pl_debates_test.py

# coding=utf-8
# pl_debates_test.py

import os
import csv
import datasets

_CITATION = """
@misc{pl_politicians_2025,
  title = {Polish Politicians Speech Dataset},
  author = {directtt},
  year = {2025}
}
"""

# Base URLs on the Hub — replace YOUR-USER/YOUR-REPO accordingly
_BASE_URL       = "https://huggingface.co/datasets/directtt/pl_debates_test/resolve/main/"
_TRANSCRIPT_URL = _BASE_URL + "transcript/pl/test.tsv"
_AUDIO_TAR_URL  = _BASE_URL + "audio_pl.tar"


class PlDebatesTest(datasets.GeneratorBasedBuilder):
    """Polish politicians’ speech audio dataset."""

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="pl_debates_test",
            version=datasets.Version("1.0.0"),
            description="Studio‐recorded short utterances by Polish politicians.",
        )
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description="Studio‐recorded short utterances by Polish politicians.",
            features=datasets.Features({
                "path":        datasets.Value("string"),
                "audio":       datasets.Audio(sampling_rate=16_000),
                "sentence":    datasets.Value("string"),
                "age":         datasets.Value("int32"),
                "gender":      datasets.Value("string"),
                "speech_type": datasets.Value("string"),
                "source":      datasets.Value("string"),
            }),
            supervised_keys=None,
            homepage="https://huggingface.co/datasets/directtt/pl_debates_test",
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        if dl_manager.is_streaming:
            # streaming: download the tar and the TSV, but don't extract to disk
            tsv_path = dl_manager.download(_TRANSCRIPT_URL)
            tar_path = dl_manager.download(_AUDIO_TAR_URL)
            archives = [dl_manager.iter_archive(tar_path)]
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TEST,
                    gen_kwargs={
                        "tsv_path":   tsv_path,
                        "archives":   archives,
                        "streaming":  True,
                    },
                ),
            ]
        else:
            # eager: download & extract everything
            downloaded = dl_manager.download_and_extract({
                "tsv": _TRANSCRIPT_URL,
                "tar": _AUDIO_TAR_URL,
            })
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TEST,
                    gen_kwargs={
                        "tsv_path":   downloaded["tsv"],
                        "audio_root": downloaded["tar"],
                        "streaming":  False,
                    },
                ),
            ]

    def _generate_examples(self, tsv_path, streaming, archives=None, audio_root=None):
        # 1) load metadata
        meta = {}
        with open(tsv_path, encoding="utf-8") as f:
            reader = csv.DictReader(f, delimiter="\t")
            for row in reader:
                meta[row["path"]] = row

        key = 0

        if streaming:
            # iterate inside the tar archive without extracting
            for archive in archives:
                for path_in_tar, fileobj in archive:
                    # path_in_tar looks like "audio/pl/<speaker>/<fname>.wav"
                    if not path_in_tar.endswith(".wav"):
                        continue
                    # build the rel_path for lookup
                    rel_path = path_in_tar.replace("\\", "/")  # normalize on Windows
                    row = meta.get(rel_path)
                    if row is None:
                        continue
                    # read the bytes and yield them
                    audio_bytes = fileobj.read()
                    yield key, {
                        "path":        rel_path,
                        "audio":       {"path": None, "bytes": audio_bytes},
                        "sentence":    row["sentence"],
                        "age":         int(row.get("age", -1)),
                        "gender":      row["gender"],
                        "speech_type": row["speech_type"],
                        "source":      row["source"],
                    }
                    key += 1

        else:
            # eager: walk the extracted folder on disk
            pl_root = os.path.join(audio_root, "audio", "pl")
            for speaker in sorted(os.listdir(pl_root)):
                sp_dir = os.path.join(pl_root, speaker)
                if not os.path.isdir(sp_dir):
                    continue
                for fname in sorted(os.listdir(sp_dir)):
                    if not fname.endswith(".wav"):
                        continue
                    rel_path = os.path.join("audio", "pl", speaker, fname)
                    row = meta.get(rel_path)
                    if row is None:
                        continue
                    yield key, {
                        "path":        rel_path,
                        "audio":       os.path.join(sp_dir, fname),
                        "sentence":    row["sentence"],
                        "age":         int(row.get("age", -1)),
                        "gender":      row.get("gender", "unknown") or "unknown",
                        "speech_type": row["speech_type"],
                        "source":      row["source"],
                    }
                    key += 1