Clarity 1 Launches Real-Time Voice Isolation

Clarity 1 tackles a more consequential problem than cosmetic audio cleanup: keeping speech-to-text, turn detection, and language models from processing the room instead of the caller. The idea is compelling, but real-world crosstalk, latency, and speaker-switching tests will matter more than benchmark claims.

• –Target-speaker extraction distinguishes Clarity 1 from basic denoisers; it can use a short reference recording, or default to the loudest speaker. [Python SDK](https://pypi.org/project/kugelaudio/)
• –The model processes 240 ms chunks with roughly 50 ms of processing time, creating a potential delay of up to about 290 ms depending on frame alignment. [Official product page](https://www.kugelaudio.com/en/clarity)
• –Streaming output makes it a natural preprocessing layer before STT and turn detection, potentially reducing false interruptions and transcript contamination.
• –KugelAudio claims the highest DNSMOS scores among its comparisons, but developers should validate performance across accents, music, overlapping speech, and poor phone audio.
• –A free trial through October 30 lowers the barrier to testing Clarity 1 against real caller recordings before production adoption. [Official product page](https://www.kugelaudio.com/en/clarity)