Update app.py

app.py

@@ -9,6 +9,8 @@ Secrets / environment:
   INTERVIEWS_DB_PATH    optional, path of the SQLite file. Point it at a PERSISTENT
                         volume in production, otherwise the data is lost on redeploy.
   AZURE_SPEECH_KEY / AZURE_SPEECH_REGION   optional, better TTS voices
+Pre-generated audio:
+  welcome_audio/interview_welcome_{en,ar}.mp3   generate with gen_welcome.py and commit them
 """
 import os, io, re, json, hmac, html, base64, sqlite3, asyncio, logging, secrets, random
 from contextlib import closing
@@ -30,6 +32,7 @@ COMPANY = {"en": "Bin Dawood Holdings", "ar": "بن داود القابضة"}
 DB_PATH = os.environ.get("INTERVIEWS_DB_PATH", "interviews.db")
 LOGO_PATH = "logo.png"
 PHOTO_PATH = "agent_photo.jpg"
+WELCOME_DIR = Path("welcome_audio")    # pre-generated intro clips, committed to the repo
 LLM_MODEL = "openai/gpt-oss-120b"
 STT_MODEL = "whisper-large-v3"
 MIN_AUDIO_BYTES = 12000
@@ -45,11 +48,14 @@ AZURE_LOCALES = {"en": "en-US", "ar": "ar-SA"}
 
 SPOKEN = {
     "en": {
-        "greeting": ("Welcome to the Bin Dawood AI Bot interview. Hello {name}, I'm {agent}, an AI "
-                     "assistant, and I will be conducting your first-round interview for the "
-                     "{title} position at {company}. I will ask you {n} questions. Take your time "
-                     "with each answer, and feel free to speak or type. Are you ready? "
-                     "Can we start?"),
+        # Generic opening: pre-recorded once (see gen_welcome.py). If you change this text,
+        # delete welcome_audio/*.mp3 and regenerate so the audio matches what is on screen.
+        "greeting_intro": "Welcome to the Bin Dawood AI Bot interview. I'm Reem, an AI assistant.",
+        # Personalised part: always generated live.
+        "greeting_body": ("Hello {name}, I will be conducting your first-round interview for the "
+                          "{title} position at {company}. I will ask you {n} questions. Take your "
+                          "time with each answer, and feel free to speak or type. Are you ready? "
+                          "Can we start?"),
         "lets_begin": "Great, let's begin.",
         "not_ready": "No problem, take your time. Just let me know when you're ready to start.",
         "question": "Question {i} of {n}. {q}",
@@ -63,10 +69,10 @@ SPOKEN = {
                         "to you for the next steps. Thank you again, and goodbye."),
     },
     "ar": {
-        "greeting": ("أهلاً بكم في مقابلة بوت بن داود للذكاء الاصطناعي. مرحباً {name}، أنا {agent}، "
-                     "مساعدة ذكاء اصطناعي، وسأجري معكم المقابلة الأولى لوظيفة {title} في {company}. "
-                     "سأطرح عليكم {n} أسئلة، وخذوا وقتكم في كل إجابة، ويمكنكم التحدث أو الكتابة "
-                     "كما تفضلون. هل أنتم مستعدون؟ هل يمكننا أن نبدأ؟"),
+        "greeting_intro": "أهلاً بكم في مقابلة بوت بن داود للذكاء الاصطناعي. أنا ريم، مساعدة ذكاء اصطناعي.",
+        "greeting_body": ("مرحباً {name}، سأجري معكم المقابلة الأولى لوظيفة {title} في {company}. "
+                          "سأطرح عليكم {n} أسئلة، وخذوا وقتكم في كل إجابة، ويمكنكم التحدث أو الكتابة "
+                          "كما تفضلون. هل أنتم مستعدون؟ هل يمكننا أن نبدأ؟"),
         "lets_begin": "رائع، لنبدأ.",
         "not_ready": "لا بأس، خذوا وقتكم. أخبروني عندما تكونون مستعدين للبدء.",
         "question": "السؤال {i} من {n}. {q}",
@@ -92,8 +98,7 @@ NOTICE = {
         "Please do not share sensitive personal information (for example health details, religion, "
         "or ID numbers) in your answers.",
         "You may stop at any time. Contact HR to ask for access to, or deletion of, your data.",
-        "Please turn your volume on. After you start, tap the 🔊 button next to the avatar once "
-        "to hear the first message.",
+        "Please turn your volume on. The interviewer will speak to you as soon as you start.",
     ], "I have read the above and agree to take this interview."),
     "ar": ("قبل أن نبدأ", [
         "تتم هذه المقابلة بواسطة مساعد ذكاء اصطناعي وليس بواسطة موظف.",
@@ -104,7 +109,7 @@ NOTICE = {
         "القرارات الوظيفية يتخذها أشخاص وليس الذكاء الاصطناعي.",
         "يرجى عدم ذكر معلومات شخصية حساسة (مثل التفاصيل الصحية أو الدين أو أرقام الهوية) في إجاباتكم.",
         "يمكنكم التوقف في أي وقت، ويمكنكم التواصل مع الموارد البشرية لطلب الاطلاع على بياناتكم أو حذفها.",
-        "يرجى تشغيل الصوت. بعد البدء، اضغطوا على زر 🔊 بجانب الصورة مرة واحدة لسماع الرسالة الأولى.",
+        "يرجى تشغيل الصوت. ستتحدث إليكم المحاورة بمجرد البدء.",
     ], "قرأتُ ما سبق وأوافق على إجراء هذه المقابلة."),
 }
 
