[Misc] Clarify mono audio requirement (#50141)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>

docs/features/multimodal_inputs.md

@@ -421,6 +421,7 @@ full_transcription = " ".join(transcriptions)
 
 The `split_audio` function:
 
+- Expects 1D mono audio (`load_audio` downmixes by default)
 - Splits audio at quiet points to avoid cutting through speech
 - Uses RMS energy to find low-amplitude regions within the overlap window
 - Preserves all audio samples (no data loss)

tests/multimodal/test_audio.py

@@ -744,6 +744,23 @@ class TestAudioChunking:
         # from start_idx, so the quietest scanned window starts at 19200.
         assert split_idx == 19200
 
+    def test_split_audio_rejects_multi_channel(self):
+        """Chunking is mono-only; stereo must fail loudly rather than silently.
+
+        find_split_point searches axis 0, so (channels, time) input would skip
+        the energy search entirely and split at fixed boundaries instead.
+        """
+        stereo = np.ones((2, 16000 * 65), dtype=np.float32)
+
+        with pytest.raises(ValueError, match="expects mono audio"):
+            split_audio(
+                audio_data=stereo,
+                sample_rate=16000,
+                max_clip_duration_s=30.0,
+                overlap_duration_s=1.0,
+                min_energy_window_size=1600,
+            )
+
     def test_split_audio_preserves_boundaries(self):
         """Verify first and last samples are preserved when chunking."""
 

vllm/entrypoints/speech_to_text/base/serving.py

@@ -178,6 +178,7 @@ class SpeechToTextBaseServing(GenerateBaseServing):
                 y, sr = load_audio(
                     buf,
                     sr=self.asr_config.sample_rate,
+                    mono=True,
                     max_duration_s=self.max_audio_decode_duration_s,
                 )
         except ValueError:

vllm/multimodal/audio.py

@@ -336,8 +336,8 @@ def split_audio(
     for splitting.
 
     Args:
-        audio_data: Audio array to split. Can be 1D (mono) or multi-dimensional.
-                   Splits along the last dimension (time axis).
+        audio_data: 1D mono audio array to split. ASR models consume mono, so
+                   callers must downmix before chunking.
         sample_rate: Sample rate of the audio in Hz.
         max_clip_duration_s: Maximum duration of each chunk in seconds.
         overlap_duration_s: Overlap duration in seconds between consecutive chunks.
@@ -345,8 +345,10 @@ def split_audio(
         min_energy_window_size: Window size in samples for finding low-energy regions.
 
     Returns:
-        List of audio chunks. Each chunk is a numpy array with the same shape
-        as the input except for the last (time) dimension.
+        List of 1D audio chunks.
+
+    Raises:
+        AssertionError: If ``audio_data`` is not 1D.
 
     Example:
         >>> audio = np.random.randn(1040000)  # 65 seconds at 16kHz
@@ -360,6 +362,11 @@ def split_audio(
         >>> len(chunks)
         3
     """
+    if audio_data.ndim > 1:
+        raise ValueError(
+            f"split_audio expects mono audio, got shape {audio_data.shape}"
+        )
+
     chunk_size = int(sample_rate * max_clip_duration_s)
     overlap_size = int(sample_rate * overlap_duration_s)
     chunks = []
@@ -402,7 +409,7 @@ def find_split_point(
     RMS energy in sliding windows.
 
     Args:
-        wav: Audio array. Can be 1D or multi-dimensional.
+        wav: 1D mono audio array.
         start_idx: Start index of search region (inclusive).
         end_idx: End index of search region (exclusive).
         min_energy_window: Window size in samples for energy calculation.