Newer
Older
# SPDX-License-Identifier: CECILL-2.1
module SpeechRecognitionFSTs
using ..TensorFSTs
using ..TensorFSTs.LinearFSTs
export EmissionMapping, lexiconfst, tokenfst
"""
struct EmissionMapping <: AbstractMatrix{Int}
numstates::Int
numtokens::Int
end
Default emission mapping which assumes the same number of emissions per token.
"""
struct EmissionMapping <: AbstractMatrix{Int}
Base.size(x::EmissionMapping) = (x.numpdfs, x.numtokens)
Base.getindex(x::EmissionMapping, i, j) = (j - 1) * x.numpdfs + i
Base.maximum(x::EmissionMapping) = (x.numsnumpdfstates * x.num_emission)
"""
tokenfst(S, topo, initweights, finalweights, tokens, mapping)
Create a FST composed of a set of smaller FSTs. Each token share the same FST
topology specified with `topo = [(src, dest, weight), ...]`. `S` is a
`Semiring` of output FST, `tokens` is a list of token IDs, and
`mapping[tokens[i]]` is the output symbol .
"""
function tokenfst(
S,
topo,
initweights,
finalweights,
tokens,
mapping = EmissionMapping(length(topo), length(tokens))
)
print(mapping[1,2])
emission_count = 1
states = Set(Int[])
arcs, init, final = [], [], []
for (i, token) in enumerate(tokens)
offset = length(states)
values = [topo_arc[2] for topo_arc in topo]
for (j, topo_arc) in enumerate(topo)
if j>1
if topo_arc[2] != values[j-1]
emission_count += 1
end
end
src, dest, weight = offset + topo_arc[1], offset + topo_arc[2], topo_arc[3]
arc = Arc(
src = src,
isym = mapping[emission_count],
dest = dest,
weight = S(weight)
)
push!(states, src)
push!(states, dest)
push!(arcs, arc)
end
for (state, weight) in initweights
state = offset + state
push!(states, state)
push!(init, state => S(weight))
end
for (state, weight) in finalweights
state = offset + state
push!(states, state)
push!(final, state => S(weight))
end
end
TensorFST(arcs, init, final)
end
include("trie.jl")
include("lexicon.jl")