Skip to content

✌️ An ASL fingerspell recognition and semantic pose retrieval interface (arXiv, GitHub, YouTube)

License

Notifications You must be signed in to change notification settings

kevinjosethomas/sign-language-processing

Repository files navigation

democut.mp4

✌️ ASL ⭤ English Translation with MediaPipe, PointNet, ThreeJS and Embeddings

American Sign Language (ASL) is a complete, natural language that exhibits the same linguistic complexities as spoken languages, including its own syntax, morphology, and grammar that significantly differ from English. However, most existing tools that aim to bridge ASL and English are often developed under the misconception that ASL is identical to English. These tools often prioritize the needs of hearing individuals, simply offering Text-to-Speech (TTS) and Speech-to-Text capabilities (STT). When Deafness is viewed as a "disability", most of these translation tools are simply developed to assist with the inability to hear, but they do not serve any purpose in actually translating between ASL and English.

This project is a prototype that enables translation between American Sign Language (ASL) and English, facilitating communication between ASL signers and individuals who do not understand ASL. It is still far from fully translating the nuances of visual language, but it is designed to respect and preserve ASL as the primary language. The interface provides two main functionalities:

ASL Fingerspelling → English Translation English → ASL Sign Translation
Translates ASL fingerspelling into written English, which is then spoken aloud. Removes the need for Deaf individuals to translate their thoughts into English, and then write them out. Translates spoken English into ASL signs, which are then signed by a moving avatar. Removes the need for Deaf individuals to read written English, and then translate it into visual language.

You can watch a full explanation of the project here. You can read my paper documenting the technical aspects on arXiv.

Table of Contents

Note

As a hearing student with limited ASL proficiency, I recognize that my perspective as a hearing person is limited. My role has been to listen carefully and integrate feedback from the Deaf community, and I have done my best to approach this project with a mindset of learning and understanding, rather than assuming. This project would not have been possible without the active involvement and advice of Deaf individuals and ASL experts who have generously shared their insights. Please see acknowledgements.

Motivation

For over eight years, I tried learning multiple languages, from Sanskrit and Spanish to Hindi and French, yet I could barely maintain a fluent conversation in any of them. When I moved to Vancouver in 2021, I joined Burnaby South Secondary School, which shares its campus with the British Columbia Secondary School for the Deaf (BCSD). This gave me the unique opportunity to study a new kind of language – a visual language – in high school.

ASL wasn't like any of the other languages I had attempted to learn before: It wasn't just about words or pronunciation, but rather learning how to fully express yourself without the tools you typically use. Over the last three years, our ASL class has shown me how I take communication for granted and also helped me notice the many hurdles that are faced by the Deaf community in our hearing-centric society. From my very first week at Burnaby South, I have had many experiences that suddenly remind me of the reasons we learn about Deaf culture and accessibility in ASL class. The mission below is ultimately what I hope to achieve with this project.

Language

Below, on the left is the typical flow of conversation between two people speaking the same language. On the right is the typical flow of a conversation between two individuals who speak different languages. This features either a tool that can translate between the two languages (Google Translate) or a person who knows both languages and serves as a translator. This is also an ASL interpreter's role in a typical conversation between a Deaf and hearing individual.

In contrast, below is the typical flow of conversation between a Deaf and hearing individual when an interpreter is not present. The most commonly utilized process is writing/typing to communicate, or using a TTS/STT tool that assists the hearing individual.

The idea that ASL is merely a visual representation of English is a widespread misconception. In reality, ASL is a distinct language with its own syntax, grammar, and cultural nuances that differ significantly from spoken English. Most existing direct translation tools, similar to the writing/typing translation process I visualized above, are built with this understanding, failing to capture the depth of ASL communication. Furthermore, these tools put an even larger burden on the Deaf individual, requiring them to accommodate and adapt to meet the needs of the hearing individual.

While ASL signers do have a grasp of English, it is often incredibly hard to constantly translate thoughts from ASL into English, and vice-versa. Essentially, most of these translation tools simply transcribe text, and don't serve any real "translation" purpose. They are built to assist deafness as a disability, but not as a culture or as a language.

On the left, I added a visual of the 5 fundamental parameters that define ASL, as well as some examples of each (there are hundreds!). On the right, I added an example of the grammatical difference between the same sentence in English and in ASL. I would recommend reading about ASL Parameters, and also about ASL Gloss & Grammar.