@@ -325,6 +330,24 @@ def tts_b64(text: str, lang: str) -> str:
     return base64.b64encode(data).decode() if data else ""
 
 
+def welcome_clip_b64(lang: str) -> str:
+    """Pre-generated intro clip (committed to the repo). If the file is missing, try to
+    generate it once; if that fails too, return '' and the caller speaks the full greeting."""
+    p = WELCOME_DIR / f"interview_welcome_{lang}.mp3"
+    if not p.exists():
+        text = SPOKEN[lang]["greeting_intro"]
+        data = _azure_tts(text, lang) or _edge_tts(text, lang) or _gtts(text, lang)
+        if not data:
+            return ""
+        try:
+            WELCOME_DIR.mkdir(exist_ok=True)
+            p.write_bytes(data)
+        except OSError as e:
+            log.warning("could not cache welcome clip: %s", e)
+            return base64.b64encode(data).decode()
+    return base64.b64encode(p.read_bytes()).decode()
+
+
 # ══════════════════════════════════════════════════════════════════════════
 #  LLM TASKS
 # ══════════════════════════════════════════════════════════════════════════
@@ -527,26 +550,43 @@ padding:9px 18px;font-size:.85rem;font-weight:600;cursor:pointer;}
 <button class="tap" id="tap">🔊 Tap to hear · اضغط للاستماع</button></div>
 <script>__SCRIPT__</script></body></html>"""
 
+# Plays a queue of base64 mp3 clips one after another (e.g. pre-recorded intro + live part).
 AUDIO_JS = """(function(){
 var a=document.getElementById('a'),c=document.getElementById('av'),
 r=document.getElementById('ring'),s=document.getElementById('status'),
 t=document.getElementById('tap');
-a.src='data:audio/mp3;base64,__AUDIO__';
-a.addEventListener('play',function(){t.style.display='none';c.classList.add('speaking');r.classList.add('active');s.textContent='Speaking...';});
-function stop(){c.classList.remove('speaking');r.classList.remove('active');s.textContent='Listening';}
-a.addEventListener('ended',stop);a.addEventListener('pause',stop);
-function p(){var x=a.play();if(x&&x.catch)x.catch(function(){t.style.display='inline-block';s.textContent='Tap to hear';});}
-t.addEventListener('click',function(){a.play().catch(function(){});});
-if(__MANUAL__){t.style.display='inline-block';s.textContent='Tap to hear';}
-else if(document.readyState==='complete')p();else window.addEventListener('load',p);})();"""
-
-
-def render_avatar(img_b64: str, name: str, audio_b64: str = "", manual: bool = False) -> str:
-    """manual=True skips autoplay and shows the tap button (browsers block the first autoplay)."""
+var srcs=__AUDIOS__.map(function(b64){
+  var bin=atob(b64),u=new Uint8Array(bin.length);
+  for(var k=0;k<bin.length;k++)u[k]=bin.charCodeAt(k);
+  return URL.createObjectURL(new Blob([u],{type:'audio/mpeg'}));
+});
+var i=0;
+function load(n){a.src=srcs[n];a.preload='auto';}
+function speaking(){t.style.display='none';c.classList.add('speaking');r.classList.add('active');s.textContent='Speaking...';}
+function idle(){c.classList.remove('speaking');r.classList.remove('active');s.textContent='Listening';}
+function needTap(){t.style.display='inline-block';s.textContent='Tap to hear';}
+a.addEventListener('play',speaking);
+a.addEventListener('ended',function(){
+  if(i+1<srcs.length){i++;load(i);a.play().catch(needTap);}else{idle();}
+});
+a.addEventListener('pause',function(){if(!a.ended&&i+1>=srcs.length)idle();});
+a.addEventListener('error',function(){s.textContent='Audio failed to load';});
+t.addEventListener('click',function(){a.play().catch(function(e){s.textContent='Cannot play: '+e.name;});});
+load(0);
+if(__MANUAL__){needTap();}
+else{var go=function(){var x=a.play();if(x&&x.catch)x.catch(needTap);};
+  if(document.readyState==='complete')go();else window.addEventListener('load',go);}
+})();"""
+
+
+def render_avatar(img_b64: str, name: str, audio=None, manual: bool = False) -> str:
+    """`audio` is a base64 string or a list of base64 strings (played in order).
+    manual=True skips autoplay and shows the tap button."""
+    clips = [audio] if isinstance(audio, str) and audio else [c for c in (audio or []) if c]
     img = (f'<img src="data:image/jpeg;base64,{img_b64}" alt="{html.escape(name)}">'
            if img_b64 else '<div class="ph">👩‍💼</div>')
-    script = (AUDIO_JS.replace("__AUDIO__", audio_b64)
-              .replace("__MANUAL__", "true" if manual else "false")) if audio_b64 else ""
+    script = (AUDIO_JS.replace("__AUDIOS__", json.dumps(clips))
+              .replace("__MANUAL__", "true" if manual else "false")) if clips else ""
     return (AVATAR_TEMPLATE.replace("__IMG__", img)
             .replace("__NAME__", html.escape(name)).replace("__SCRIPT__", script))
 
