KugelAudio
Real-time text-to-speech model you can self-host
Start new thread

Kugel Ear - Speech-to-text that only asks when it's unsure

by
•
Speech-to-text for voice agents that only asks about the word it's unsure of. Your bot gets that word flagged, with the other spellings it heard, and asks once, like a person would, instead of making callers spell everything. €0.48 per hour.

Add a comment

Replies

Best
Maker
📌
Another week, another KugelAudio release, this time its we’re revolutionizing Speech-to-Text. Voice agents either guess, or make you spell everything out letter by letter: Kugel Ear knows when it's unsure. Instead of guessing, your agent asks about just that word, or offers the options it heard: “Viktor with a K or a C?” In our tests it flagged 86% of the words it got wrong. It's also one of the most accurate speech-to-text models out there, built for European languages: on customer-support calls it got 93.4% of words right across 14 accents, ahead of ElevenLabs, Speechmatics, OpenAI and Deepgram. That accuracy helps any transcript; the asking back is for live calls, where there's someone to ask. All for €0.48 per hour. Try it on your hardest names and numbers and tell us where it breaks.


How is a correction utterance (e.g. spelling out a word) processed? Does STT confidence directly influence the second transcription, or only indirectly through dialogue management?

Love the idea of a speech model that flags only the word it is unsure about and asks once, just like a person would, instead of making callers spell everything out.

The "ask when unsure" thing is so obvious in hindsight. Every transcription tool I use confidently butchers names. How often does it actually interrupt you in a normal meeting?