Source code
Revision control
Copy as Markdown
Other Tools
#pragma once
#include <cstdint>
namespace pk {
// Per-emitted-token decode metadata produced by the greedy decoders when an
// optional out-vector is requested. Mirrors NeMo's per-token timestamp +
// confidence (timestamps=True, confidence method 'max_prob').
//
// id : the emitted token id (always < blank_id; same as the id-only path).
// frame : the encoder time frame attributed to the token (see per-head note).
// * CTC -> the frame where the token's consecutive RUN STARTS
// (NeMo `start_offset` for the collapsed token).
// * RNNT -> the encoder frame `t` at the emission step.
// * TDT -> the encoder frame `t` at the emission step.
// Time(seconds) = frame * frame_sec (= hop * subsampling / sample_rate).
// conf : NeMo's rescaled `max_prob` confidence in (0, 1]:
// conf = (N * p_max - 1) / (N - 1) (alpha == 1.0)
// where p_max is the softmax probability of the emitted (argmax)
// token over the SAME logit slice NeMo log-softmaxes, and N is the
// number of classes in that slice (= vocab_size + 1, incl. blank).
// span : the number of encoder frames this token spans, for word-end timing.
// * CTC -> 1
// * RNNT -> 1
// * TDT -> durations[d_k] (the predicted duration/skip applied
// for this token).
struct TokenInfo {
int32_t id;
int32_t frame;
float conf;
int32_t span;
};
} // namespace pk