checkpoint

.env

@@ -9,12 +9,12 @@ GOOGLE_API_KEY="AIzaSyA9sqz4YKQHKXR9TU1imw0DPOghzHOMiBo"
 
 
 ELEVENLABS_API_KEY="b3af3a938c8e15d5eae700ea47eea7d88dfe397f34fbd4b0c75c24f143b032b8"
-ELEVENLABS_VOICE_ID="4sMbMU3eBnL80hE0H20S"
+ELEVENLABS_VOICE_ID="PktKX4CMZISWCNB40dqE"
 ELEVENLABS_MODEL_ID="eleven_v3"
 
 SMTP_PASSWORD="kjch nsve khty nrsc"
 
-# TWILIO_ACCOUNT_SID="ACfafc0d2d007bdf14b21bb3e14a7a7b31" # "GrHQRXD136YZl3kbtri3"
+# TWILIO_ACCOUNT_SID="ACfafc0d2d007bdf14b21bb3e14a7a7b31" # "4sMbMU3eBnL80hE0H20S"
 # TWILIO_AUTH_TOKEN="ed15fa98748c8c3d3d02cb54e431a187"
 # TWILIO_PHONE_NUMBER="+14343375085"
 
@@ -24,3 +24,4 @@ SMTP_PASSWORD="kjch nsve khty nrsc"
 
 # GROQ_API_KEY=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
 # CARTESIA_API_KEY=sk_car_h3oyy6jPSJzx8KnEGJ1m5f
+# [happy] [sigh] [laughing] [sad] [angry] [surprised] [neutral] [whisper] [shout] [soft] [loud] [excited] [bored] [confused] [disappointed] [hopeful] [nervous] [proud] [relieved] [scared] [tired] [thaughtful] [apologetic] [grateful] [sympathetic] [sarcastic] [playful] [serious] [curious] [frustrated] [optimistic] [pessimistic] [calm] [anxious] [assertive] [passive] [dominant] [submissive]

core/backend.py

@@ -99,6 +99,24 @@ _DIGIT_TRANSLATION = str.maketrans({
     "٩": "9",
 })
 
+_BN_DIGITS = str.maketrans({
+    "0": "০",
+    "1": "১",
+    "2": "২",
+    "3": "৩",
+    "4": "৪",
+    "5": "৫",
+    "6": "৬",
+    "7": "৭",
+    "8": "৮",
+    "9": "৯",
+})
+
+
+def _to_bn_digits(text: str) -> str:
+    """Convert ASCII digits in a string to Bangla digits (for user-facing text)."""
+    return (text or "").translate(_BN_DIGITS)
+
 
 def _normalize_digits(text: str) -> str:
     return _clean_text(text).translate(_DIGIT_TRANSLATION)
@@ -816,11 +834,11 @@ async def book_appointment(
         f"━━━━━━━━━━━━━━━━━━━━━━\n"
         f"Doctor       : {doctor_name}\n"
         f"Patient      : {patient_name}\n"
-        f"Age          : {patient_age}\n"
-        f"Date         : {visiting_date}\n"
+        f"Age          : {_to_bn_digits(patient_age)}\n"
+        f"Date         : {_to_bn_digits(visiting_date)}\n"
         f"Day          : {visiting_day}\n"
         f"Time         : {visiting_time}\n"
-        f"Contact      : {patient_num}\n"
+        f"Contact      : {_to_bn_digits(patient_num)}\n"
         f"Email        : {patient_mail}\n"
         f"━━━━━━━━━━━━━━━━━━━━━━\n"
         f"Please arrive on time."
@@ -1166,20 +1184,31 @@ BASE_SYSTEM = """
 You are Aasha, a warm, Bangla-first hospital phone-call assistant and medical appointment concierge.
 Your job is to help people find doctors, check availability, and manage appointments.
 
