Replace gTTS realtime speech with Piper

CODEX_TASK_PIPER.md

@@ -0,0 +1,34 @@
+# Codex task: replace gTTS realtime speech synthesis with Piper
+
+Repository: trinitron88/ChinTranslator
+Base branch: main
+Issue: #11
+
+Goal
+Replace per-utterance gTTS MP3 generation in the realtime Hugging Face Space with local Piper TTS to reduce latency, network dependency, and phone battery drain.
+
+Context
+hf_space/app.py currently calls gTTS, writes an MP3 temp file, reloads it via librosa, converts it to PCM, and yields it back to FastRTC. This works but is slow and network-bound. The project roadmap already calls out Piper as the next step for low-latency local TTS.
+
+Suggested implementation
+- Work primarily in hf_space/app.py and hf_space/requirements.txt.
+- Add backend config with environment variables such as TTS_BACKEND, PIPER_MODEL, and PIPER_CONFIG.
+- Keep gTTS as fallback while Piper is being tested.
+- Implement a speak_piper(text, lang) path for English output in Chin to English mode.
+- Return FastRTC-compatible sample rate and int16 PCM audio.
+- Avoid shelling out per phrase if a Python API or persistent process is practical; otherwise document the tradeoff.
+- English to Chin can remain text-only unless a suitable Chin voice exists.
+- Update REALTIME.md and/or README.md with required Space variables and model setup.
+
+Acceptance criteria
+- Chin to English mode speaks English using Piper when configured.
+- App still boots if Piper is unavailable and falls back cleanly to gTTS or text-only.
+- Per-utterance latency is materially lower than gTTS in local or HF testing.
+- No temp MP3 round-trip is required for the Piper path.
+- Docs explain how to configure Piper in the Space.
+
+Suggested validation
+- Run or import hf_space/app.py enough to confirm syntax and imports.
+- Test the gTTS fallback still behaves like current production.
+- Test text-only mode if implemented.
+- Test Piper path on a short English phrase and confirm PCM output shape and sample rate are accepted by FastRTC.

hf_space/app.py

@@ -61,6 +61,9 @@
 
 CHIN_CODE = "cnh"
 
+APP_VERSION = os.environ.get("APP_VERSION", "v5.0.1")
+print(f"[app] version={APP_VERSION}", flush=True)
+
 # Skip translating/speaking when the input is detected as English. Besides being
 # the requested behavior, this breaks the TTS→mic feedback loop (the English we
 # speak won't get re-transcribed and echoed). Tunable via env: SKIP_ENGLISH=0
@@ -332,7 +335,7 @@ def _update_transcript(current: str, chin: str, english: str) -> str:
     server_rtc_configuration=server_rtc_configuration,
     # Short, one-line title so it doesn't wrap and overlap the record button
     # on mobile.
-    ui_args={"title": "Hahka Chin Audio Translator"},
+    ui_args={"title": "Hakha Chin Audio Translator"},
 )
 
 # Spaces (gradio SDK) serves this `demo` object. Append a mic-sensitivity slider;
@@ -375,8 +378,11 @@ def _update_transcript(current: str, chin: str, english: str) -> str:
         ".button-wrap .mute-button {"
         " padding: var(--size-2) var(--size-4) !important;"
         " border-radius: var(--radius-lg) !important; }"
+        ".app-version { font-size: 0.85rem; opacity: 0.7; text-align: center;"
+        " margin: -0.25rem 0 0.75rem; }"
         "</style>"
     )
+    gr.HTML(f"<div class='app-version'>Build: {APP_VERSION}</div>")
     direction = gr.Radio(
         choices=[("Hakha Chin → English", "cnh2en"),
                  ("English → Hakha Chin", "en2cnh")],