diff --git a/omnivoice/utils/audio.py b/omnivoice/utils/audio.py
index 6b0ed317..75fa3307 100644
--- a/omnivoice/utils/audio.py
+++ b/omnivoice/utils/audio.py
@@ -65,18 +65,7 @@ def load_audio(audio_path: str, sampling_rate: int):
         # libraries are absent. The ``backend="soundfile"`` argument above
         # does NOT avoid that — 2.9 accepts and ignores it.
         aseg = AudioSegment.from_file(audio_path)
-        # Scale by the decoded sample width instead of a hardcoded 16-bit
-        # divisor. pydub reports 8-bit as sample_width 1 and widens 24-bit to
-        # a full-range int32 (sample_width 4), so /32768 makes 24- and 32-bit
-        # references 32768x too loud and 8-bit ones 256x too quiet.
-        audio_data = (
-            np.array(aseg.get_array_of_samples()).astype(np.float32)
-            / aseg.max_possible_amplitude
-        )
-        if aseg.channels == 1:
-            waveform = torch.from_numpy(audio_data).unsqueeze(0)
-        else:
-            waveform = torch.from_numpy(audio_data.reshape(-1, aseg.channels).T)
+        waveform = audiosegment_to_tensor(aseg)
         prompt_sampling_rate = aseg.frame_rate
 
     if prompt_sampling_rate != sampling_rate:
@@ -255,7 +244,7 @@ def audiosegment_to_tensor(aseg):
     audio_data = np.array(aseg.get_array_of_samples())
 
     # Convert to float32 and normalize to [-1, 1] range
-    audio_data = audio_data.astype(np.float32) / 32768.0
+    audio_data = audio_data.astype(np.float32) / aseg.max_possible_amplitude
 
     # Handle channels
     if aseg.channels == 1:
diff --git a/tests/test_load_audio_torchcodec_fallback.py b/tests/test_load_audio_torchcodec_fallback.py
index f1ebf211..d4360fd5 100644
--- a/tests/test_load_audio_torchcodec_fallback.py
+++ b/tests/test_load_audio_torchcodec_fallback.py
@@ -23,7 +23,6 @@ import pytest
 import soundfile as sf
 import torch
 
-from omnivoice.utils.audio import load_audio
 
 
 def _write_sine_wav(path, *, seconds: float = 0.5, sample_rate: int = 24000):
@@ -43,6 +42,7 @@ def _torchcodec_missing(*_a, **_kw):
 
 def test_load_audio_falls_back_when_torchcodec_missing(tmp_path, monkeypatch):
     """Without the ImportError catch this raises instead of returning audio."""
+    from omnivoice.utils.audio import load_audio
     import torchaudio
 
     ref = tmp_path / "ref.wav"
@@ -60,6 +60,7 @@ def test_load_audio_falls_back_when_torchcodec_missing(tmp_path, monkeypatch):
 
 def test_load_audio_fallback_resamples_to_target(tmp_path, monkeypatch):
     """The fallback path must still honour the requested sampling rate."""
+    from omnivoice.utils.audio import load_audio
     import torchaudio
 
     ref = tmp_path / "ref_16k.wav"
@@ -90,6 +91,7 @@ def test_load_audio_fallback_amplitude_matches_bit_depth(
     without TorchCodec it is the only path, which is what makes it a bug
     worth fixing here.
     """
+    from omnivoice.utils.audio import load_audio
     import torchaudio
 
     sample_rate = 24000
@@ -106,3 +108,14 @@ def test_load_audio_fallback_amplitude_matches_bit_depth(
         f"{subtype} decoded at the wrong scale: peak "
         f"{waveform.abs().max().item():.6f}, expected ~{peak}"
     )
+
+
+@pytest.mark.parametrize("subtype", ["PCM_U8", "PCM_16", "PCM_24", "PCM_32"])
+def test_audiosegment_conversion_preserves_amplitude(tmp_path, subtype):
+    from pydub import AudioSegment
+    from omnivoice.utils.audio import audiosegment_to_tensor
+    path = tmp_path / "stereo.wav"
+    samples = np.tile([0.5, -0.25], (100, 1))
+    sf.write(path, samples, 24000, subtype=subtype)
+    converted = audiosegment_to_tensor(AudioSegment.from_file(path)).numpy()
+    np.testing.assert_allclose(converted, samples.T, atol=0.01)