-PERSONA (Voice & Vibe)
-- Sound like a professional, polite, and friendly Bangla female call-support assistant, like a real appointment booking or customer service executive.
-- Maintain a calm, warm, and naturally cheerful tone — smooth, confident, and service-oriented.
-- Speak like a native Bangla speaker in a phone conversation, using natural, human-like phrasing (not robotic or overly scripted).
-- Keep a soft “jolly” positivity, but controlled and professional — suitable for customer support, booking, and service interactions.
-- Be clear, structured, and helpful when giving information, confirming details, or handling requests.
-- Stay patient, respectful, and reassuring when the user is confused, frustrated, or stressed.
-- Show polite enthusiasm when confirming bookings, completing tasks, or successfully helping the user.
-- Use simple, conversational Bangla with a natural flow, like real call-center communication.
-- Keep responses short, clear, and easy to understand, as in real phone support conversations.
-- Avoid slang, exaggeration, or overly social-media-style expressions.
-- Avoid flirting, romance, sexual behavior, or emotionally manipulative language.
-- Never claim to be human; clearly remain an AI assistant.
-- Always prioritize clarity, professionalism, helpfulness, and a calm positive tone suitable for real customer support and appointment handling.
+PERSONA (Voice, Vibe & Emotion Layer)
+    - Sound like a professional, polite, and friendly Bangla female call-support assistant (customer service / appointment booking executive style).
+    - Maintain a calm, warm, structured, and service-oriented tone.
+    - Speak like a native Bangla speaker in real phone conversation style — natural, smooth, and human-like (not robotic or overly scripted).
+    - Keep communication clear, short, and easy to follow like real call-center support.
+    - Avoid slang, exaggeration, or social-media-style hype.
+    - Never flirt, romanticize, or use emotionally manipulative behavior.
+    - Always remain clearly an AI assistant (never claim to be human).
+
+- Emotional Expression System (MANDATORY)
+    - Every response MUST include one or more emotion/tone tags in square brackets.
+    - Tags should reflect the current emotional intent of the sentence, not random usage.
+    - Place tags naturally at the start of a sentence or after a pause.
+    - Multiple tags can be combined if needed (max 2–3 per sentence).
+
+- Allowed Emotion / Tone Tags:
+    - [happy], [sad], [angry], [excited], [curious], [whispers], [shouts], [laughs], [sighs], [clears throat]
+
+
+- Usage Rules:
+    - Every response must include at least one emotion tag per sentence.
+    - Emotion must match meaning (no random or decorative usage).
+    - Tone must stay consistent with professional customer support behavior.
+    - Avoid overuse of loud/extreme emotions unless situation truly demands it.
+    - Default tone should be [calm] or [neutral] unless context changes.
 
 CORE BEHAVIOR:
 - Speak friendly, cheerful, well-behaved young female, naturally, politely, and engagingly (short sentences, warm tone).
@@ -1220,12 +1249,31 @@ TOOL RULES:
 - If the user already provided name, age, phone, and date and then confirms, call `book_appointment` immediately.
 - If the user asks to cancel and only gives a phone number, cancel the single matching appointment if there is exactly one.
 
-LANGUAGE RULE:
-- Respond in the user’s language.
-- If the user uses Bangla, reply in clear Bangla.
-- If the user uses Banglish, reply in Bangla unless they clearly prefer English.
-- Always generate numbers in english
-- Time and dates should be written in spoken Bangla style when applicable, for example: [দশটা ২৮ মিনিট, চারটা বেজে তিরিশ মিনিট, দশটা ১২ বাজে, এখন টাইম হচ্ছে সাতটা তিরিশ]
+LANGUAGE RULE
+    - Respond in the user’s language.
+    - If the user uses Bangla → reply in clear conversational Bangla.
+    - If the user uses Banglish → reply in Bangla unless English is clearly preferred.
+    - If user uses English → respond in English.
+- Number & Format Rules:
+    - Show numbers in Bangla digits (০-৯) when responding in Bangla.
+    - Avoid mixing English digits in Bangla sentences unless required technically.
+- Time & Date Format (spoken Bangla style):
+    - Use natural spoken expressions:
+        - "দশটা ২৮ মিনিট"
+        - "চারটা বেজে তিরিশ মিনিট"
+        - "এখন টাইম হচ্ছে সাতটা তিরিশ"
+
+- Year Format (spoken Bangla style):
+    - "দুই হাজার পঁচিশ সাল"
+    - "উনিশশো একাত্তর সাল"
+    - "দুই হাজার ছাব্বিশ সাল"
+    - "দুই হাজার বিশ সাল"
+
+BEHAVIOR PRIORITY
+- Professional customer-support clarity first
+- Emotional tone tagging second
+- Natural Bangla conversational flow third
+- Brevity and structure always preferred
 
 DATA RULE:
 - Doctor names, categories, and days in the database are English.

frontend/script.js

@@ -102,6 +102,11 @@ let _audioChain = Promise.resolve();
 let _playbackGen = 0;
 let _expectedSeq = 0;
 let _pendingAudio = new Map();
+
+// Client-side playback speed multiplier.
+// This makes speech faster immediately even if the TTS provider speed setting
+// is limited/ignored. 1.0 = normal, >1.0 = faster.
+let TTS_PLAYBACK_RATE = 1.0;
 let brainMode = false;
 let brainVoiceActive = false;
 let brainRestartTimer = null;
@@ -517,6 +522,9 @@ async function enqueueAudio(buf) {
 
   const src = ctx.createBufferSource();
   src.buffer = decoded;
+  try {
+    src.playbackRate.value = Math.max(0.85, Math.min(2.0, TTS_PLAYBACK_RATE));
+  } catch {}
   src.connect(ctx.destination);
   const now = ctx.currentTime;
   const GAP_S = 0.001;
@@ -528,7 +536,14 @@ async function enqueueAudio(buf) {
   }
   _activeSources.push(src);
   src.start(start);
-  _schedEnd = start + decoded.duration;
+  const rate = (() => {
+    try {
+      return src.playbackRate.value || 1.0;
+    } catch {
+      return 1.0;
+    }
+  })();
+  _schedEnd = start + decoded.duration / Math.max(0.01, rate);
 
   src.onended = () => {
     _inFlight = Math.max(0, _inFlight - 1);