speechbrain · knQzx · Apr 4, 2026
diff --git a/speechbrain/dataio/dataio.py b/speechbrain/dataio/dataio.py
@@ -99,7 +99,7 @@ def _recursive_format(data, replacements):
             # If not dict, list or str, do nothing
 
 
-def load_data_csv(csv_path, replacements=None):
+def load_data_csv(csv_path, replacements=None, col_types=None):
     """Loads CSV and formats string values.
 
     Uses the SpeechBrain legacy CSV data format, where the CSV must have an
@@ -108,6 +108,11 @@ def load_data_csv(csv_path, replacements=None):
     The rest of the fields are left as they are (legacy _format and _opts fields
     are not used to load the data in any special way).
 
+    When ``col_types`` is provided, the specified columns are explicitly cast
+    to the given types. This is useful when numeric columns (e.g. duration,
+    speaker id) need to be stored as a specific Python type rather than as
+    strings.
+
     Bash-like string replacements with $to_replace are supported.
 
     Arguments
@@ -117,6 +122,10 @@ def load_data_csv(csv_path, replacements=None):
     replacements : dict
         (Optional dict), e.g., {"data_folder": "/home/speechbrain/data"}
         This is used to recursively format all string values in the data.
+    col_types : dict, optional
+        A mapping from column name to a callable type, e.g.
+        ``{"duration": float, "spk_id": int}``.  Only the listed columns
+        are cast; all others are left as strings.
 
     Returns
     -------
@@ -169,6 +178,11 @@ def load_data_csv(csv_path, replacements=None):
             # Duration:
             if "duration" in row:
                 row["duration"] = float(row["duration"])
+            # Explicit column type casting:
+            if col_types:
+                for col, typ in col_types.items():
+                    if col in row:
+                        row[col] = typ(row[col])
             result[data_id] = row
     return result
 

diff --git a/speechbrain/dataio/dataset.py b/speechbrain/dataio/dataset.py
@@ -453,10 +453,31 @@ def from_json(
 
     @classmethod
     def from_csv(
-        cls, csv_path, replacements={}, dynamic_items=[], output_keys=[]
+        cls,
+        csv_path,
+        replacements={},
+        dynamic_items=[],
+        output_keys=[],
+        col_types=None,
     ):
-        """Load a data prep CSV file and create a Dataset based on it."""
-        data = load_data_csv(csv_path, replacements)
+        """Load a data prep CSV file and create a Dataset based on it.
+
+        Parameters
+        ----------
+        csv_path : str
+            Path to the CSV file.
+        replacements : dict
+            Substitutions for $-variables in the CSV values.
+        dynamic_items : list
+            Dynamic items to add to the dataset.
+        output_keys : list
+            Keys to include in the output.
+        col_types : dict, optional
+            Mapping from column name to type, e.g.
+            ``{"duration": float, "spk_id": int}``.
+            Passed through to :func:`load_data_csv`.
+        """
+        data = load_data_csv(csv_path, replacements, col_types=col_types)
         return cls(data, dynamic_items, output_keys)
 
     @classmethod

diff --git a/tests/unittests/test_data_io.py b/tests/unittests/test_data_io.py
@@ -3,6 +3,34 @@
 import torch
 
 
+def test_load_data_csv_col_types(tmpdir):
+    """col_types should cast only the specified columns."""
+    from speechbrain.dataio.dataio import load_data_csv
+
+    csv_content = "ID,duration,spk_id,transcript\nutt1,1.45,007,hello world\nutt2,2.0,42,bye\n"
+    csv_path = os.path.join(tmpdir, "test_col_types.csv")
+    with open(csv_path, "w", encoding="utf-8") as f:
+        f.write(csv_content)
+
+    # without col_types, duration is float (legacy), rest are strings
+    data = load_data_csv(csv_path)
+    assert isinstance(data["utt1"]["duration"], float)
+    assert isinstance(data["utt1"]["spk_id"], str)
+    assert data["utt1"]["spk_id"] == "007"  # leading zero preserved
+
+    # with col_types, specified columns get cast
+    data = load_data_csv(csv_path, col_types={"spk_id": int})
+    assert data["utt1"]["spk_id"] == 7
+    assert isinstance(data["utt1"]["spk_id"], int)
+    assert isinstance(data["utt1"]["duration"], float)  # still float (legacy)
+    assert isinstance(data["utt1"]["transcript"], str)
+
+    # col_types can override duration type too
+    data = load_data_csv(csv_path, col_types={"duration": str})
+    # col_types is applied after the legacy duration cast, so it re-casts
+    assert isinstance(data["utt1"]["duration"], str)
+
+
 def test_read_audio_info(tmpdir, device):
     from speechbrain.dataio.dataio import read_audio_info, write_audio