fix(asr): load audio from in-memory buffer instead of disk

Replaced file-based audio loading with io.BytesIO to handle uploaded audio directly in memory.
Librosa/SoundFile reads the buffer, resamples to 16kHz if needed, and feeds Whisper ASR model.

app/routes/asr.py

@@ -1,6 +1,8 @@
 from transformers import WhisperForConditionalGeneration, WhisperProcessor
 from fastapi import APIRouter, Request, UploadFile, File
 import librosa
+import io
+import soundfile as sf
 import os
 
 router = APIRouter()
@@ -15,12 +17,14 @@ async def asr(request: Request, audio: UploadFile = File(...)):
     # Get the loaded ASR model and processor
     processor, model = request.app.state.processor_asr, request.app.state.model_asr
     # Audio file path
-    audio_path = os.path.join(request.app.state.AUDIO_DIR, "temp", audio.filename)
-    with open(audio_path, "wb") as f:
-        f.write(await audio.read())
+    audio_bytes = await audio.read()
+    buffer = io.BytesIO(audio_bytes)
 
     # Loading audio file 
-    audio_data, sampling_rate = librosa.load(audio_path, sr=16000)
+    audio_data, sampling_rate = sf.read(buffer, dtype="float32")
+    if sampling_rate != 16000:
+        audio_data = librosa.resample(audio_data, orig_sr=sampling_rate, target_sr=16000)
+        sampling_rate = 16000
 
     # Preparing input data
     inputs = processor(audio_data, return_tensors="pt", sampling_rate=sampling_rate)