自动语音识别

Automatic Speech Recognition

Convolutional front ends

Encoder-decoder and self-supervised recognition

CTC alignment and error rates

Which of these recognition settings should have the highest word error rate for a system trained on standard read speech?

Whisper's 30-second audio context yields 3,000 log mel frames at one per 10 ms, and its second convolutional layer uses a stride of 2. How many embeddings per context window leave the two-layer front end?

Against a $200$-word reference, a recognizer produces a hypothesis with $14$ substitutions, $6$ insertions and $3$ deletions. What is its word error rate, to three decimal places?

In the first stage of HuBERT pretraining, what serves as the gold target at each masked frame?

A CTC decoder emits the frame-level path [·, a, b, b, ·, b, a] over alphabet {a, b, ·} with · the blank. Apply the collapsing function B, merging consecutive duplicates then deleting blanks. What text does the path decode to?