Upload national_speech_corpus_sg_imda.py with huggingface_hub

national_speech_corpus_sg_imda.py

@@ -0,0 +1,301 @@
+import os
+import zipfile
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+try:
+    import audiosegment
+except:
+    print("Please install audiosegment to use the `national_speech_corpus_sg_imda` dataloader.")
+import datasets
+import pandas as pd
+
+try:
+    import textgrid
+except:
+    print("Please install textgrid to use the `national_speech_corpus_sg_imda` dataloader.")
+
+from seacrowd.utils import schemas
+from seacrowd.utils.configs import SEACrowdConfig
+from seacrowd.utils.constants import Licenses, Tasks
+
+_CITATION = """\
+@inproceedings{koh19_interspeech,
+  author={Jia Xin Koh and Aqilah Mislan and Kevin Khoo and Brian Ang and Wilson Ang and Charmaine Ng and Ying-Ying Tan},
+  title={{Building the Singapore English National Speech Corpus}},
+  year=2019,
+  booktitle={Proc. Interspeech 2019},
+  pages={321--325},
+  doi={10.21437/Interspeech.2019-1525},
+  issn={2308-457X}
+}
+"""
+
+_DATASETNAME = "national_speech_corpus_sg_imda"
+
+_DESCRIPTION = """\
+The National Speech Corpus (NSC) is the first large-scale Singapore English corpus spearheaded by the Info-communications and Media Development Authority (IMDA) of Singapore.
+It aims to become an important source of open speech data for automatic speech recognition (ASR) research and speech-related applications.
+The NSC improves speech engines’ accuracy of recognition and transcription for locally accented English.
+The NSC is also able to contribute to speech synthesis technology where an AI voice can be produced that is more familiar to Singaporeans, with local terms pronounced more accurately.
+"""
+
+_HOMEPAGE = "https://www.imda.gov.sg/how-we-can-help/national-speech-corpus"
+
+_LANGUAGES = ["eng"]  # We follow ISO639-3 language code (https://iso639-3.sil.org/code_tables/639/data)
+
+_LICENSE = f"{Licenses.OTHERS.value} | Singapore Open Data Licence V1.0"
+
+_LOCAL = True
+
+_URLS = {}
+
+# paths of all file locations, presented in a list to support different operating systems
+_PATHS = {
+    "read_balanced": {
+        "metadata": ["PART1", "DOC", "Speaker Information (Part 1).XLSX"],
+        "data": {
+            "standing_mic": {
+                "audio": ["PART1", "DATA", "CHANNEL0", "WAVE"],
+                "text": ["PART1", "DATA", "CHANNEL0", "SCRIPT"],
+            },
+            "boundary_mic": {
+                "audio": ["PART1", "DATA", "CHANNEL1", "WAVE"],
+                "text": ["PART1", "DATA", "CHANNEL1", "SCRIPT"],
+            },
+            "phone": {
+                "audio": ["PART1", "DATA", "CHANNEL2", "WAVE"],
+                "text": ["PART1", "DATA", "CHANNEL2", "SCRIPT"],
+            },
+        },
+    },
+    "read_pertinent": {
+        "metadata": ["PART2", "DOC", "Speaker Information (Part 2).XLSX"],
+        "data": {
+            "standing_mic": {
+                "audio": ["PART2", "DATA", "CHANNEL0", "WAVE"],
+                "text": ["PART2", "DATA", "CHANNEL0", "SCRIPT"],
+            },
+            "boundary_mic": {
+                "audio": ["PART2", "DATA", "CHANNEL1", "WAVE"],
+                "text": ["PART2", "DATA", "CHANNEL1", "SCRIPT"],
+            },
+            "phone": {
+                "audio": ["PART2", "DATA", "CHANNEL2", "WAVE"],
+                "text": ["PART2", "DATA", "CHANNEL2", "SCRIPT"],
+            },
+        },
+    },
+    "conversational_f2f": {
+        "metadata": ["PART3", "Documents", "Speakers (Part 3).XLSX"],
+        "data": {
+            "close_mic": {
+                "audio": ["PART3", "Audio Same CloseMic"],
+                "text": ["PART3", "Scripts Same"],
+            },
+            "boundary_mic": {
+                "audio": ["PART3", "Audio Same BoundaryMic"],
+                "text": ["PART3", "Scripts Same"],
+            },
+        },
+    },
+    "conversational_telephone": {
+        "metadata": ["PART3", "Documents", "Speakers (Part 3).XLSX"],
+        "data": {
+            "ivr": {
+                "audio": ["PART3", "Audio Separate IVR"],
+                "text": ["PART3", "Scripts Separate"],
+            },
+            "standing_mic": {
+                "audio": ["PART3", "Audio Separate StandingMic"],
+                "text": ["PART3", "Scripts Separate"],
+            },
+        },
+    },
+}
+
+_SUPPORTED_TASKS = [Tasks.SPEECH_RECOGNITION]
+
+_SOURCE_VERSION = "2.0.8"  # should be 2.08 but HuggingFace does not allow
+
+_SEACROWD_VERSION = "2024.06.20"
+
+
+class NationalSpeechCorpusSgIMDADataset(datasets.GeneratorBasedBuilder):
+    """The National Speech Corpus (NSC), the first large-scale Singapore English corpus spearheaded by the Info-communications and Media Development Authority (IMDA) of Singapore."""
