mini_librispeech.jl

# SPDX-License-Identifier: CECILL-2.1

#######################################################################

const MINILS_URL = Dict(
    "dev" => "https://www.openslr.org/resources/31/dev-clean-2.tar.gz",
    "train" => "https://www.openslr.org/resources/31/train-clean-5.tar.gz"
)
const MINILS_SUBSETS = Dict(
    "train" => "train-clean-5",
    "dev" => "dev-clean-2"
)

#######################################################################

struct MINILIBRISPEECH <: SpeechCorpus
    recordings
    train
    dev
    test
end

function minils_recordings(dir, subset)
    subsetdir = joinpath(dir, "LibriSpeech", MINILS_SUBSETS[subset])
    recs = Dict()

    for d1 in readdir(subsetdir; join = true)
        for d2 in readdir(d1; join = true)
            for path in readdir(d2; join = true)
                endswith(path, ".flac") || continue
                id = replace(basename(path), ".flac" =>  "")
                r = Recording(
                    id,
                    CmdAudioSource(`sox $path -t wav -`);
                    channels = [1],
                    samplerate = 16000
                )
                recs[r.id] = r
            end
        end
    end
    recs
end

function minils_supervisions(dir, subset)
    subsetdir = joinpath(dir, "LibriSpeech", MINILS_SUBSETS[subset])
    sups = Dict()
    for d1 in readdir(subsetdir; join = true)
        for d2 in readdir(d1; join = true)
            k1 = d1 |> basename
            k2 = d2 |> basename
            open(joinpath(d2, "$(k1)-$(k2).trans.txt"), "r") do f
                for line in eachline(f)
                    tokens = split(line)
                    s = Supervision(
                        tokens[1], # supervision id
                        tokens[1]; # recording id
                        channels = [1],
                        data = Dict("text" => join(tokens[2:end], " "))
                    )
                    sups[s.id] = s
                end
            end
        end
    end

    sups
end

function minils_download(dir)
    donefile = joinpath(dir, ".download.done")
    if ! isfile(donefile)
        run(`mkdir -p $dir`)
        @debug "downloading the corpus"
        for subset in ["train", "dev"]
            run(`wget --no-check-certificate -P $dir $(MINILS_URL[subset])`)
            tarpath = joinpath(dir, "$(MINILS_SUBSETS[subset]).tar.gz")
            @debug "extracting"
            run(`tar -xf $tarpath -C $dir`)
            run(`rm $tarpath`)
        end

        run(pipeline(`date`, stdout = donefile))
    end
    @debug "dataset in $dir"
end

function minils_prepare(dir)
    # 1. Recording manifest.
    out = joinpath(dir, "recordings.jsonl")
    if ! isfile(out)
        open(out, "w") do f
            for subset in ["train", "dev"]
                @debug "preparing recording manifest ($subset) $out"
                recs = minils_recordings(dir, subset)
                writemanifest(f, recs)
            end
        end
    end

    # 2. Supervision manifests.
    for (subset, name) in [("train", "train"), ("dev", "dev"), ("dev", "test")]
        out = joinpath(dir, "supervisions-$name.jsonl")
        if ! isfile(out)
            @debug "preparing supervision manifest ($subset) $out"
            sups = minils_supervisions(dir, subset)
            open(out, "w") do f
                writemanifest(f, sups)
            end
        end
    end
end


function MINILIBRISPEECH(dir, subset)
    minils_download(dir)
    minils_prepare(dir)
    dataset(dir, subset)
end