Improve Bambara/Fula quality + add full correction UI

voice_responder.py:
- Every template now returns (native_text, english_translation) tuple
- Added greeting/farewell/thanks/soil_ok/weather_ok/pest_ok templates
- Short MMS-TTS-optimised sentences kept (≤6 words)

intent_parser.py:
- Added greeting, thanks, farewell intents with Bambara/Fula/French/English keywords
- Added French and English keywords to agricultural intents

app.py:
- _run_pipeline returns (transcript, english_translation, response_text, audio)
- Tab 1: added English translation box; added "Send to Correction" button
- Tab 2: added English translation field + corrected_english field + corrected_response field
- "Send to Correction" auto-populates all Tab 2 fields from Tab 1 in one click
- _save_feedback_to_hub stores corrected_english and corrected_response

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -250,12 +250,12 @@ def _run_pipeline(audio_path: str, language_code: str):
         })
 
     responder = VoiceResponder(language=language_code)
-    response_text = responder.generate_response(intent, sensor_data)
+    response_text, english_translation = responder.generate_response(intent, sensor_data)
 
     # ── 3. MMS-TTS (GPU) ──────────────────────────────────────────────────────
     wav_np, sample_rate = _tts.synthesize(response_text, language_code, device=device)
 
-    return transcript, response_text, (sample_rate, wav_np)
+    return transcript, english_translation, response_text, (sample_rate, wav_np)
 
 
 # ── HF Hub feedback persistence ───────────────────────────────────────────────
@@ -264,7 +264,10 @@ def _save_feedback_to_hub(
     audio_path: str | None,
     transcript: str,
     corrected_text: str,
+    english_translation: str,
+    corrected_english: str,
     response_text: str,
+    corrected_response: str,
     rating: int,
     notes: str,
     language_label: str,
@@ -272,7 +275,7 @@ def _save_feedback_to_hub(
     language_code = SUPPORTED_LANGUAGES.get(language_label, "bam")
 
     if not corrected_text.strip():
-        return "⚠️ Corrected text is empty."
+        return "⚠️ Corrected transcription is empty — please fill in what was actually said."
 
     timestamp = datetime.now(timezone.utc).strftime("%Y%m%d_%H%M%S_%f")
 
@@ -283,7 +286,10 @@ def _save_feedback_to_hub(
         "audio_file": f"audio/{language_code}_{timestamp}.wav",
         "whisper_output": transcript,
         "corrected_text": corrected_text.strip(),
+        "english_translation": english_translation.strip(),
+        "corrected_english": corrected_english.strip() or english_translation.strip(),
         "response_text": response_text,
+        "corrected_response": corrected_response.strip() or response_text.strip(),
         "rating": rating,
         "notes": notes.strip(),
         "is_correction": transcript.strip() != corrected_text.strip(),
@@ -424,19 +430,19 @@ def _get_adapter_status() -> str:
 
 def handle_ask(audio_path, language_label):
     if audio_path is None:
-        return "⚠️ No audio — press Record or upload a file.", "", None
+        return "⚠️ No audio — press Record or upload a file.", "", "", None
 
     language_code = SUPPORTED_LANGUAGES.get(language_label, "bam")
     status = _ensure_whisper_loaded()
 
     if _whisper_model is None:
-        return f"⏳ Model loading ({status}). Wait a moment and try again.", "", None
+        return f"⏳ Model loading ({status}). Wait a moment and try again.", "", "", None
 
     try:
-        transcript, response_text, audio_out = _run_pipeline(audio_path, language_code)
-        return transcript, response_text, audio_out
+        transcript, english_translation, response_text, audio_out = _run_pipeline(audio_path, language_code)
+        return transcript, english_translation, response_text, audio_out
     except Exception as e:
-        return f"❌ {e}", "", None
+        return f"❌ {e}", "", "", None
 
 
 # ── Gradio UI ─────────────────────────────────────────────────────────────────
@@ -458,10 +464,10 @@ def build_ui() -> gr.Blocks:
         status_timer = gr.Timer(value=3)
         status_timer.tick(fn=get_model_status, outputs=model_status_box)
 
-        with gr.Tabs():
+        with gr.Tabs() as tabs:
 