@@ -574,15 +614,24 @@ def begin_interview(inv, pos):
         "INSERT INTO interviews(invite_code, status, consent_at, started_at, questions_json) "
         "VALUES (?,?,?,?,?)",
         (inv["code"], "in_progress", now_iso(), now_iso(), json.dumps(questions, ensure_ascii=False)))
-    greeting = SPOKEN[lang]["greeting"].format(
-        name=inv["candidate_name"], agent=AGENT_NAME[lang], title=pos["title"],
-        company=COMPANY[lang], n=len(questions))
-    save_turn(iid, "interviewer", "greeting", None, greeting)
+
+    S = SPOKEN[lang]
+    intro = S["greeting_intro"]
+    body = S["greeting_body"].format(
+        name=inv["candidate_name"], title=pos["title"], company=COMPANY[lang], n=len(questions))
+    greeting = intro + " " + body
+    save_turn(iid, "interviewer", "greeting", None, greeting)   # full text is what HR and the candidate see
     ss.iv = {"id": iid, "lang": lang, "name": inv["candidate_name"], "title": pos["title"],
              "questions": questions, "q_index": 0, "followups_used": 0, "done": False,
              "stage": "ready"}
-    ss.pending_audio = tts_b64(greeting, lang)
-    ss.manual_audio = True   # first audio of the session: browsers block autoplay, show tap button
+
+    # Pre-recorded intro first; then only the personalised part live.
+    # If the clip is unavailable, speak the whole greeting live instead.
+    clip = welcome_clip_b64(lang)
+    live = tts_b64(body if clip else greeting, lang)
+    ss.pending_audio = [c for c in (clip, live) if c]
+    ss.manual_audio = False   # the candidate just clicked "Start", so autoplay is allowed;
+                              # the iframe falls back to the tap button if the browser refuses
     st.rerun()
 
 
@@ -597,7 +646,7 @@ def resume_interview(inv, pos, row):
              "questions": questions, "q_index": min(row["q_index"], len(questions) - 1),
              "followups_used": row["followups_used"], "done": False, "stage": stage}
     ss.pending_audio = tts_b64(last[-1]["content"], lang) if last else ""
-    ss.manual_audio = True   # fresh page load: no prior user gesture, autoplay will be blocked
+    ss.manual_audio = False   # the candidate just clicked "Resume"
     st.rerun()
 
 
@@ -1044,6 +1093,8 @@ def main():
                        initial_sidebar_state="collapsed")
     st.markdown(CSS, unsafe_allow_html=True)
     init_db()
+    for _lang in LANGS:          # creates welcome_audio/*.mp3 on first run if missing
+        welcome_clip_b64(_lang)
 
     logo = load_b64(LOGO_PATH)
     logo_html = f'<img class="hero-logo" src="data:image/png;base64,{logo}" alt="logo">' if logo else ""