Fix LUNA voice + transcription: edge-tts 7.2.8 (fixes 403), Whisper base model + en/VAD for accuracy

backend/services/interview_engine.py

@@ -104,7 +104,7 @@ def load_whisper_model():
             # lightweight model to improve startup times. Available options
             # include: tiny, base, base.en, small, medium, large. See
             # https://huggingface.co/ggerganov/whisper.cpp for details.
-            model_name = os.getenv("WHISPER_MODEL_NAME", "tiny")
+            model_name = os.getenv("WHISPER_MODEL_NAME", "base")
             whisper_model = WhisperModel(model_name, device=device, compute_type=compute_type)
             logging.info(f"Whisper model '{model_name}' loaded on {device} with {compute_type}")
         except Exception as e:
@@ -417,7 +417,16 @@ def whisper_stt(audio_path):
             return ""
 
         model = load_whisper_model()
-        segments, _ = model.transcribe(wav_path)
+        # language="en" avoids misdetecting the language; vad_filter drops
+        # silence so Whisper doesn't hallucinate phrases on quiet gaps;
+        # condition_on_previous_text=False keeps each answer independent.
+        segments, _ = model.transcribe(
+            wav_path,
+            language="en",
+            beam_size=5,
+            vad_filter=True,
+            condition_on_previous_text=False,
+        )
         transcript = " ".join(segment.text for segment in segments)
         return transcript.strip()
     except Exception as e:

requirements.txt

@@ -24,7 +24,7 @@ accelerate==0.29.3
 huggingface_hub==0.20.3
 faster-whisper==1.1.1
 av==13.1.0
-edge-tts==6.1.2
+edge-tts==7.2.8
 gunicorn
 python-dotenv
 chromadb>=0.4.0