Expand description
The acoustic model’s output, seen as an FST.
A neural acoustic model hands back a T × V matrix: for each of T frames,
a score for each of V symbols. Decoding is the composition of that matrix
with a decoding graph, so the matrix has to be an FST first.
As an FST it is a chain of T + 1 states, with V arcs from frame t to
frame t + 1, one per symbol, weighted by that symbol’s score in that
frame. It is an acceptor: the symbol is both input and output label.
Nothing is materialised. DenseFst borrows the matrix and computes each
arc as it is asked for, so composing against it costs no copy of the
acoustic scores. It corresponds to a single, unbatched item in k2’s
DenseFsaVec.
Structs§
- Dense
ArcIter - The arcs of one frame: one per column, in column order.
- Dense
Fst - A
T × Vmatrix of acoustic scores, read as an FST. - Dense
State Iter - The states of a
DenseFst, which are the frame boundaries.
Traits§
- From
Score - A weight that can be built from one acoustic score.