The goal of this project has been to eliminate the extra steps that Deaf people have to take because of how hearing-centric our society is. From the 4 extra steps visualized above, our tool eliminates three 3 steps to a sufficient degree. Here is the new flow below:

Technology

This project uses computer vision, machine learning, and web animation to create a two-way ASL-English translation system. There are two main components to the project:

  • Receptive: Ability to interpret fingerspelling and express as spoken English
  • Expressive: Ability to interpret spoken English and express as ASL signs

Receptive

The receptive component of the project focuses on translating ASL fingerspelling into English. It involves several stages:

  1. Detection: Identifying and tracking hands within the frame
  2. Classification: Recognizing ASL alphabets with normalized points of the hand
  3. Synthesis: Synthesizing singular alphabets to reduce inaccuracies and form complete sentences

This is what the overall flow looks like:

Inference

Detection

This stage involves the real-time recognition identification and tracking of hand movements and positions using the Google MediaPipe Hand Landmark model. The model captures 21 3D hand key points per frame, and it provides detailed information on hand orientation and finger position. Furthermore, it can run entirely locally and does not require significant computing resources to run in a realtime setting. Below is an image of the 21 points detected by the MediaPipe Hand Landmark model.

Detection

By using these points instead of images of hands (like I previously tried!), the classification process becomes significantly more powerful:

  1. It is not affected by different backgrounds, hand sizes, skin tones and other factors that will make a typical image classification model significantly more incapable
  2. It only needs to process a set of 63 numbers (3 for each point!) for each frame instead of an entire image, making it significantly more efficient for real-time use
  3. It looks a lot cooler

To ensure that the model is not affected by distance from the camera, I normalize each point to be relative to the bounds of the hand itself. When training the model, this provides more standardized data that will help increase accuracy and reliability.

Classification

Once hand landmarks are captured, the data is fed into a Keras PointNet model, which I trained on over 120,000 labelled images of ASL fingerspelling. PointNet is a deep learning model architecture developed with the intent of classifying 3D point clouds, similar to how the detected hands are now represented.

The PointNet model classifies the input data into one of the ASL alphabet signs (except for J and Z, which include movement of the hand to properly express). Below is a demonstration of the training process of the PointNet model.

Classification

I started out by downloading multiple ASL Fingerspell datasets from Kaggle: One Two Three

I used MediaPipe to identify hand landmarks on every image in the three datasets. Subsequently, I normalized the landmarks and stored the point clouds in NumPy files, creating a combined augmented dataset. Below is a visualized form of the point cloud dataset image

Once I had the augmented dataset, I trained a PointNet model on it. So far, I've trained the model around 6 times, playing with parameters, mostly adjusting epoch count, bath size, and learning rate. Here is training information from the most accurate model currently trained:

image image

As seen in the confusion matrix, similar fingerspelled alphabets like K-V, U-R and E-S are often misclassified for each other. This is expected and can be improved with additional training data. There also seems to be an issue in the validation data, as seen in a sudden spike in validation loss in one of the epochs.

Synthesis

The final stage synthesizes the classified characters into coherent words and sentences. This involves error correction to adjust for common misrecognition, as well as contextual synthesis to form sentences based on the classified letters.

First, the program uses conditionals to differentiate between commonly misrecognized letters. For instance, the letters A-T-M-N-S are commonly mistaken for each other, and this can be fixed by checking the relative positioning of certain key coordinates (like the thumb). After ensuring the classified letter is accurate, the program ensures that the recognized letter has been demonstrated for multiple consecutive frames, to ensure that the alphabet is properly recognized. The program also ensures that the same letter is not recognized more than two times consecutively, preventing each individual frame from adding a letter to the synthesized text. Finally, it uses an LLM to synthesize the cleaned information into a meaningful sentence. This synthesis process also applies grammatical rules to form sentences that are syntactically correct in English.

Expressive

The expressive component of this project focuses on translating spoken English into ASL, which is visually represented through a 2D animated avatar using ThreeJS.