+
+    SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
+    SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
+
+    BUILDER_CONFIGS = [
+        SEACrowdConfig(
+            name="national_speech_corpus_sg_imda_source",
+            version=SOURCE_VERSION,
+            description="national_speech_corpus_sg_imda source schema",
+            schema="source",
+            subset_id="national_speech_corpus_sg_imda",
+        ),
+        SEACrowdConfig(
+            name="national_speech_corpus_sg_imda_seacrowd_sptext",
+            version=SEACROWD_VERSION,
+            description="national_speech_corpus_sg_imda SEACrowd schema",
+            schema="seacrowd_sptext",
+            subset_id="national_speech_corpus_sg_imda",
+        ),
+    ]
+
+    DEFAULT_CONFIG_NAME = "national_speech_corpus_sg_imda_source"
+
+    def _info(self) -> datasets.DatasetInfo:
+        if self.config.schema == "source":
+            features = datasets.Features(
+                {
+                    "id": datasets.Value("string"),
+                    "speaker_id": datasets.Value("string"),
+                    "path": datasets.Value("string"),
+                    "audio": datasets.Audio(sampling_rate=16_000),
+                    "text": datasets.Value("string"),
+                }
+            )
+        elif self.config.schema == "seacrowd_sptext":
+            features = schemas.speech_text_features
+
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=features,
+            homepage=_HOMEPAGE,
+            license=_LICENSE,
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
+        """Returns SplitGenerators."""
+        if self.config.data_dir is None:
+            raise ValueError(f"This is a local dataset. Please download the data from {_HOMEPAGE} and pass the path as the data_dir kwarg to load_dataset.")
+        else:
+            data_dir = self.config.data_dir
+
+        splits = []
+        for split_name, data in _PATHS.items():
+            metadata_path = os.path.join(*data["metadata"])
+            for mic_type, data_path in data["data"].items():
+                splits.append(
+                    datasets.SplitGenerator(
+                        name=f"{split_name}_{mic_type}",
+                        gen_kwargs={"filepath": data_dir, "audio_dir": os.path.join(*data_path["audio"]), "text_dir": os.path.join(*data_path["text"]), "split": split_name, "mic_type": mic_type, "metadata_path": metadata_path},
+                    )
+                )
+        return splits
+
+    def _read_part1_part2_text(self, text_path):
+        text_data = {}
+        with open(text_path, encoding="utf-8-sig") as f:
+            for line in f:
+                comp = line.split("\t")
+                text_id = comp[0].strip()
+                if len(text_id) > 1:
+                    text_data[text_id] = comp[1].strip()
+        return text_data
+
+    def _generate_examples(self, filepath: Path, audio_dir: Path, text_dir: Path, split: str, mic_type: str, metadata_path: str, tmp_cache="~/.cache/nsc") -> Tuple[int, Dict]:
+        """Yields examples as (key, example) tuples."""
