Page 730 - AI for Good Innovate for Impact
P. 730

AI for Good Innovate for Impact



                      2      Use Case Description


                      2�1     Description


                      For more than 27.8 million DHH individuals in China, sign language interpretation is an essential
                      service. However, the number of certificated sign language interpreters in China was less than
                      500 in 2023, representing a huge talent gap in the sign language interpretation market.

                      Meanwhile, developing a sign language translation system is uniquely challenging. Sign
                      language differs fundamentally from spoken Chinese in grammar, structure, and expression,
                      requiring key information extraction rather than direct translation. As a low-resource language, it
                      also lacks sufficient data for AI model training. Moreover, natural sign language communication
                      relies not only on hand gestures but also on mouth movements and body language, demanding
                      complex multi-channel synthesis.

                      To address this issue, Baidu AI Cloud's XiLing digital avatar platform launched solutions
                                                                 [1]
                      including the Baidu AI Sign Language Platform , AI Sign Language All-in-One Machine, and
                      digital avatar, aiming to empower the DHH community to integrate into social life and equally
                      benefit from technological advancements. Unlike earlier solutions that could only translate
                      speech into sign language, Baidu AI sign language interpretation tools realize the bidirectional
                      translation between sign language and text, speech and regular videos, enabling truly natural
                      two-way communication.

                      AI Sign Language Platform provides on-demand automated sign language translation and
                      interpretation through digital avatars [2], while ensuring consistency and accuracy. The Platform
                      supports typical sign language translation scenarios such as bilingual video production and
                      sign language livestream. It integrates AI capabilities like automatic speech recognition (ASR)
                      and sign language translation with a modular business platform and a digital avatar rendering
                      engine[3], enabling efficient data flow from audio/video/text input to final media output via a
                      unified pipeline.

                      The use case addresses these challenges with a three-step approach: converting speech to
                      text, translating text into sign language codes, and generating expressive sign language via
                      digital avatars, thanks to its robust suite of AI capabilities.

                      Additionally, the device is trained on a dataset comprising nearly ten million "natural sign
                      language samples," ensuring its ability to provide accurate and nuanced sign language
                      interpretation. It can achieve a near-field speech recognition rate on mobile phones of over 98%.
                      The AI sign language virtual interpreter also has 98.5% accuracy of mouth shape generation.

                      AI Sign Language All-in-One Machine is the latest offering in XiLing's comprehensive suite of
                      solutions. With a creative and interactive design of dual screens, the Machine achieves two-
                      way real-time translation between sign language and text. As a typical application integrating
                      multimodal AI capabilities, especially in visual recognition and natural language processing,
                      the Machine tackles challenges such as accurately tracking and recognizing sign language
                      against an unsharp background and fast motion of sign language movements.

                      To address the common problem of low recognition rates for long sentences in attention
                      models, XiLing has uniquely innovated with the introduction of the world's first online attention
                      model for voice interaction. This model's streaming attention capabilities not only outperform





                  694
   725   726   727   728   729   730   731   732   733   734   735