Fix: accept scalar audio formats in Voxtral transcription requests

src/transformers/models/voxtral/processing_voxtral.py

@@ -329,10 +329,20 @@ def apply_transcription_request(
             ]
         else:
             audio = make_list_of_audio(audio)
+            if format is None:
+                raise ValueError("`format` must be provided when passing audio arrays to VoxtralProcessor.")
+
+            if isinstance(format, str):
+                format = [format] * len(audio)
+
             if len(audio) != len(format):
                 raise ValueError(
                     f"When passed as a list of audio, the length ({len(audio)}) must match the number of format ({len(format)})"
                 )
+
+            if not is_soundfile_available():
+                raise ImportError("Please install `soundfile` to encode audio arrays with VoxtralProcessor.")
+
             audio_buffers = []
             for array, f in zip(audio, format):
                 # Create new BytesIO object and write audio data to it

tests/models/voxtral/test_modeling_voxtral.py

@@ -15,6 +15,8 @@
 
 import unittest
 
+import numpy as np
+
 from transformers import (
     AutoProcessor,
     LlamaConfig,
@@ -27,10 +29,12 @@
 from transformers.testing_utils import (
     Expectations,
     cleanup,
+    require_mistral_common,
     require_torch,
     slow,
     torch_device,
 )
+from transformers.utils import is_soundfile_available
 
 from ...alm_tester import ALMModelTest, ALMModelTester
 
@@ -380,6 +384,20 @@ def test_mini_multi_turn_text_and_audio(self):
         ]
         self.assertEqual(decoded_outputs, EXPECTED_OUTPUT)
 
+    @slow
+    @require_mistral_common
+    @unittest.skipUnless(is_soundfile_available(), "test requires soundfile")
+    def test_transcription_request_accepts_single_format_for_array_audio(self):
+        processor = AutoProcessor.from_pretrained(self.checkpoint_name)
+        audio = np.zeros(16000, dtype=np.float32)
+
+        inputs = processor.apply_transcription_request(
+            audio=audio, model_id=self.checkpoint_name, sampling_rate=16000, format="wav"
+        )
+
+        self.assertIn("input_features", inputs)
+        self.assertEqual(inputs["input_features"].shape[0], 1)
+
     @slow
     def test_transcribe_mode_audio_input(self):
         """