HSI1 24,805.63 -148.84 228.43B
HSCEI1 8,246.33 -28.45 62.15B
Back    Zoom +    Zoom - Block Traded
XIAOMI-W Open-Sources Industrial-Grade Target Speaker Speech Recognition Large Model
2026-09-11 16:07:48
XIAOMI-W (01810.HK) officially launched and open-sourced the industrial-grade target speaker speech recognition large model Xiaomi-CocktailASR-1, aiming to solve the "cocktail party" problem.

The model adopts an end-to-end LLM architecture and uses a segment of reference audio from the target speaker as a voiceprint prompt, enabling precise extraction and transcription only of the target user's speech in complex multi-speaker environments where multiple people are speaking simultaneously.

It achieved SOTA performance across multiple mainstream multi-speaker benchmark datasets, significantly outperforming existing solutions.
~

AASTOCKS Financial News
Website: www.aastocks.com