Fix audio input parsing for browser/client variants

app.py

@@ -2,16 +2,17 @@ import numpy as np
 import torch
 import gradio as gr
 from transformers import WhisperForConditionalGeneration, WhisperProcessor
+from transformers.pipelines.audio_utils import ffmpeg_read
 
 MODEL_ID = "openai/whisper-small"
+TARGET_SAMPLE_RATE = 16000
 
 processor = WhisperProcessor.from_pretrained(MODEL_ID)
 model = WhisperForConditionalGeneration.from_pretrained(MODEL_ID)
 model.eval()
 
 
-def _to_float32_audio(audio: tuple[int, np.ndarray]) -> tuple[int, np.ndarray]:
-    sample_rate, data = audio
+def _normalize_waveform(data: np.ndarray) -> np.ndarray:
     if data.ndim > 1:
         data = data.mean(axis=1)
 
@@ -24,14 +25,35 @@ def _to_float32_audio(audio: tuple[int, np.ndarray]) -> tuple[int, np.ndarray]:
         if peak > 1.0:
             data = data / peak
 
-    return sample_rate, data
+    return data
+
+
+def _extract_audio(audio_input):
+    # Gradio may send audio as (sample_rate, np.ndarray), a filepath string, or a FileData-like dict.
+    if isinstance(audio_input, tuple) and len(audio_input) == 2:
+        sample_rate, data = audio_input
+        return int(sample_rate), _normalize_waveform(data)
+
+    path = None
+    if isinstance(audio_input, str):
+        path = audio_input
+    elif isinstance(audio_input, dict):
+        path = audio_input.get("path")
+
+    if path:
+        with open(path, "rb") as f:
+            audio_bytes = f.read()
+        data = ffmpeg_read(audio_bytes, TARGET_SAMPLE_RATE)
+        return TARGET_SAMPLE_RATE, _normalize_waveform(data)
+
+    raise ValueError("Unsupported audio input format. Please upload a valid audio file.")
 
 
 def transcribe_audio(audio, task):
     if audio is None:
         return "Please upload or record audio first."
 
-    sample_rate, data = _to_float32_audio(audio)
+    sample_rate, data = _extract_audio(audio)
 
     inputs = processor(
         data,