CTC sums alignment paths whose blanks and repeats collapse in a defined order

For each encoder time step, a CTC model defines a distribution over its label alphabet plus blank. Collapse first merges adjacent repetitions of the same path label, then removes blanks. Thus a,a,blank,a,b,b becomes a,a,b, preserving two output a labels separated in the path. Repeated output labels need separating blank positions in a minimal alignment. These labels may be characters or other tokens rather than words, and encoder steps can represent subsampled feature frames rather than individual waveform samples.

The original Graves and colleagues paper defines the path-summing objective and its dynamic program. The target probability includes every compatible alignment, not only the most likely framewise path. A greedy path decoder is convenient, but its collapsed output need not maximize the summed target probability. A prefix beam can retain several output hypotheses and combine equivalent alignment contributions. Keep blank-ending and nonblank-ending contributions distinct where the recurrence requires them; otherwise repeated-label handling can silently change even when the top-level transcript looks plausible on easy examples.

CTC can be used with a causal or bounded-lookahead encoder, but the loss itself does not restrict the encoder's receptive field. An offline bidirectional model may use the entire recording to predict an early frame and cannot become a genuine live recognizer merely by receiving audio in smaller arrays. Specify feature centering, encoder right context, subsampling, and chunk cache state. Emformer is a primary example of explicitly designed streaming context and memory; its latency settings are architectural and execution choices rather than consequences of the letters CTC.

A partial transcript can change when new evidence arrives, even when the acoustic computation is causal. Define whether the application receives unstable hypotheses, a committed prefix, or only a final result. A token spike is not automatically a calibrated word boundary or a definitive end of turn. Across chunks, carry the decoder state needed to distinguish a repeated label from a continuation of the same path run. Independently collapsing each chunk and concatenating the resulting strings can duplicate or delete labels at the boundary.

Use tiny enumerated paths to verify collapse and dynamic-programming logic, then compare chunked recognition with a reference under the same allowed context. Test repeated characters, long blank runs, boundary-spanning label runs, and empty outputs. Record whether scores are summed in probability space or combined with stable log-sum-exp arithmetic. Beam pruning and language-model additions change the search contract and may change scores' interpretation. Evaluate word error, partial revision rate, and emission delay separately; a low final error rate does not establish low live interaction latency.

The probability lab counts label-like successes but discards their order. Reset, explicitly set BOTH p=0.6 and n=3, predict mean count 1.8, then Play or Step. Reset and re-set BOTH p=0.8 and n=3; predict 2.4. Edits retain t and playback highlights exact binomial counts. There is no CTC decoder or blank alphabet. Two three-step paths with two a labels, a,a,blank and a,blank,a, have the same count yet collapse differently. This limitation shows why aggregate label counts cannot replace an ordered alignment calculation.