diff --git a/chirp/audio_utils.py b/chirp/audio_utils.py
index 4ab8fb6c..96909381 100644
--- a/chirp/audio_utils.py
+++ b/chirp/audio_utils.py
@@ -88,18 +88,21 @@ def load_audio_file(
   # Handle other audio formats.
   # Because librosa passes file handles to soundfile, we need to copy the file
   # to a temporary file before passing it to librosa.
-  with tempfile.NamedTemporaryFile(mode='w+b', suffix=extension) as f:
+  with tempfile.NamedTemporaryFile(
+      mode='w+b', suffix=extension, delete=False
+  ) as f:
     with filepath.open('rb') as sf:
       f.write(sf.read())
-    # librosa outputs lots of warnings which we can safely ignore when
-    # processing all Xeno-Canto files and PySoundFile is unavailable.
-    with warnings.catch_warnings():
-      warnings.simplefilter('ignore')
-      audio, _ = librosa.load(
-          f.name,
-          sr=target_sample_rate,
-          res_type=resampling_type,
-      )
+  # librosa outputs lots of warnings which we can safely ignore when
+  # processing all Xeno-Canto files and PySoundFile is unavailable.
+  with warnings.catch_warnings():
+    warnings.simplefilter('ignore')
+    audio, _ = librosa.load(
+        f.name,
+        sr=target_sample_rate,
+        res_type=resampling_type,
+    )
+  os.unlink(f.name)
   return audio
 
 
@@ -252,20 +255,22 @@ def load_xc_audio(xc_id: str, sample_rate: int) -> jnp.ndarray:
     raise requests.exceptions.RequestException(
         f'Failed to load audio from Xeno-Canto {xc_id}'
     ) from e
-  with tempfile.NamedTemporaryFile(suffix='.mp3', mode='wb') as f:
+  with tempfile.NamedTemporaryFile(suffix='.mp3', mode='wb', delete=False) as f:
     f.write(data)
     f.flush()
-    audio = load_audio_file(f.name, target_sample_rate=sample_rate)
+  audio = load_audio_file(f.name, target_sample_rate=sample_rate)
+  os.unlink(f.name)
   return audio
 
 
 def load_url_audio(url: str, sample_rate: int) -> jnp.ndarray:
   """Load audio from a URL."""
   data = requests.get(url).content
-  with tempfile.NamedTemporaryFile(mode='wb') as f:
+  with tempfile.NamedTemporaryFile(mode='wb', delete=False) as f:
     f.write(data)
     f.flush()
-    audio = load_audio_file(f.name, target_sample_rate=sample_rate)
+  audio = load_audio_file(f.name, target_sample_rate=sample_rate)
+  os.unlink(f.name)
   return audio
 
 
@@ -631,7 +636,7 @@ def slice_peaked_audio(
 
   # Ensure that the peak locations are such that
   # `audio[peak - left_shift: peak + right_shift]` is a non-truncated slice.
-  peaks = jnp.clip(peaks, left_shift, audio.shape[0] - right_shift)
+  peaks = jnp.clip(peaks, left_shift, jnp.array(audio).shape[0] - right_shift)
   # As a result, it's possible that some (start, stop) pairs become identical;
   # eliminate duplicates.
   start_stop = jnp.unique(