Wav2Letter++

Provides recipes and pre-trained models for multiple published ASR papers, such as those by Pratap et al. (2020) and Synnaeve et al. (2020), ensuring exact reproducibility with specified Flashlight versions.

Trains models directly from audio to text without intermediate phonetic representations, supporting lexicon-free recognition as highlighted in the key features for modern architectures.

Enables real-time speech recognition with low-latency ConvNet models, specifically in recipes like Pratap et al. (2020) for online ASR applications.

Implements state-of-the-art models like time-depth separable convolutions and ConvNets, based on research papers included in the repository for cutting-edge accuracy.