             # ── Tab 1: Voice Assistant ────────────────────────────────────────
-            with gr.TabItem("🎙️ Voice Assistant"):
+            with gr.TabItem("🎙️ Voice Assistant", id="tab_voice"):
                 with gr.Row():
                     with gr.Column(scale=1):
                         language_dd = gr.Dropdown(
@@ -478,14 +484,20 @@ def build_ui() -> gr.Blocks:
 
                     with gr.Column(scale=1):
                         transcript_box = gr.Textbox(
-                            label="Whisper heard",
-                            lines=3,
+                            label="Whisper heard (transcription)",
+                            lines=2,
                             placeholder="Your words will appear here…",
                             interactive=False,
                         )
+                        translation_box = gr.Textbox(
+                            label="English translation",
+                            lines=2,
+                            placeholder="English meaning will appear here…",
+                            interactive=False,
+                        )
                         response_box = gr.Textbox(
-                            label="Response / Jaabi",
-                            lines=3,
+                            label="Response in your language",
+                            lines=2,
                             placeholder="Agricultural advice will appear here…",
                             interactive=False,
                         )
@@ -494,18 +506,23 @@ def build_ui() -> gr.Blocks:
                             autoplay=True,
                             interactive=False,
                         )
+                        correct_btn = gr.Button(
+                            "✏️ Something wrong? Send to Correction tab",
+                            variant="secondary",
+                            size="sm",
+                        )
 
                 ask_btn.click(
                     fn=handle_ask,
                     inputs=[audio_input, language_dd],
-                    outputs=[transcript_box, response_box, audio_output],
+                    outputs=[transcript_box, translation_box, response_box, audio_output],
                 )
 
             # ── Tab 2: Feedback & Correction ─────────────────────────────────
-            with gr.TabItem("📝 Feedback & Correction"):
+            with gr.TabItem("📝 Feedback & Correction", id="tab_feedback"):
                 gr.Markdown(
-                    "Help improve the model by correcting transcription errors. "
-                    "Your audio and corrections are saved to the training dataset."
+                    "Correct what Whisper heard, the English translation, and the response. "
+                    "All corrections are saved to the training dataset to improve future accuracy."
                 )
                 with gr.Row():
                     with gr.Column():
@@ -517,35 +534,53 @@ def build_ui() -> gr.Blocks:
                         fb_audio = gr.Audio(
                             sources=["microphone", "upload"],
                             type="filepath",
-                            label="Audio (re-record or upload)",
+                            label="Audio",
                         )
+                        gr.Markdown("**Step 1 — Fix the transcription**")
                         fb_transcript = gr.Textbox(
-                            label="Whisper output (what it heard)",
-                            lines=3,
-                            placeholder="Paste or type what Whisper said…",
+                            label="What Whisper heard",
+                            lines=2,
+                            placeholder="Auto-filled from Tab 1…",
                         )
                         fb_corrected = gr.Textbox(
-                            label="Corrected transcription (what was actually said)",
-                            lines=3,
-                            placeholder="Type the correct text here…",
+                            label="✏️ What was actually said (in Bambara/Fula)",
+                            lines=2,
+                            placeholder="Type the correct transcription here…",
                         )
 
                     with gr.Column():
+                        gr.Markdown("**Step 2 — Fix the English translation**")
+                        fb_english = gr.Textbox(
+                            label="Auto-generated English translation",
+                            lines=2,
+                            placeholder="Auto-filled from Tab 1…",
+                        )
+                        fb_corrected_english = gr.Textbox(
+                            label="✏️ Correct English translation",
+                            lines=2,
+                            placeholder="Type the correct English meaning here…",
+                        )
+                        gr.Markdown("**Step 3 — Fix the response**")
                         fb_response = gr.Textbox(
-                            label="Response text (optional — for rating)",
+                            label="Auto-generated response",
                             lines=2,
-                            placeholder="Copy the response from Tab 1…",
+                            placeholder="Auto-filled from Tab 1…",
+                        )
+                        fb_corrected_response = gr.Textbox(
+                            label="✏️ Better response (in farmer's language)",
+                            lines=2,
+                            placeholder="Type a better response here…",
                         )
                         fb_rating = gr.Slider(
                             minimum=1, maximum=5, step=1, value=3,
-                            label="Response quality (1 = poor, 5 = excellent)",
+                            label="Overall quality (1 = poor, 5 = excellent)",
                         )
                         fb_notes = gr.Textbox(
                             label="Notes (optional)",
                             lines=2,
                             placeholder="e.g. noisy background, strong accent…",
                         )
-                        save_btn = gr.Button("💾 Save to Dataset", variant="secondary")
+                        save_btn = gr.Button("💾 Save to Dataset", variant="primary")
                         save_status = gr.Textbox(
                             label="Save status", interactive=False, lines=2
                         )
@@ -554,11 +589,21 @@ def build_ui() -> gr.Blocks:
                     fn=_save_feedback_to_hub,
                     inputs=[
                         fb_audio, fb_transcript, fb_corrected,
-                        fb_response, fb_rating, fb_notes, fb_lang,
+                        fb_english, fb_corrected_english,
+                        fb_response, fb_corrected_response,
+                        fb_rating, fb_notes, fb_lang,
                     ],
                     outputs=[save_status],
                 )
 
