Create distil_raw_ncc_speech_v7_compact1.py

distil_raw_ncc_speech_v7_compact1.py

@@ -0,0 +1,197 @@
+import argparse
+import io
+import json
+import tarfile
+import datasets
+from huggingface_hub import hf_hub_download
+
+_CITATION = """\
+# Citation details
+"""
+
+_DESCRIPTION = """\
+This database was created from NB deposit recordings
+"""
+
+_HOMEPAGE = "https://ai.nb.no"
+
+# The repository ID for the dataset
+_REPO_ID = "NbAiLab/distil_raw_ncc_speech_v7_compact1"
+
+_SHARDS = {
+    "no": {
+        datasets.Split.TRAIN: 1,  # Updated from 256 to 8
+        datasets.Split.VALIDATION: 1,
+        datasets.Split.TEST: 1,
+    },
+}
+
+_SOURCES = ["audio_books_nn", "audio_books_no", "clean_audio_books_no", 
+            "clean_stortinget_no", "norwegian_fleurs", "nrk_no", "nst", 
+            "stortinget_no"]
+
+_SHARDS["no"].update({f"validation_{source}": 1 for source in _SOURCES})
+_SHARDS["no"].update({f"test_{source}": 1 for source in _SOURCES})
+
+
+class distil_raw_ncc_speech_v7_compact1Config(datasets.BuilderConfig):
+    def __init__(self, *args, **kwargs):
+        super(distil_raw_ncc_speech_v7_compact1Config, self).__init__(*args, **kwargs)
+
+
+class distil_raw_ncc_speech_v7_compact1(datasets.GeneratorBasedBuilder):
+    DEFAULT_WRITER_BATCH_SIZE = 1000
+    BUILDER_CONFIGS = [
+        distil_raw_ncc_speech_v7_compact1Config(
+            name="no",
+            version=datasets.Version("1.0.1"),
+            description="ncc_speech Norwegian Compact 8",
+        ),
+    ]
+
+    def __init__(self, *args, post_processors=None, **kwargs):
+        if not isinstance(post_processors, (tuple, list)):
+            post_processors = [post_processors]
+        self.post_processors = post_processors
+        super().__init__(*args, **kwargs)
+
+    def _info(self):
+        sampling_rate = 16000
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features({
+                "id": datasets.Value("string"),
+                # ... (your existing features)
+            }),
+            supervised_keys=None,
+            homepage=_HOMEPAGE,
+            citation=_CITATION,
+        )
+
+    def _info(self):
+        sampling_rate = 16000
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features({
+                "id": datasets.Value("string"),
+                "group_id": datasets.Value("string"),
+                "source": datasets.Value("string"),
+                "audio_language": datasets.Value("string"),
+                "audio": datasets.features.Audio(sampling_rate=sampling_rate),
+                "audio_duration": datasets.Value("int32"),
+                "previous_text": datasets.Value("string"),
+                "text_en": datasets.Value("string"),
+                "text_language": datasets.Value("string"),
+                "text": datasets.Value("string"),
+                "timestamped_text_en": datasets.Value("string"),
+                "text_en": datasets.Value("string"),
+                "timestamped_text": datasets.Value("string"),
+                "wav2vec_wer": datasets.Value("float32"),
+                "whisper_wer": datasets.Value("float32"),
+                "verbosity_level": datasets.Value("int32"),
+                "file": datasets.Value("string"),
+                "channels": datasets.Value("int32"),
+                "frequency": datasets.Value("int32"),
+                "language": datasets.Value("string"),
+                "task": datasets.Value("string"),
+                "_post_processor": datasets.Value("string"),
+            }),
+            supervised_keys=None,
+            homepage=_HOMEPAGE,
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager):
+        data_urls = {}
+        splits = _SHARDS[self.config.name].keys()
+        for split in splits:
+            data_urls[split] = []
+            shard_total = _SHARDS["no"][split]
+            for shard_idx in range(1, shard_total + 1):
+                string_formatting = dict(
+                    split=split,
+                    lang_code="no",
+                    shard_idx=shard_idx,
+                    shard_total=shard_total
+                )
+                folder = f"data/{split}"
+                metadata_filename = f"ncc_speech_v7-{string_formatting['lang_code']}-{string_formatting['shard_idx']:04d}-{string_formatting['shard_total']:04d}.json"
+                archive_filename = f"ncc_speech_v7-{string_formatting['lang_code']}-{string_formatting['shard_idx']:04d}-{string_formatting['shard_total']:04d}.tar.gz"
+                metadata_path = f"{folder}/{metadata_filename}"
+                archive_path = f"{folder}/{archive_filename}"
+                data_urls[split].append((metadata_path, archive_path))
+
+        return [
+            datasets.SplitGenerator(
+                name=split, gen_kwargs={
+                    "filepaths": data_urls[split],
+                    "dl_manager": dl_manager
+                }
+            ) for split in splits
+        ]
+
+    def _generate_examples(self, filepaths, dl_manager):
+        data_fields = list(self._info().features.keys())
+        data_fields.remove("audio")
+
+        for metadata_path, archive_path in filepaths:
+            metadata_content = {}
+
+            # Download metadata using the Hugging Face Hub API
+            metadata_file_path = hf_hub_download(
+                repo_id=_REPO_ID,
+                filename=metadata_path,
+                repo_type="dataset",
+                use_auth_token=True
+            )
+
+            with open(metadata_file_path, "r", encoding="utf-8") as metadata_file:
+                metadata_records = metadata_file.read().splitlines()
+                for record in metadata_records:
+                    record = record.strip()
+                    if not record:
+                        continue  # Skip empty lines
+                    metadata_object = json.loads(record)
+                    metadata_key = metadata_object["id"]
+                    metadata_content[metadata_key] = metadata_object
+
+            # Stream audio archives using iter_archive
+            archive_file_path = hf_hub_download(
+                repo_id=_REPO_ID,
+                filename=archive_path,
+                repo_type="dataset",
+                use_auth_token=True
+            )
+
+            for filename, file_handle in dl_manager.iter_archive(archive_file_path):
+                if filename.endswith(".mp3"):
+                    metadata_key = filename.replace(".mp3", "")
+                    if metadata_key not in metadata_content:
+                        continue
+
+                    # Extract relevant fields for each entry
+                    fields = {key: metadata_content[metadata_key].get(key, "") for key in data_fields}
+                    fields["file"] = fields["id"] + ".mp3"
+                    fields["channels"] = 1
+                    fields["frequency"] = 16000
+                    fields["task"] = "transcribe"
+                    fields["language"] = fields["text_language"]
+                    fields["_post_processor"] = None
+
+                    audio_bytes = file_handle.read()
+                    audio_dict = {"bytes": audio_bytes, "path": filename}
+                    metadata_dict = {
+                        "id": metadata_key,
+                        "audio": audio_dict,
+                        **fields
+                    }
+
+                    for func in self.post_processors:
+                        if func is None:
+                            yield metadata_key, metadata_dict
+                        else:
+                            func_name = func.__name__ if func.__name__ else hex(id(func)).replace("0x", "lambda-")
+                            result = func(metadata_dict)
+                            if result:
+                                result["_post_processor"] = func_name
+                                yield f"{metadata_key}_{func_name}", result