[transducer] decoding strategy

[Mddct]

Open this to track different decoding strategies of the transducer model. I will update some materials and code snippets here Suggestions welcome and hope to collaborate

[Mddct]

• espnet optimize thanks @b-flo
• Sequence Transduction with Recurrent Neural Networks
• Alignment-Length Synchronous Decoding for RNN Transducer
• Accelerating RNN Transducer Inference via One-Step Constrained Beam Search
• Accelerating RNN Transducer Inference via Adaptive Expansion Search
• LESS IS MORE: IMPROVED RNN-T DECODING USING LIMITED LABEL CONTEXT AND PATH MERGING
• ALIGNMENT-LENGTH SYNCHRONOUS DECODING FOR RNN TRANSDUCER
• https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9250505

[b-flo]

Hi @Mddct ,

For alignment-length synchronous (ALSD), time-synchronous decodng (TSD) and modified Adaptive Expansion Search (mAES) in ESPnet, please refer to this. The version you linked is missing some fixes and optimizations!

Btw, you can drop 1-step/N-step constrained beam search from your list as it was superseded by AES/mAES!

[Mddct]

Hi @Mddct ,

For alignment-length synchronous (ALSD), time-synchronous decodng (TSD) and modified Adaptive Expansion Search (mAES) in ESPnet, please refer to this. The version you linked is missing some fixes and optimizations!

Btw, you can drop 1-step/N-step constrained beam search from your list as it was superseded by AES/mAES!

Thank you very much for your advice， I'll do some experiments later

[Mddct]

If we want to use one-step decoding in the inference stage, can we try the implementation of this rnhnt loss later? paper implement

[github-actions[bot]]

This issue has been automatically closed due to inactivity.