+
+        # get speaker info from Excel file
+        all_speaker_info = {}
+        if split.startswith("read"):
+            excel_file = pd.read_excel(open(os.path.join(filepath, metadata_path), "rb"), sheet_name="Speakers", dtype="object")
+            column_name = "SCD/PART" + "1" if split == "read_balanced" else 2
+            for idx, row in excel_file.iterrows():
+                all_speaker_info[str(row[column_name])] = {"gender": row["SEX"]}
+        elif split == "conversational_f2f":
+            excel_file = pd.read_excel(open(os.path.join(filepath, metadata_path), "rb"), sheet_name="Same Room", dtype="object")
+            for idx, row in excel_file.iterrows():
+                all_speaker_info[row["SCD"]] = {"age": row["AGE"], "gender": row["SEX"]}
+        elif split == "conversational_telephone":
+            excel_file = pd.read_excel(open(os.path.join(filepath, metadata_path), "rb"), sheet_name="Separate Room", dtype="object")
+            for idx, row in excel_file.iterrows():
+                all_speaker_info[f"{row['Conference ID']}_{row['Speaker ID']}"] = {"age": row["Age"], "gender": row["Gender"]}
+
+        for rel_text_path in os.listdir(os.path.join(filepath, text_dir)):
+            text_path = os.path.join(filepath, text_dir, rel_text_path)
+            text_name, _ext = os.path.splitext(rel_text_path)
+            if not os.path.isfile(text_path):
+                continue
+            if split.startswith("conversational"):
+                if split == "conversational_telephone":
+                    # conf_{confid}_{confid}_{speaker_id} -> {confid}_{speaker_id}
+                    speaker_id = text_name.split("_", 2)[-1]
+                    pass
+                else:
+                    speaker_id = text_name
+                speaker_info = all_speaker_info.get(speaker_id, {})
+                speaker_info["id"] = speaker_id
+
+                if mic_type in ["close_mic", "standing_mic"]:
+                    audio_filename = text_name
+                    audio_filepath = audio_filename + ".wav"
+                    audio_path = os.path.join(audio_dir, audio_filepath)
+                elif mic_type == "boundary_mic":
+                    audio_filename = text_name.split("-")[0]
+                    audio_filepath = audio_filename + ".wav"
+                    audio_path = os.path.join(audio_dir, audio_filepath)
+                elif mic_type == "ivr":
+                    audio_subdir, audio_filename = text_name.rsplit("_", 1)
+                    audio_filepath = audio_filename + ".wav"
+                    audio_path = os.path.join(audio_dir, audio_subdir, audio_filepath)
+
+                audio_file = audiosegment.from_file(os.path.join(filepath, audio_path)).resample(sample_rate_Hz=16000)
+                text_spans = textgrid.TextGrid.fromFile(text_path)
+                audio_name, _ = os.path.splitext(os.path.basename(audio_dir))
+                for text_span in text_spans[0]:
+                    start, end, text = text_span.minTime, text_span.maxTime, text_span.mark
+                    key = f"{audio_name}_{start}_{end}"
+
+                    start_sec, end_sec = int(start * 1000), int(end * 1000)
+                    segment = audio_file[start_sec:end_sec]
+                    export_dir = os.path.join(tmp_cache, "segmented", audio_name)
+                    os.makedirs(export_dir, exist_ok=True)
+                    segement_filename = os.path.join(export_dir, f"{audio_filename}-{round(start, 0)}-{round(end, 0)}.wav")
+                    segment.export(segement_filename, format="wav")
+
+                    example = {
+                        "id": key,
+                        "speaker_id": speaker_info["id"],
+                        "path": segement_filename,
+                        "audio": segement_filename,
+                        "text": text,
+                    }
+                    if self.config.schema == "seacrowd_sptext":
+                        example["metadata"] = {
+                            "speaker_gender": speaker_info.get("gender", None),  # not all speaker details are available
+                            "speaker_age": None,  # speaker age only available in age groups, but SEACrowd schema requires int64
+                        }
+                    yield key, example
+
+            else:
+                text_data = self._read_part1_part2_text(text_path)
+                audio_name, session = text_name[1:-1], text_name[-1]
+                speaker_id = audio_name
+                speaker_info = all_speaker_info.get(speaker_id, {})
+                speaker_info["id"] = speaker_id
+                for text_id, text in text_data.items():
+                    with zipfile.ZipFile(os.path.join(filepath, audio_dir, f"SPEAKER{audio_name}.zip")) as zip_file:
+                        zip_path = os.path.join(f"SPEAKER{audio_name}", f"SESSION{session}", f"{text_id}.WAV")
+                        extract_path = os.path.join(tmp_cache, audio_dir)
+                        os.makedirs(extract_path, exist_ok=True)
+                        audio_path = zip_file.extract(zip_path, path=extract_path)
+
+                        key = text_id
+                        example = {
+                            "id": key,
+                            "speaker_id": speaker_info["id"],
+                            "path": audio_path,
+                            "audio": audio_path,
+                            "text": text,
+                        }
+                        if self.config.schema == "seacrowd_sptext":
+                            example["metadata"] = {
+                                "speaker_gender": speaker_info.get("gender", None),
+                                "speaker_age": None,  # speaker age only available in age groups, but SEACrowd schema requires int64
+                            }
+                        yield key, example