8 hours ago
1
ARTICLE AD BOX
Talksign, a Nigeria- and UK-based artificial intelligence (AI) research and product company, has launched two AI-powered models that enable real-time translation between American Sign Language (ASL) and text or speech.
Palm 1.0 interprets ASL into text or speech with 84.2% semantic accuracy, while Echo 1.0 converts written or spoken language into photorealistic ASL video, generating avatars in real time with minimal delay.
The two translators, launched on May 20, build on Talksign’s first foundation model, Talksign-1, introduced in February, which provided basic bidirectional communication by translating 250 ASL signs into speech or text and converting spoken or typed words into sign language video sequences.
Talksign-1, however, was limited to isolated signs and could not interpret continuous sentences or fingerspelling, a limitation that Palm 1.0 and Echo 1.0 aim to overcome.
The launch comes as part of a growing effort to address a global communication gap affecting millions of people with hearing loss. According to the World Health Organisation, over 430 million people worldwide have disabling hearing loss, and tens of millions use sign language as their primary mode of communication.
Most digital tools—from video conferencing platforms to public service kiosks—still assume users can hear and speak, creating barriers to everyday participation in society.
Palm 1.0 is Talksign’s sign-to-text model, designed to translate ASL into written or spoken language in real time. The company said the model achieves 84.2% semantic accuracy, meaning it captures the intended meaning of a signer’s gestures, and 79.6% word-level accuracy, approaching fluency for a visual language that has historically lacked large-scale training data.
It was trained on over 71,000 ASL samples and was built on a transformer-based architecture with a system called SAGE (Spatial Attention Graph Encoder), which tracks 133 anatomical landmarks on the body, including hands, head, and shoulders. According to Talksign, this allows Palm 1.0 to interpret signing as a continuous, contextual conversation rather than as isolated gestures.
“Palm 1.0 is the first model we are confident putting into the hands of Deaf users at scale,” said Edidiong Ekong, Talksign’s CEO and co-founder. “The next step is putting it everywhere a Deaf person needs to communicate: on phones, smart glasses, in classrooms and hospitals.”
The company said Echo 1.0, the second product, was trained on 94,410 ASL sentence pairs, running through the dataset 15 times to improve accuracy and fluency. It converts written or spoken language into ASL video using photorealistic avatars. The company said it outputs 30 frames per second video with translation latency of approximately 29 milliseconds, making the signing appear in real time.
The model translates English text into ASL gloss, preserving ASL grammar and word order rather than producing a literal, word-for-word transcription. Each gloss token is then matched to a high-fidelity 3D motion sequence and rendered by a neural engine. Echo 1.0 also allows personalised avatars to be generated from a single reference photo, which the company said helps make signing feel natural.
The translation process for Echo 1.0 operates in three stages: speech recognition or direct text input, conversion of English text to ASL gloss, and rendering of gloss tokens into photorealistic video. This workflow allows the model to produce fluent, visually coherent signing rather than a word-for-word transcription of English.
Kazi Mahathir Rahman, Talksign’s co-founder and CTO, said the models also open the door to “signing as a first-class interface” for human-computer interaction, enabling new ways for Deaf users to interact with AI systems without relying on speech or keyboards.
Talksign said the new models are intended for contexts where professional sign language interpreters are scarce or unavailable, including emergency alerts and live news broadcasts. The company said both models were trained on datasets reviewed with Deaf advisors, educators, and accessibility advocates. Landmark extraction occurs on the user’s device, meaning only processed data points—not raw video—are sent to company servers.
The company acknowledged that both models currently have limitations. Echo 1.0 accepts only English input, with Spanish, French, and Arabic planned for future versions.
Specialised vocabulary in fields such as medicine, law, and engineering requires additional fine-tuning, and multi-word ASL phrases are only partially modeled. Palm 1.0, while capable of sentence-level interpretation, is not yet optimised for all continuous signing contexts.
Talksign said it plans to expand the models’ capabilities in upcoming versions, including support for additional sign languages such as British Sign Language, Deutsche Gebärdensprache, and Nigerian Sign Language.
The company said the full rollout of Palm 1.0 and Echo 1.0 on the desktop app and Meta Ray-Ban smart glasses is scheduled for August 20, 2026.
With the launch of the 2 products, Talksign joins other AI-powered accessibility platforms, including SignVrse, whose flagship product Terp 360 provides real-time translation of spoken language into sign language using a hyper-realistic 3D avatar, using artificial intelligence to reduce communication barriers for people with hearing impairments.
English (US) ·