcement, describing it as the fastest speech-to-speech model.
The benchmark offers one measure of responsiveness, but it does not by itself establish how well a system handles every live exchange or performs across all European languages. Deepslate says its model achieved the best error rate for European languages in the CoVoST2 comparison field, while the company’s funding plans point to further work on regional data and voice quality. For now, the concrete deployment claims are production use by unnamed insurers, contact centres and platforms, alongside EU hosting and self-hosting options.
From benchmark speed to customer deployments
The seed round backs both the model and the operating capacity around it: training data, European compute infrastructure and a larger commercial team. Those investments target the practical requirements of voice agents that handle real customer calls, where a quick first response has to be paired with accurate recognition and dependable operation. Deepslate has not disclosed customer names or revenue, leaving the scale of its current deployments unspecified.
The immediate measure of the funding will be whether the company can expand those deployments while improving performance on the languages and accents it says it is targeting. The model’s 440-millisecond benchmark result is a defined speed measurement; the company’s next work includes German names, addresses and dialects, plus more infrastructure for production call volumes.