+            # Wire "Send to Correction" button — populates Tab 2 fields from Tab 1
+            correct_btn.click(
+                fn=lambda t, tr, r, lang: (t, t, tr, tr, r, r, lang),
+                inputs=[transcript_box, translation_box, response_box, language_dd],
+                outputs=[fb_transcript, fb_corrected, fb_english, fb_corrected_english,
+                         fb_response, fb_corrected_response, fb_lang],
+            )
+
             # ── Tab 3: Training Status ────────────────────────────────────────
             with gr.TabItem("🔧 Training Status"):
                 gr.Markdown(

src/iot/intent_parser.py

@@ -18,25 +18,56 @@ class Intent:
 
 # Intent keyword taxonomy for Bambara (bam) and Fula (ful)
 INTENT_KEYWORDS: dict[str, dict[str, list[str]]] = {
+    # Social / conversational
+    "greeting": {
+        "bam": ["i ni ce", "i ni sogoma", "i ni wula", "ani sogoma", "an ka"],
+        "ful": ["jam waali", "jam hiiri", "jam na", "no mbadda"],
+        "fr":  ["bonjour", "bonsoir", "salut", "bonne nuit"],
+        "en":  ["hello", "hi", "good morning", "good evening", "hey"],
+    },
+    "thanks": {
+        "bam": ["aw ni ce", "i ni ce barika", "a ni barika"],
+        "ful": ["jaraama", "barakallahu"],
+        "fr":  ["merci", "je vous remercie"],
+        "en":  ["thank", "thanks"],
+    },
+    "farewell": {
+        "bam": ["kana tɛmɛ", "i ka taa", "sini"],
+        "ful": ["yahdu jam", "o yahdu", "jemma"],
+        "fr":  ["au revoir", "à bientôt", "bonne journée"],
+        "en":  ["goodbye", "bye", "see you"],
+    },
+    # Agricultural
     "check_soil": {
         "bam": ["bunding", "nɔgɔ", "dugu", "foro", "sani"],
         "ful": ["leydi", "ngesa", "ladde"],
+        "fr":  ["sol", "terre", "humidité"],
+        "en":  ["soil", "ground", "moisture", "dirt"],
     },
     "check_weather": {
         "bam": ["teliman", "sanji", "dibi", "sira"],
-        "ful": ["yeeso", "fuɗorde"],
+        "ful": ["yeeso", "fuɗorde", "ndiyam"],
+        "fr":  ["météo", "temps", "pluie", "chaleur"],
+        "en":  ["weather", "rain", "temperature", "hot"],
     },
     "irrigation_status": {
-        "bam": ["ji", "sanji", "foro"],
-        "ful": ["ndiyam", "ngesa"],
+        "bam": ["ji", "sanji", "foro ji"],
+        "ful": ["ndiyam", "ngesa ndiyam"],
+        "fr":  ["irrigation", "arrosage", "eau"],
+        "en":  ["irrigation", "water", "watering"],
     },
     "pest_alert": {
-        "bam": ["kungoloni", "suruku"],
-        "ful": ["biñ-biñ"],
+        "bam": ["kungoloni", "suruku", "dɔgɔw"],
+        "ful": ["biñ-biñ", "kuuje"],
+        "fr":  ["insecte", "nuisible", "ravageur"],
+        "en":  ["pest", "insect", "bug"],
     },
 }
 
