Deepslate Speech AI Software

Deepslate is one of a small number of European companies training a native speech-to-speech model. Most voice agents today turn the caller's words into text, run that text through a language model, and then turn the reply back into audio. Deepslate's model skips the text step. It hears the audio, reasons over it, and answers in audio, which the company argues makes replies faster and keeps information such as tone, emphasis, and accent that a transcript loses.

The model, Opal, has three parts that Deepslate trains itself. A speech encoder reads the incoming audio, a reasoning core based on an open-weights language model is post-trained for particular languages, and a speech decoder produces the spoken answer. Because the language model sits in the middle, Deepslate says it can swap it out without retraining the whole system. In a September 2026 Artificial Analysis benchmark, Opal reached first audio in 440 milliseconds.

Contact centers are the main market. Deepslate pitches the model for European call center reality: German and English code-switching, regional dialects, and data such as IBANs, policy numbers, meter readings, and street names with umlauts. It works with existing CCaaS platforms, SIP trunks, and telephony, and the company says it is in production with large insurers, contact centers, and platforms.

Data location is the other half of the pitch. Deepslate runs in Deutsche Telekom's cloud in Germany, holds ISO 27001 certification, and lets enterprise customers self-host, including in air-gapped environments. Paskal Paesler and Jan Brachthäuser, who met more than ten years ago running large Minecraft servers, lead the company. Its 7.7 million euro seed round in October 2026 was led by 42CAP, with Alstin Capital, SIVentures, and business angels.

1. Native speech-to-speech model (Opal) with no separate STT, LLM, or TTS stage

2. 440 ms time to first audio in Artificial Analysis testing, September 2026

3. 27 languages with training aimed at European accents, dialects, names, and addresses

4. Reliable capture of IBANs, policy numbers, meter readings, and other structured data

5. Server-side voice activity detection and native interruption handling

6. Function tools and webhooks for live data lookups during calls

7. Hosted voices and voice cloning

8. LiveKit and Pipecat plugins, Python and Node SDKs, WebSocket, WebRTC, and SIP

9. EU hosting in Germany, ISO 27001, GDPR compliance, and optional Zero Data Retention

10. Full self-hosting, including air-gapped and on-premise deployment

Deepslate sells pay-as-you-go access to its full speech-to-speech model. The regular price is 10 euro cents per minute, and a launch offer cuts that to 2 cents per minute for early sign-ups for a limited time. New accounts get 15 euros of free credit (750 minutes at the launch price) with no credit card, and 10 concurrent calls are included. Extra capacity costs 99 euros a month for 10 more lines or 249 euros a month for 30 lines in total. Zero Data Retention is an add-on. Enterprise plans with guaranteed capacity, an SLA, Zero Data Retention, and self-hosting are priced through sales. Prices exclude VAT.

Deepslate does not name its customers publicly. It says the model is in production with large insurers, contact centers, and platforms.