Modification of the structure

apps/stt/.env.example

@@ -1,11 +1,10 @@
-# Local Faster-Whisper (same engine as Pipecat local STT) — no GapGPT.
-# Sizes: tiny | base | small | medium | large-v3
-# Do NOT use cloud names like whisper-1
-# On low-RAM servers prefer: base or tiny
-STT_MODEL=base
+# Local Faster-Whisper — no GapGPT.
+# Persian dictation needs at least small (base/tiny are too inaccurate).
+# medium is better if the server has RAM.
+STT_MODEL=small
 STT_DEVICE=cpu
 STT_COMPUTE_TYPE=int8
-STT_PROMPT=یادداشت و مکالمهٔ فروش به زبان فارسی. اسامی مشتریان و اصطلاحات کسب‌وکار را دقیق بنویس.
+STT_PROMPT=این یک یادداشت فارسی است. لطفاً این مورد را بررسی کنید.
 
 STT_HOST=127.0.0.1
 STT_PORT=3040

apps/stt/main.py

@@ -46,20 +46,21 @@ def _require_token(header_token: str | None) -> None:
 
 def _model_name() -> str:
     """Map env / legacy cloud names to Faster-Whisper local sizes."""
-    raw = (_env("STT_MODEL", "base") or "base").lower()
+    raw = (_env("STT_MODEL", "small") or "small").lower()
     legacy = {
-        "whisper-1": "base",
-        "whisper": "base",
+        "whisper-1": "small",
+        "whisper": "small",
         "gpt-4o-transcribe": "small",
-        "gpt-4o-mini-transcribe": "base",
+        "gpt-4o-mini-transcribe": "small",
+        # tiny/base are too weak for Persian dictation
+        "tiny": "small",
+        "tiny.en": "small",
+        "base": "small",
+        "base.en": "small",
     }
     if raw in legacy:
         return legacy[raw]
     allowed = {
-        "tiny",
-        "tiny.en",
-        "base",
-        "base.en",
         "small",
         "small.en",
         "medium",
@@ -73,7 +74,7 @@ def _model_name() -> str:
         "distil-large-v3",
         "large-v3-turbo",
     }
-    return raw if raw in allowed else "base"
+    return raw if raw in allowed else "small"
 
 
 def _get_model() -> WhisperModel:
@@ -211,19 +212,38 @@ def _ensure_16k(wav_bytes: bytes) -> np.ndarray:
     return _resample_audio(audio, rate, 16000)
 
 
+def _normalize(audio: np.ndarray) -> np.ndarray:
+    if audio.size == 0:
+        return audio
+    peak = float(np.max(np.abs(audio)))
+    if peak < 1e-4:
+        return audio
+    # Bring quiet mic levels up without clipping
+    gain = min(8.0, 0.9 / peak)
+    if gain <= 1.05:
+        return audio
+    return np.clip(audio * gain, -1.0, 1.0).astype(np.float32)
+
+
 def _transcribe_local(wav_bytes: bytes) -> str:
     model = _get_model()
-    audio = _ensure_16k(wav_bytes)
+    audio = _normalize(_ensure_16k(wav_bytes))
     initial_prompt = _env(
         "STT_PROMPT",
-        "یادداشت و مکالمهٔ فروش به زبان فارسی. اسامی مشتریان و اصطلاحات کسب‌وکار را دقیق بنویس.",
+        "این یک یادداشت فارسی است. لطفاً این مورد را بررسی کنید.",
     ) or None
 
     segments, _info = model.transcribe(
         audio,
         language="fa",
+        task="transcribe",
         initial_prompt=initial_prompt,
-        vad_filter=True,
+        beam_size=5,
+        best_of=5,
+        temperature=0.0,
+        condition_on_previous_text=False,
+        vad_filter=False,
+        no_speech_threshold=0.45,
     )
     parts = [seg.text.strip() for seg in segments if getattr(seg, "text", None)]
     return " ".join(p for p in parts if p).strip()

apps/web/src/components/VoiceToTextButton.vue

@@ -244,7 +244,11 @@ async function start() {
       stream = null;
       return;
     }
+    try {
+      ctx = new Ctor({ sampleRate: 16000 });
+    } catch {
       ctx = new Ctor();
+    }
     await ctx.resume();
     sampleRate = ctx.sampleRate || 16000;
     source = ctx.createMediaStreamSource(stream);