Baidu Launches Speech to Sign Language Translation

Article Featured Image

Baidu AI Cloud today launched an artificial intelligence sign language platform able to generate digital avatars for sign language translation and live interpretation within minutes.

Released as part of Baidu AI Cloud's digital avatar platform XiLing, this platform aims to help break down communication barriers for the deaf and hard-of-hearing.

Also released with the platform are two all-in-one AI sign language translators. The translators have been designed for a wide range of use scenarios such as hospitals, banks, airports, bus stations and other public areas.

With the technology enablement brought by AI, the production and operational costs of digital avatars have been reduced to a significant degree, making it possible for AI sign language to go scale and serve more deaf and hard-of-hearing individuals, said Tian Wu, Baidu's corporate vice president..

Deaf or hard-of-hearing individuals who want to study or socialize online without barriers, the platform can be quickly integrated into commonly used mobile applications, websites, and mini programs within a few hours, performing functions like sign language video synthesis and livestream synthesis, text-to-sign language translation, and audio-to-sign language translations. The all-in-one translators are tailored for offline scenarios to improve the accessibility of public services.

Baidu's translators come with two models: a full offline version V3 and a cloud-connected version P3. Both are embedded with core functions of the AI sign-language platform, able to realize speech recognition, speech translation, and portrait rendering.

This full range of functions offers incredible potential for empowering the hearing impaired, who can now visit the hospital and manage the complicated process of registration, consultation, payment, and medicine collection without further assistance. Additional applications will help the hearing impaired travel, dine, and even work independently.

Compared to translations between spoken languages, the sign language translation is more complicated mainly because it is not translated word by word from verbal speech. Instead, the language refinement and word order must be adjusted to show the actual meaning of the sentence. As a relatively rarely-used language, a very limited amount of data on sign language is available for machine learning. It also requires lip language and facial expressions to assist understanding. In real-world settings, solutions are often faced with complex environmental factors making them difficult to deploy. All these practical barriers have posed numerous challenges to the development of AI sign language.

Baidu scientists had to resolve three key challenges: the clarity of speech recognition, the accuracy of sign language translation, and the fluency of sign language movements.

To address speech recognition clarity, the XiLing AI sign language platform uses Baidu's home-grown SMLTA speech recognition model to achieve end-to-end modeling speech recognition integrating acoustics and language. Based on Baidu's deep learning algorithm, targeted training can enable word accuracy in a wide range of fields, such as tourism, medical care, and legal proceedings.

In terms of the accuracy and refinement of sign language translation, Baidu built the first neural network-based sign language translation model that can automatically learn sign language translation knowledge from real data, such as word order adjustment, word mapping, and length control to generate natural sign language that conforms to the habits of hard-of-hearing people.

To ensure the accuracy of the sign language translation, Baidu invited more than 500 scholars and students with hearing loss in China to help enlarge and vet the sign language corpus, with many joining the project as volunteers.

To ensure the fluency of sign language actions, the AI ??sign language platform has sorted nearly 11,000 actions based on the National Universal Sign Language Dictionary with its action fusion algorithm so that all digital sign language gestures have the degree of coherency and expression as human sign language. In addition, with the help of 4D scanning technology, the accuracy of mouth shape generation has been optimized up to 98.5 percent.

SpeechTek Covers
for qualified subscribers
Subscribe Now Current Issue Past Issues