 INTENT_ENTITIES = {
+    "greeting": "social",
+    "thanks": "social",
+    "farewell": "social",
     "check_soil": "soil",
     "check_weather": "weather",
     "irrigation_status": "irrigation",

src/iot/voice_responder.py

@@ -1,7 +1,8 @@
 """
 Generates voice response text from sensor data in the farmer's own language.
 Supports Bambara (bam), Fula (ful), French (fr), and English (en).
-Bambara/Fula templates use short sentences (≤15 words) for best MMS-TTS quality.
+Bambara/Fula templates use short sentences (≤6 words) for best MMS-TTS quality.
+Each template has an English equivalent used for the translation display in the UI.
 """
 from __future__ import annotations
 
@@ -20,55 +21,140 @@ TEMP_HIGH = 38.0
 PEST_ALERT_HIGH = 2         # Alert level ≥ 2 → warning
 
 # ── Bambara templates (≤6 words per sentence for clear MMS-TTS output) ───────
-BAMBARA_TEMPLATES = {
-    "soil_moisture_low":   "Bunding ji dɔgɔ. I ka foro ji.",
-    "soil_moisture_high":  "Ji ca kojugu. Foro ma fɛ.",
-    "soil_ph_low":         "Bunding kɔnɔ jugu. Kalisi fara a kan.",
-    "soil_ph_high":        "Bunding kɔnɔ tɛmɛ. Soufre fara a kan.",
-    "weather_hot":         "Teliman gbɛlɛ. Tile ma sigi.",
-    "rain_likely":         "Sanji bɛ na. Sɔrɔ jɔ.",
-    "pest_high":           "Dɔgɔw bɛ foro kɔnɔ. Bɔ u.",
-    "irrigation_needed":   "Foro fɛ ji. Ji sira yɔrɔ.",
-    "irrigation_active":   "Ji bɛ taa. A bɛ kɛ cogo di.",
-    "default":             "Kabako jumanw sɔrɔla.",
+# Format: (bambara_text, english_translation)
+BAMBARA_TEMPLATES: dict[str, tuple[str, str]] = {
+    # Greetings / social
+    "greeting":            ("I ni ce. N bɛ i dɛmɛ.",
+                            "Hello. I am here to help you."),
+    "greeting_morning":    ("I ni sogoma. Sɔrɔ ka ɲi.",
+                            "Good morning. May the harvest be good."),
+    "greeting_evening":    ("I ni wula. Kɛnɛya.",
+                            "Good evening. Good health to you."),
+    "thanks":              ("Aw ni ce. A ka ɲi.",
+                            "Thank you. That is good."),
+    "farewell":            ("Kana tɛmɛ. Ne bɛ i kɔnɔ.",
+                            "Goodbye. I will be thinking of you."),
+    "not_understood":      ("N ma a faamu. A fɔ.",
+                            "I did not understand. Please repeat."),
+    # Soil
+    "soil_moisture_low":   ("Bunding ji dɔgɔ. I ka foro ji.",
+                            "Soil moisture is low. Irrigate your field."),
+    "soil_moisture_high":  ("Ji ca kojugu. Foro ma fɛ.",
+                            "Too much water. The field is flooded."),
+    "soil_ph_low":         ("Bunding kɔnɔ jugu. Kalisi fara a kan.",
+                            "Soil is too acidic. Apply lime."),
+    "soil_ph_high":        ("Bunding kɔnɔ tɛmɛ. Soufre fara a kan.",
+                            "Soil is too alkaline. Apply sulphur."),
+    "soil_ok":             ("Bunding ka ɲi. Foro sɔrɔ.",
+                            "Soil conditions are good. The field is ready."),
+    # Weather
+    "weather_hot":         ("Teliman gbɛlɛ. Tile ma sigi.",
+                            "It is very hot. Do not work in the midday sun."),
+    "rain_likely":         ("Sanji bɛ na. Sɔrɔ jɔ.",
+                            "Rain is coming. Protect the harvest."),
+    "weather_ok":          ("Yɔrɔ min ɲi. Kɛ.",
+                            "Weather is fine. You can work."),
+    # Pest
+    "pest_high":           ("Dɔgɔw bɛ foro kɔnɔ. Bɔ u.",
+                            "Pests are in the field. Drive them out."),
+    "pest_ok":             ("Dɔgɔw tɛ yen. Foro ka ɲi.",
+                            "No pests detected. The field looks healthy."),
+    # Irrigation
+    "irrigation_needed":   ("Foro fɛ ji. Ji sira yɔrɔ.",
+                            "The field needs water. Open the irrigation."),
+    "irrigation_active":   ("Ji bɛ taa. A bɛ kɛ cogo di.",
+                            "Irrigation is running. It is working well."),
+    # Default
+    "default":             ("Kabako jumanw sɔrɔla.",
+                            "Sensor data received. No alerts at this time."),
 }
 
