Kugel Ear

Kugel Ear 1 is a speech‑to‑text tool designed for voice agents that need to ask callers only when the model is unsure. It streams audio to a single WebSocket and returns a final transcript when the turn ends, along with word alternatives when the model cannot be confident. The service is priced at €0.48 per hour, exclusive of VAT.

You send the call audio in real time, using a JSON configuration that sets the sample rate to 16 kHz and chooses the “kugelturn” rule for turn detection. During the stream you transmit base‑64‑encoded audio chunks. Kugel Ear 1 immediately signals the end of a turn, which includes the final transcript and the exact moment of termination. After that, it delivers a separate message that lists the words the model marked as unsure, along with the possible spellings and their probabilities. For example, if a caller says “Viktor Brandt,” the tool may flag the name with alternatives such as Viktor (55 %), Victor (31 %), and Wiktor (14 %). The first transcript arrives after a median 72 ms, which is faster than most competitor models measured on the same 280–turn dataset.

The product works in nine European languages, and the language can be specified per turn or left to detection. It is especially useful for customer‑support calls, where the accuracy of names, addresses, and customer numbers matters. On AppTek’s public call‑center dataset of 280 English turns, Kugel Ear 1 achieved 93.4 % correct words, outperforming ElevenLabs Scribe v2 and others. When the audio comes over a phone line, the accuracy remains high at 88.5 %. The tool also reduces false cut‑offs; it cuts callers off least at both 300 ms and 700 ms wait times compared to other systems.

Key features

• Real‑time WebSocket API with single‑stream audio input and final transcript output
• End‑of‑turn detection that waits for the last spoken word before signaling completion
• Word‑level uncertainty flags with alternate spellings and probabilities
• Supports nine European languages, with optional language detection per turn
• Median transcript latency of 72 ms, faster than most benchmarked models
• Pricing of €0.48 per hour, no upfront costs