Facebook releases low-latency online speech recognition framework

2020-01-14

Facebook's artificial intelligence (AI) division announced today that its deep learning-based inference framework, Wav2letter@anywhere, can now achieve fast online automatic speech recognition (ASR) in cloud or embedded edge environments. Wav2letter@anywhere consists of the neural network-based language models wav2letter and wav2letter++, which were released in December 2018 and hailed as the most available and fastest open-source speech recognition systems. ASR is the technology that converts a speaker's voice into text and infers the speaker's intentions. The wav2letter++ knowledge base, available on GitHub, supports concurrent audio streams and popular deep learning speech recognition models such as convolutional neural networks (CNN) or recurrent neural networks (RNN), meeting the scalability requirements for online ASR. According to a recent paper published by eight fair research scientists from the New York City lab and Menlo Park headquarters, Wav2letter@anywhere outperforms two baseline models created by bidirectional LSTM RNNs in terms of word error rate. Bidirectional LSTM RNNs are a popular method for controlling latency, but Wav2letter@anywhere has surpassed them. "The system's throughput is nearly three times that of the tuned hybrid ASR baseline, while also achieving lower latency and better word error rate. Although delay-controlled bidirectional LSTMs are commonly used in online speech recognition, combining future context with convolutions can generate more accurate and lower latency models. We found that TDS convolutions can maintain lower expectations in a limited future environment," the researchers stated in the paper. These advancements were achieved through improvements in time-depth separable (TBS) convolutional acoustic models, a method proposed by Facebook at Interspeech 2019 last autumn. This method not only reduces latency but also provides state-of-the-art performance on LibriSpeech. CNNs used for speech inference deviate from the trend of natural language models, which focus on recurrent neural networks or transformer-based models such as Google's transformer and bidirectional encoders. Separable models are most famous for their applications in computer vision, such as Google's MobileNet. The launch of wav2letter@anywhere comes after the release of Pythia's image and language model framework, as well as novel works like wav2vec online speech recognition and RoBERTa. RoBERTa, a model based on Google's BERT, rose to the top of the GLUE benchmark rankings this summer but has since fallen to eighth place.

RECENT AI NEWS

RECENT AI TOOLS