Dataset

Project.toml

@@ -7,6 +7,7 @@ version = "0.7.0"
 Base64 = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f"
 HTTP = "cd3eb016-35fb-5094-929b-558a96fad6f3"
 JSON = "682c06a0-de6a-54ab-a142-c8b1cf79cde6"
+MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 SpeechFeatures = "6f3487c4-5ca2-4050-bfeb-2cf56df92307"
 WAV = "8149f6b0-98f6-5db9-b78f-408fbbb8ef88"
 

src/SpeechCorpora.jl

@@ -7,7 +7,9 @@ using HTTP
 using JSON
 using WAV
 
-using SpeechFeatures: AbstractAudioSource, CmdAudioSource, FileAudioSource, URLAudioSource, loadaudio
+using SpeechFeatures
+
+import MLUtils
 
 export
     # ManifestItem
@@ -28,11 +30,17 @@ export
     # Corpora
     MultilingualLibriSpeech,
     MiniLibriSpeech,
-    TIMIT
+    TIMIT,
+
+    # Dataset
+    dataset
 
 include("speechcorpus.jl")
 include("manifest_item.jl")
 include("manifest_io.jl")
+include("dataset.jl")
+
+# Supported corpora
 include("corpora/multilingual_librispeech.jl")
 include("corpora/mini_librispeech.jl")
 include("corpora/timit.jl")

src/dataset.jl

+# SPDX-License-Identifier: CECILL-2.1
+
+"""
+FastDataset(supervisions, recordings, partition)
+
+Constructor for dataset represented as JSONL files (a.k.a. manifests).
+"""
+struct SpeechDataset <: MLUtils.AbstractDataContainer
+    idxs::Vector{AbstractString}
+    superivions::Dict{AbstractString, Supervision}
+    recordings::Dict{AbstractString, Recording}
+    partition::Symbol
+end
+
+"""
+dataset(manifestroot, subset)
+
+Load `SpeechDataset` from manifest files stored in `manifestroot`.
+Partition is specified by `subset`, e.g. `:train`, `:test`.
+
+Each item of the dataset is a nested tuple `((samples, sampling_rate), Supervision.data)`.
+
+See also [`Supervision`](@ref).
+
+# Examples
+```julia-repl
+julia> ds = dataset("./manifests", :train)
+SpeechDataset(
+    ...
+)
+julia> ds[1]
+(
+    (samples=[...], sampling_rate=16_000),
+    Dict(
+        "text" => "Supervision text here"
+    )
+)
+```
+"""
+function dataset(manifestroot::AbstractString, subset)
+    supervisions = load(Supervision, manifestroot, subset)
+    recordings = load(Recording, manifestroot, subset)
+    idxs = collect(keys(supervisions))
+    SpeechDataset(idxs, supervisions, recordings, Symbol(subset))
+end
+
+function Base.getindex(d::SpeechDataset, key::AbstractString)
+    sup = d.superivions[key]
+    rec = d.recordings[sup.recording_id]
+    samples, sr = load(rec, sup)
+    (samples=samples, sampling_rate=sr), sup.data
+end
+Base.getindex(d::SpeechDataset, idx::Integer) = getindex(d, d.idxs[idx])
+# Fix1 -> partial funcion with fixed 1st argument
+Base.getindex(d::SpeechDataset, idxs::AbstractVector) = map(Base.Fix1(getindex, d), idxs)
+
+Base.length(d::SpeechDataset) = length(d.idxs)