Reproduce “Transformer Transducer: One Model Unifying Streaming and Non-streaming Speech Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anshuman Tripathi, Jaeyoung Kim, Qian Zhang, Han Lu, Hasim Sak
Google researchers developed the Transformer Transducer (T-T), a unified model architecture that bridges the gap between streaming and non-streaming Automatic Speech Recognition (ASR). This single model achieves accuracy comparable to systems with unlimited context using just 1-2 seconds of future context, simplifying development and deployment while providing both immediate feedback and high-quality final transcriptions.
Key insights
- A single Transformer Transducer model can effectively operate in both streaming and non-streaming automatic speech recognition modes by varying the amount of right context during training and inference.
- Achieving accuracy comparable to full-utterance models is possible with only 1-2 seconds of future audio context.
- Combining a Y-model architecture with constrained alignment training allows for simultaneous low-latency streaming outputs and higher-accuracy non-streaming refinements from a single system.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.