 # ── Fula templates (≤6 words per sentence for clear MMS-TTS output) ──────────
-FULA_TEMPLATES = {
-    "soil_moisture_low":   "Leydi ndiyam famɗi. Wado ngesa.",
-    "soil_moisture_high":  "Ndiyam heewi. Leydi famɗaali.",
-    "soil_ph_low":         "Leydi suurii. Waɗ kalisi.",
-    "soil_ph_high":        "Leydi alkalii. Waɗ soufre.",
-    "weather_hot":         "Nguleeki heewi. Muusal.",
-    "rain_likely":         "Ndiyam wadata. Loosu ngesa.",
-    "pest_high":           "Biñ-biñ ngesa nder. Fiil ɗen.",
-    "irrigation_needed":   "Ngesa fɛɗɛli ndiyam. Wado.",
-    "irrigation_active":   "Ndiyam wona jooni.",
-    "default":             "Humpito juuti waɗaama.",
+# Format: (fula_text, english_translation)
+FULA_TEMPLATES: dict[str, tuple[str, str]] = {
+    # Greetings / social
+    "greeting":            ("Jam waali. Mi woni ɗoo.",
+                            "Hello. I am here."),
+    "greeting_morning":    ("Jam waali. Yoo barke.",
+                            "Good morning. May there be blessings."),
+    "greeting_evening":    ("Jam hiiri. Jam waɗaa.",
+                            "Good evening. Peace be with you."),
+    "thanks":              ("Jaraama. A weli.",
+                            "Thank you. That is good."),
+    "farewell":            ("Yahdu jam. Mi anndii.",
+                            "Go in peace. I will remember you."),
+    "not_understood":      ("Mi faamii wanaa. Hol ɗoo.",
+                            "I did not understand. Please say again."),
+    # Soil
+    "soil_moisture_low":   ("Leydi ndiyam famɗi. Wado ngesa.",
+                            "Soil moisture is low. Water the field."),
+    "soil_moisture_high":  ("Ndiyam heewi. Leydi famɗaali.",
+                            "Too much water. Drainage is needed."),
+    "soil_ph_low":         ("Leydi suurii. Waɗ kalisi.",
+                            "Soil is acidic. Add lime."),
+    "soil_ph_high":        ("Leydi alkalii. Waɗ soufre.",
+                            "Soil is alkaline. Add sulphur."),
+    "soil_ok":             ("Leydi weli. Ngesa yoodi.",
+                            "Soil is good. The field is ready."),
+    # Weather
+    "weather_hot":         ("Nguleeki heewi. Muusal.",
+                            "It is very hot. Rest during midday."),
+    "rain_likely":         ("Ndiyam wadata. Loosu ngesa.",
+                            "Rain is coming. Protect the harvest."),
+    "weather_ok":          ("Jawdi weli. Waɗ golle.",
+                            "Weather is fine. Go work."),
+    # Pest
+    "pest_high":           ("Biñ-biñ ngesa nder. Fiil ɗen.",
+                            "Pests are in the field. Remove them."),
+    "pest_ok":             ("Biñ-biñ alaa. Ngesa weli.",
+                            "No pests found. Field looks healthy."),
+    # Irrigation
+    "irrigation_needed":   ("Ngesa fɛɗɛli ndiyam. Wado.",
+                            "Field needs water. Start irrigation."),
+    "irrigation_active":   ("Ndiyam wona jooni.",
+                            "Irrigation is running now."),
+    # Default
+    "default":             ("Humpito juuti waɗaama.",
+                            "Sensor data received. No alerts."),
 }
 
 
 class VoiceResponder:
-    """Converts sensor readings into actionable voice messages in the farmer's language."""
+    """Converts sensor readings into actionable voice messages in the farmer's language.
+
+    generate_response() returns (native_text, english_translation).
+    For French/English, english_translation == native_text.
+    """
 
     def __init__(self, language: str = "fr") -> None:
         self.language = language
 
