Dolphin's Positioning and Core Technology Strengths
Dolphin is an intelligent speech processing system jointly developed by DataoceanAI and Tsinghua University, and its core positioning is to solve the speech recognition challenges in complex Asian language environments. The model adopts an advanced CTC-Attention hybrid architecture, in which the encoder uses the innovative E-Branchformer structure and the decoder is based on the Transformer framework, which is specifically optimized for the acoustic and grammatical features of Asian languages.
The major breakthroughs at the technical level are reflected in the processing power to support 40 Asian languages and 22 Chinese dialects; based on more than 210,000 hours of multi-source training data (including both proprietary and public datasets); and the use of a unique two-layer tagging system (e.g., ) to accurately differentiate between linguistic and regional variants. Compared to general-purpose speech recognition models, Dolphin's recognition accuracy in Asian languages, especially Chinese dialects, improves significantly, with the SMALL model reducing the error rate to 25.21 TP3T.
This answer comes from the articleDolphin: Asian Language Recognition and Speech-to-Text Modeling for Asian LanguagesThe
English 
简体中文 
日本語 
Deutsch 
Português do Brasil