Fix percent-encoded audio file URIs

src/mistral_common/tokens/tokenizers/audio.py

@@ -8,6 +8,7 @@
 from enum import Enum
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
+from urllib.parse import unquote
 
 import numpy as np
 import requests as _requests_lib
@@ -153,8 +154,8 @@ def from_file(file: str, strict: bool = True) -> "Audio":
         r"""Create an Audio instance from an audio file.
 
         Args:
-            file: Path to the audio file. A file:// URI prefix is accepted
-                and stripped.
+            file: Path to the audio file. A file:// URI prefix is accepted and
+                stripped, and percent-encoded path components are decoded.
             strict: If `True`, strictly enforce mono audio; multi-channel audio
                 raises an error. If `False`, extra channels are dropped.
 
@@ -167,7 +168,7 @@ def from_file(file: str, strict: bool = True) -> "Audio":
         assert_soundfile_installed()
 
         if isinstance(file, str) and file.startswith("file://"):
-            file = file[7:]
+            file = unquote(string=file[7:])
 
         if not Path(file).exists():
             raise FileNotFoundError(f"{file=} does not exist")

tests/unit/tokens/tokenizers/test_audio_file_uri.py

@@ -0,0 +1,84 @@
+from pathlib import Path
+
+import numpy as np
+import pytest
+import soundfile as sf
+
+from mistral_common.protocol.instruct.chunk import AudioURLChunk
+from mistral_common.tokens.tokenizers.audio import (
+    Audio,
+    AudioConfig,
+    AudioEncoder,
+    AudioSpectrogramConfig,
+    SpecialAudioIDs,
+)
+
+_AUDIO_FILENAMES = [
+    "clip one.wav",
+    "clip%20one.wav",
+    "café.wav",
+    "clip+one.wav",
+    "clip.wav",
+]
+
+
+def _write_audio_file(path: Path) -> None:
+    """Write a finite mono WAV fixture to a path."""
+    samples = np.linspace(start=-0.5, stop=0.5, num=800, dtype=np.float32)
+    sf.write(file=path, data=samples, samplerate=16_000)
+
+
+def _audio_encoder() -> AudioEncoder:
+    """Build a local audio encoder without model or network dependencies."""
+    return AudioEncoder(
+        audio_config=AudioConfig(
+            sampling_rate=16_000,
+            frame_rate=100.0,
+            encoding_config=AudioSpectrogramConfig(
+                num_mel_bins=80,
+                hop_length=160,
+                window_size=400,
+            ),
+        ),
+        special_ids=SpecialAudioIDs(
+            audio=1,
+            begin_audio=2,
+            streaming_pad=None,
+            text_to_audio=None,
+            audio_to_text=None,
+        ),
+    )
+
+
+@pytest.mark.parametrize(argnames="filename", argvalues=_AUDIO_FILENAMES)
+def test_audio_from_file_uri_matches_plain_path(tmp_path: Path, filename: str) -> None:
+    """Decode a file URI once while preserving plain local paths."""
+    path = tmp_path / filename
+    _write_audio_file(path=path)
+
+    plain_audio = Audio.from_file(file=str(path))
+    uri_audio = Audio.from_file(file=path.as_uri())
+
+    assert uri_audio.sampling_rate == plain_audio.sampling_rate
+    np.testing.assert_allclose(actual=uri_audio.audio_array, desired=plain_audio.audio_array)
+
+
+def test_audio_url_chunk_file_uri_loads_through_audio_encoder(tmp_path: Path) -> None:
+    """Load a percent-encoded file URI through the AudioURLChunk workflow."""
+    path = tmp_path / "clip one% café.wav"
+    _write_audio_file(path=path)
+
+    plain_audio = Audio.from_file(file=str(path))
+    encoded_audio = _audio_encoder()(content=AudioURLChunk(audio_url=path.as_uri())).audio
+
+    assert encoded_audio is not None
+    assert encoded_audio.sampling_rate == plain_audio.sampling_rate
+    np.testing.assert_allclose(actual=encoded_audio.audio_array, desired=plain_audio.audio_array)
+
+
+def test_missing_audio_file_uri_raises_file_not_found(tmp_path: Path) -> None:
+    """Raise FileNotFoundError when a decoded file URI has no target."""
+    missing_path = tmp_path / "missing clip.wav"
+
+    with pytest.raises(FileNotFoundError):
+        Audio.from_file(file=missing_path.as_uri())