-    def generate_response(self, intent: "Intent", sensor_data: "SensorData") -> str:
+    def generate_response(self, intent: "Intent", sensor_data: "SensorData") -> tuple[str, str]:
+        """Return (response_in_native_language, english_translation)."""
         if self.language == "bam":
-            return self._bambara_response(sensor_data)
+            return self._bambara_response(intent, sensor_data)
         elif self.language == "ful":
-            return self._fula_response(sensor_data)
+            return self._fula_response(intent, sensor_data)
         else:
-            return self._french_response(sensor_data)
+            text = self._french_response(sensor_data)
+            return text, text
 
     # ── Bambara ──────────────────────────────────────────────────────────────
 
-    def _bambara_response(self, sensor_data: "SensorData") -> str:
+    def _bambara_response(self, intent: "Intent", sensor_data: "SensorData") -> tuple[str, str]:
         t = sensor_data.sensor_type
         v = sensor_data.values
         T = BAMBARA_TEMPLATES
 
+        # Greeting intents — checked first regardless of sensor type
+        if intent.action == "greeting":
+            key = "greeting_morning" if v.get("is_morning") else "greeting"
+            return T[key]
+        if intent.action == "thanks":
+            return T["thanks"]
+        if intent.action == "farewell":
+            return T["farewell"]
+
         if t == "soil":
             moisture = v.get("moisture_pct")
             if moisture is not None:
@@ -82,6 +168,7 @@ class VoiceResponder:
                     return T["soil_ph_low"]
                 elif ph > SOIL_PH_HIGH:
                     return T["soil_ph_high"]
+            return T["soil_ok"]
 
         elif t == "weather":
             temp = v.get("temperature_c")
@@ -90,10 +177,11 @@ class VoiceResponder:
                 return T["weather_hot"]
             if rain is not None and rain > 70:
                 return T["rain_likely"]
+            return T["weather_ok"]
 
         elif t == "irrigation":
-            last = v.get("last_irrigation_h_ago")
             active = v.get("active")
+            last = v.get("last_irrigation_h_ago")
             if active:
                 return T["irrigation_active"]
             if last is not None and last > 24:
@@ -103,16 +191,25 @@ class VoiceResponder:
             level = int(v.get("alert_level", 0))
             if level >= PEST_ALERT_HIGH:
                 return T["pest_high"]
+            return T["pest_ok"]
 
         return T["default"]
 
     # ── Fula ─────────────────────────────────────────────────────────────────
 
-    def _fula_response(self, sensor_data: "SensorData") -> str:
+    def _fula_response(self, intent: "Intent", sensor_data: "SensorData") -> tuple[str, str]:
         t = sensor_data.sensor_type
         v = sensor_data.values
         T = FULA_TEMPLATES
 
+        # Greeting intents — checked first regardless of sensor type
+        if intent.action == "greeting":
+            return T["greeting"]
+        if intent.action == "thanks":
+            return T["thanks"]
+        if intent.action == "farewell":
+            return T["farewell"]
+
         if t == "soil":
             moisture = v.get("moisture_pct")
             if moisture is not None:
@@ -126,6 +223,7 @@ class VoiceResponder:
                     return T["soil_ph_low"]
                 elif ph > SOIL_PH_HIGH:
                     return T["soil_ph_high"]
+            return T["soil_ok"]
 
         elif t == "weather":
             temp = v.get("temperature_c")
@@ -134,6 +232,7 @@ class VoiceResponder:
                 return T["weather_hot"]
             if rain is not None and rain > 70:
                 return T["rain_likely"]
+            return T["weather_ok"]
 
         elif t == "irrigation":
             active = v.get("active")
@@ -147,6 +246,7 @@ class VoiceResponder:
             level = int(v.get("alert_level", 0))
             if level >= PEST_ALERT_HIGH:
                 return T["pest_high"]
+            return T["pest_ok"]
 
         return T["default"]