CTC
Connectionist Temporal Classification (CTC) is a loss function and decoding method for sequence-to-sequence learning that enables end-to-end training of speech recognition and other sequence labeling models without requiring explicit alignment annotations between input and output sequences.
Timeline (last 90 days)
CTC introduces a blank symbol, requires no prior knowledge of the alignment between phonemes and audio frames, and during decoding produces text by merging, deduplicating and removing blanks along alignment paths.
Pure CTC models often lag behind autoregressive large models in WER, and in engineering terms the two are more complementary than substitutes.
CTC features frame-level independent output and parallelizable computation, making it naturally suited to streaming recognition and low-latency scenarios.
Compared to the autoregressive decoder used by Whisper, CTC has more predictable computational cost and lower latency, which is one of the key reasons it can be deployed on resource-constrained microcontrollers.
CTC (Connectionist Temporal Classification) is a decoding strategy that requires no explicit alignment between input frames and output characters, and can produce output frame by frame during inference without waiting for a complete sentence, making it suitable for low-latency streaming scenarios.
In ASR scenarios, the teacher model is typically an end-to-end speech recognition model based on CTC or an Attention-based Encoder-Decoder.
The researchers selected two independently trained Conformer-Large recognizers as their study subjects, using CTC and RNN-T decoders respectively.
CTC was proposed by Alex Graves in 2006 and assumes that output labels are conditionally independent given the input.
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.
4 more timeline events
All Facts (14)
CTC introduces a blank symbol, requires no prior knowledge of the alignment between phonemes and audio frames, and during decoding produces text by merging, deduplicating and removing blanks along alignment paths.
50%UnverifiedPure CTC models often lag behind autoregressive large models in WER, and in engineering terms the two are more complementary than substitutes.
50%UnverifiedCTC features frame-level independent output and parallelizable computation, making it naturally suited to streaming recognition and low-latency scenarios.
50%UnverifiedCompared to the autoregressive decoder used by Whisper, CTC has more predictable computational cost and lower latency, which is one of the key reasons it can be deployed on resource-constrained microcontrollers.
50%UnverifiedCTC (Connectionist Temporal Classification) is a decoding strategy that requires no explicit alignment between input frames and output characters, and can produce output frame by frame during inference without waiting for a complete sentence, making it suitable for low-latency streaming scenarios.
50%UnverifiedIn ASR scenarios, the teacher model is typically an end-to-end speech recognition model based on CTC or an Attention-based Encoder-Decoder.
50%UnverifiedThe researchers selected two independently trained Conformer-Large recognizers as their study subjects, using CTC and RNN-T decoders respectively.
50%UnverifiedCTC was proposed by Alex Graves in 2006 and assumes that output labels are conditionally independent given the input.
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 introduction of the CTC (Connectionist Temporal Classification) loss function made end-to-end text recognition possible.
50%UnverifiedThe CTC loss function was proposed by Alex Graves in 2006, solving the problem of misalignment between input sequence length and output sequence length.
50%UnverifiedThe three mainstream architectures for end-to-end speech recognition are CTC, attention-based, and RNN-Transducer.
50%UnverifiedDeep Speech was pioneered by Baidu in 2014, being the first to demonstrate that a single neural network could directly map from audio waveforms to text.
50%