Restructuring STT

apps/stt/main.py

@@ -51,21 +51,36 @@ def _stt_service() -> WhisperSTTService:
 
     device = _env("STT_DEVICE", "cpu") or "cpu"
     compute = _env("STT_COMPUTE_TYPE", "int8") or "int8"
-    prompt = _env(
-        "STT_PROMPT",
-        "یادداشت و مکالمهٔ فروش به زبان فارسی. اسامی مشتریان و اصطلاحات کسب‌وکار را دقیق بنویس.",
-    )
 
     try:
+        # Settings fields differ by Pipecat version — keep to model + language only
+        settings_kwargs: dict = {"model": _model_name()}
+        try:
+            settings_kwargs["language"] = Language.FA
+            settings = WhisperSTTService.Settings(**settings_kwargs)
+        except TypeError:
+            settings_kwargs.pop("language", None)
+            settings = WhisperSTTService.Settings(**settings_kwargs)
+
         _stt = WhisperSTTService(
             device=device,
             compute_type=compute,
-            settings=WhisperSTTService.Settings(
+            settings=settings,
-                model=_model_name(),
-                language=Language.FA,
-                initial_prompt=prompt or None,
-            ),
         )
+    except TypeError:
+        # Older constructor: model=/language= kwargs instead of settings=
+        try:
+            _stt = WhisperSTTService(
+                model=_model_name(),
+                device=device,
+                compute_type=compute,
+                language=Language.FA,
+            )
+        except Exception as exc:
+            raise HTTPException(
+                status_code=503,
+                detail=f"بارگذاری مدل Whisper محلی ناموفق بود: {exc}",
+            ) from exc
     except Exception as exc:
         raise HTTPException(
             status_code=503,
@@ -178,11 +193,20 @@ def _transcribe_local(wav_bytes: bytes) -> str:
             audio, rate = _wav_to_float32(dst.read_bytes())
 
     settings = getattr(service, "_settings", None)
-    language = getattr(settings, "language", Language.FA)
+    language = getattr(settings, "language", None) or Language.FA
     lang_code = getattr(language, "value", language) if language is not None else "fa"
     if isinstance(lang_code, str) and "-" in lang_code:
         lang_code = lang_code.split("-")[0]
-    initial_prompt = getattr(settings, "initial_prompt", None)
+    # Language enum may be "fa" or "FA" or full name — faster-whisper wants ISO code
+    if isinstance(lang_code, str):
+        lang_code = lang_code.lower()
+        if lang_code in ("persian", "farsi"):
+            lang_code = "fa"
+
+    initial_prompt = _env(
+        "STT_PROMPT",
+        "یادداشت و مکالمهٔ فروش به زبان فارسی. اسامی مشتریان و اصطلاحات کسب‌وکار را دقیق بنویس.",
+    ) or None
 
     segments, _info = model.transcribe(
         audio,