RNN-T
RNN-T is an end-to-end model architecture for streaming speech recognition. It combines a recurrent neural network with a transducer structure and supports online decoding.
Timeline (last 90 days)
Traditional streaming speech recognition solutions typically rely on structures such as CTC (Connectionist Temporal Classification) or RNN-T, decoding progressively by frame or phoneme block, but the decoding unit still must wait until the corresponding phoneme is essentially complete before it is output.
On-device speech recognition typically relies on lightweight models that have been quantized and pruned, such as end-to-end models based on CTC or RNN-T architectures.
RNN-T was proposed by Graves in 2012 and adds a prediction network on top of CTC to capture internal dependencies within the output sequence.
CTC is an alignment-free discriminative model whose decoding can be fully parallelized with extremely low latency; RNN-T introduces an autoregressive mechanism requiring sequential decoding, making its computational complexity significantly higher than CTC.
The researchers selected two independently trained Conformer-Large recognizers as their study subjects, using CTC and RNN-T decoders respectively.
All Facts (5)
Traditional streaming speech recognition solutions typically rely on structures such as CTC (Connectionist Temporal Classification) or RNN-T, decoding progressively by frame or phoneme block, but the decoding unit still must wait until the corresponding phoneme is essentially complete before it is output.
50%UnverifiedOn-device speech recognition typically relies on lightweight models that have been quantized and pruned, such as end-to-end models based on CTC or RNN-T architectures.
50%UnverifiedRNN-T was proposed by Graves in 2012 and adds a prediction network on top of CTC to capture internal dependencies within the output sequence.
50%UnverifiedCTC is an alignment-free discriminative model whose decoding can be fully parallelized with extremely low latency; RNN-T introduces an autoregressive mechanism requiring sequential decoding, making its computational complexity significantly higher than CTC.
50%UnverifiedThe researchers selected two independently trained Conformer-Large recognizers as their study subjects, using CTC and RNN-T decoders respectively.
50%