Google's Live Caption adds real-time subtitles to videos.

2019-10-31

The subtitles of audio content are not only crucial for deaf and hearing-impaired individuals, but also beneficial for everyone. Watching videos on mute is a common practice, whether it's on a train, in a meeting, at night, or when children are sleeping. Research has shown that subtitles can increase the viewing time of users by nearly 40%. However, subtitle support plugins are scattered across various applications and even within applications, making it difficult to access a large amount of audio content.


Recently, the Google AI team introduced Live Caption, a new feature that automatically adds subtitles to the media you play on your phone. The subtitles are added in real-time and completely on the device, without the need for network resources, thus protecting privacy and reducing latency. This feature is currently only available on Google Pixel 4 and Pixel 4 XL, and will be launched on Pixel 3 models later this year, and will soon be widely available on other Android devices.



Building Accurate and Efficient Speech Recognition Models


Real-time captions work by combining three deep learning models on the following devices: a recursive neural network (RNN) sequence transduction model for speech recognition (RNN-T), a text-based recursive neural network model for speech recognition without punctuation, and a convolutional neural network (CNN) model for sound event classification. Real-time captions integrate signals from the three models to create a subtitle track where the process of speech recognition results is not interrupted when sound event labels appear. It automatically predicts punctuation while updating the text in parallel.


For speech recognition, the team utilizes models built on the AudioSet dataset to prepare for sound event detection. The speech recognition model is not only used to generate popular sound effect labels, but also to detect speech segments. The automatic speech recognition (ASR) RNN-T is only active during speech segments to minimize memory and battery usage. For example, when music is detected and there is no speech in the audio stream, the [MUSIC] label appears on the screen and the ASR model is unloaded. The ASR model is only loaded back into memory when speech reappears in the audio stream. To enable Live Caption to run continuously for a long time, the ASR model of Live Caption is optimized for edge devices using various techniques such as neural connection pruning, reducing power consumption by 50% compared to full-size speech models. However, despite being much more energy-efficient, the model still performs well in various scenarios and is robust to background noise.


The text-based punctuation model is optimized to run continuously on devices using a smaller architecture than cloud computing. As subtitles are formed, the speech recognition results are updated several times per second. To save computational resources and provide a smooth user experience, punctuation prediction is performed at the end of the text in the most recently recognized sentence, and if the next updated ASR result does not change the text, the previous punctuation result is retained and reused.



Future Outlook


Live Caption is now available in English on Pixel 4 and will soon be available on Pixel 3 and other Android devices. We hope to bring this feature to more users by expanding support for other languages and further improving the formatting to enhance the perceived accuracy and coherence of the subtitles.

RECENT AI NEWS

RECENT AI TOOLS