diff --git a/backend/backends/pytorch_backend.py b/backend/backends/pytorch_backend.py
index af0f4a78..7e0040d2 100644
--- a/backend/backends/pytorch_backend.py
+++ b/backend/backends/pytorch_backend.py
@@ -10,6 +10,10 @@ import numpy as np
 
 logger = logging.getLogger(__name__)
 
+WHISPER_SAMPLE_RATE = 16000
+# Whisper encodes one 30s window per pass; longer audio needs sequential long-form decoding.
+WHISPER_WINDOW_SAMPLES = 30 * WHISPER_SAMPLE_RATE
+
 from . import TTSBackend, STTBackend, LANGUAGE_CODE_TO_NAME, WHISPER_HF_REPOS
 from .base import (
     is_model_cached,
@@ -337,7 +341,7 @@ class PyTorchSTTBackend:
         def _transcribe_sync():
             """Run synchronous transcription in thread pool."""
             # Load audio
-            audio, _sr = load_audio(audio_path, sample_rate=16000)
+            audio, _sr = load_audio(audio_path, sample_rate=WHISPER_SAMPLE_RATE)
 
             # Inference runs with the process's default HF_HUB_OFFLINE
             # state — forcing offline here (issue #462) broke online users
@@ -355,7 +359,7 @@ class PyTorchSTTBackend:
             # ("expects the mel input features to be of length 3000") when
             # generate() runs language detection, i.e. whenever no language
             # is forced.
-            is_long_form = len(audio) > 30 * 16000
+            is_long_form = len(audio) > WHISPER_WINDOW_SAMPLES
             processor_kwargs = (
                 {"truncation": False, "padding": "longest", "return_attention_mask": True}
                 if is_long_form
@@ -363,7 +367,7 @@ class PyTorchSTTBackend:
             )
             inputs = self.processor(
                 audio,
-                sampling_rate=16000,
+                sampling_rate=WHISPER_SAMPLE_RATE,
                 return_tensors="pt",
                 **processor_kwargs,
             )