To begin, I created a database of over 9000+ words and their corresponding ASL signs. Once again, I used MediaPipe's Pose and Hand Landmark Models to identify body points for each frame in videos of these signs. Furthermore, I also used the all-MiniLM-L6-v2 model to create embeddings for each word. I stored all words, their corresponding embeddings, and their corresponding ASL sign point animations in a PostgreSQL database using pgvector. Although the database only has about 9500 words, using cosine similarity for semantic search allows me to drastically increase the word count by using contextually similar signs to replace certain words. Here is a demonstration of how the expressive aspect of the project works:

The interface begins by using react-speech-recognition to transcribe spoken text into words (this can be replaced with OpenAI Whisper for more accuracy). When the speaker stops speaking for a certain amount of time, it transmits the transcription to the backend through websockets. The backend iterates through the words and creates embeddings for each of them. For each word, it queries the database using the cosine similarity function to fetch the animation of the sign with the closest meaning to the spoken word. If there are no similar signs in the database, it generates a fingerspelling animation using the individual letters in the word. After fetching all the animation points, it transmits them back to the client through websockets. The client then uses ThreeJS to accurately animate the points to control the 2D avatar that signs each word correspondingly.

Future Work

Limitations

  • As seen above, out of the four fundamental obstacles that are present in ASL ⭤ English translation, this project only solves three to a sufficient degree. ASL signers still need to fingerspell words to actually communicate information. Fingerspelling only accounts for 25-30% of signed ASL
  • The receptive model is still not perfect, and can definitely be improved with more training data and ML expertise
  • The expressive part of the project still doesn't convey all five parameters of ASL, and is simply a stick figure that might be hard to comprehend

Vision

  • Use RNNs/LSTMs might be a potential way to interpret ASL signs (instead of just fingerspelling)
  • Look into using ASL parameters and databases like ASL-LEX or SLPAA to narrow down a human signing into specific parameters (like location, orientation, erc) and then use a model to classify the sign
  • Rig and control a 3D model that is used to express ASL, instead of a simple 2D stick figure

My goal for this project has changed as I have progressed, and my vision for it has grown over time. When I initially started working on simply recognizing individual ASL alphabets, I did not expect to get very far, let alone develop something capable of sustaining two-way communication between ASL and English. Regardless, I had one main goal for the project:

  • Once this project is sufficiently capable, I want to set up a desk/TV somewhere between the BCSD and Burnaby South hallways, or maybe even at the main entrance to our school. With the interface running on this device, I hope that students from BCSD and Burnaby South stop by to talk to each other without a human interpreter assisting the conversation. While this project is far from properly capturing all expressive aspects of ASL, I hope that the novelty of the fingerspell recognition and avatar visualization will bring some students together.

Now, since the project is more capable, my vision for it has also grown:

  • I want to run an instance that is publicly accessible, so anyone in the world can just open a website and mess around with it. Ideally, it will draw more attention to ASL translation itself and people take the technology further. I also hope it encourages more people to look into ASL and consider learning the language. Maybe the platform will provide people with a fun and interactive method to learn the language.
  • I also want to modularize the code and make both the receptive and expressive aspects of the project function by themselves. While this project is mostly a proof of concept, I see many potential tools arising from this concept of ASL translation, and I want to work on developing some of them. Hopefully, it will also encourage more people to develop viable accessibility technology. If you happen to be interested in developing such technology, here are some of the potential projects that I have thought of during the development process:
    • ASL Interpreter Webcam Client: A lightweight modular application that adds an ASL interpreter avatar to the top right of a user's webcam. It will sign all input from the microphone. This will allow people to join meetings or make YouTube videos with live English → ASL translation at no cost. This is already possible for more tech-savvy individuals, i.e. I implement it by adding an OBS Browser Source and serving it through a Virtual Camera.
    • ASL YouTube Captions: A browser extension that adds an ASL option for YouTube captions, essentially adding a 2d avatar to the top right that simply signs everything that is said in the YouTube video. This will allow Deaf individuals to better experience YouTube videos if they prefer signs over captions.

Acknowledgements

People

  • My Computer Science Teacher, who is always excited about my projects and inspires me to do more, for sharing this project and doing their very best to find more opportunities for me to grow this project
  • My ASL Teacher, for always taking the time out of their day to provide me with feedback, insight, and guidance for improving this project. And also for finding amazing people and opportunities where I can get more feedback!
  • Ishir Rao, my (goated) childhood friend, for getting me interested in the AI/ML space and providing me with a starting point to get ahead!

Projects

  • These Kaggle datasets, where all my training data was augmented from: One, Two, Three
  • The HandSpeak Project, for a lot of additional knowledge about the workings of ASL, as well as a source of sign language videos for the Expressive aspect of this project
  • The SignWave Project, an amazing tool developed by high-schoolers that provided me with my initial approach towards the Expressive aspect of this project