@@ -199,6 +199,26 @@ def normalize_bcp47_language(value: str) -> str:
 DEFAULT_ELEVENLABS_STT_MODEL = "scribe_v2"
 DEFAULT_ELEVENLABS_TTS_MODEL = "eleven_v3"
 
+# Kaapi BCP-47 tag -> Gemini TTS language. Full locales, not bare ISO codes:
+# en-IN keeps the Indian accent where "en" does not. Unlisted tags fall back
+# to auto-detect.
+BCP47_LOCALE_TO_GEMINI_LANG: dict[str, str] = {
+    "en-IN": "en-IN",
+    "hi-IN": "hi-IN",
+    "bn-IN": "bn-IN",
+    "ta-IN": "ta-IN",
+    "te-IN": "te-IN",
+    "mr-IN": "mr-IN",
+    "gu-IN": "gu-IN",
+    "kn-IN": "kn-IN",
+    "ml-IN": "ml-IN",
+    "pa-IN": "pa-IN",
+    # Kaapi uses the non-standard "od-IN" for Odia; the real tag is "or-IN".
+    "od-IN": "or-IN",
+    # "as-IN": not supported by Gemini
+    "sd-IN": "sd-IN",
+}
+
 # BCP-47 language tag → ElevenLabs ISO 639-1 code (Indic + English)
 BCP47_TO_ELEVENLABS_LANG: dict[str, str] = {
     "en-IN": "en",
 
@@ -8,6 +8,7 @@
 from app.models.llm import KaapiCompletionConfig, NativeCompletionConfig
 from app.models.llm.request import STTLLMParams, TextLLMParams, TTSLLMParams
 from app.models.llm.constants import (
+    BCP47_LOCALE_TO_GEMINI_LANG,
     BCP47_TO_ELEVENLABS_LANG,
     DEFAULT_ELEVENLABS_STT_MODEL,
     DEFAULT_ELEVENLABS_TTS_MODEL,
@@ -268,7 +269,7 @@ def map_kaapi_to_google_params(
         - instructions → instructions (for STT prompts, if available)
         - temperature / top_p → not sent (deprecated by Gemini)
         - max_output_tokens → same name (text only)
-        - language → full BCP-47 locale passed through (TTS only)
+        - language → full Gemini locale via BCP47_LOCALE_TO_GEMINI_LANG; unsupported → auto-detect (TTS only)
         - thinking_level → thinking_config.thinking_level (text only)
         - output_schema → output_schema, converted to Gemini's shape (text only)
         - knowledge_base_ids → FileSearch tool store names (text only)
@@ -342,10 +343,16 @@ def map_kaapi_to_google_params(
         response_format = params.get("response_format") or "wav"
         google_params["response_format"] = response_format
 
-        # Full BCP-47 locale, not the bare ISO code: collapsing en-IN to en loses the accent.
+        # Language: Only set if explicitly provided (None/missing = auto-detect)
         language = params.get("language")
         if language:
-            google_params["language"] = language
+            google_lang = BCP47_LOCALE_TO_GEMINI_LANG.get(language)
+            if not google_lang:
+                warnings.append(
+                    f"Unsupported language '{language}' for Gemini TTS, using auto-detect"
+                )
+            else:
+                google_params["language"] = google_lang
 
     elif completion_type == CompletionType.STT:
         # STT mode - instructions, input_language, output_language, response_format
 
@@ -408,11 +408,12 @@ def test_tts_language_missing_uses_auto_detect(self):
         assert "language" not in result  # Not set = auto-detect
         assert warnings == []
 
-    def test_tts_non_indic_language_passes_through(self):
+    @pytest.mark.parametrize("language", ["fr-FR", "as-IN"])
+    def test_tts_unsupported_language_falls_back_to_auto_detect(self, language):
         kaapi_params = TTSLLMParams(
             model="gemini-2.5-pro",
             voice="Kore",
-            language="fr-FR",
+            language=language,
         )
 
         result, warnings = map_kaapi_to_google_params(
@@ -422,7 +423,22 @@ def test_tts_non_indic_language_passes_through(self):
         assert result["model"] == "gemini-2.5-pro"
         assert result["voice"] == "Kore"
         assert result["response_format"] == "wav"
-        assert result["language"] == "fr-FR"
+        assert "language" not in result
+        assert warnings == [
+            f"Unsupported language '{language}' for Gemini TTS, using auto-detect"
+        ]
+
+    def test_tts_odia_maps_to_standard_locale(self):
+        # Kaapi's canonical Odia tag is the non-standard od-IN.
+        kaapi_params = TTSLLMParams(
+            model="gemini-2.5-pro", voice="Kore", language="odia"
+        )
+
+        result, warnings = map_kaapi_to_google_params(
+            kaapi_params.model_dump(exclude_none=True), completion_type="tts"
+        )
+
+        assert result["language"] == "or-IN"
         assert warnings == []
 
     def test_tts_without_language_leaves_it_unset(self):
@@ -1198,7 +1214,7 @@ def test_text_completion_maps_via_google_mapper(self, db: Session):
         assert native_config.provider == "google-native"
         assert native_config.params["model"] == "gemini-2.5-pro"
 
-    def test_unlisted_language_passes_through_without_warning(self, db: Session):
+    def test_unlisted_language_falls_back_to_auto_detect(self, db: Session):
         kaapi_config = build_kaapi_completion_config(
             provider="google",
             type="tts",
@@ -1213,8 +1229,10 @@ def test_unlisted_language_passes_through_without_warning(self, db: Session):
         )
 
         assert native_config.provider == "google-native"
-        assert native_config.params["language"] == "xx-YY"
-        assert warnings == []
+        assert "language" not in native_config.params
+        assert warnings == [
+            "Unsupported language 'xx-YY' for Gemini TTS, using auto-detect"
+        ]
 
 
 class TestTransformGoogleGCPRouting:
 
@@ -34,7 +34,7 @@ All paths relative to `backend/app/`.
 
 ## Services / CRUD
 - `services/llm/` — `mappers.py` (Kaapi params → provider API), `providers/`, `chain/`, `guardrails.py`, `jobs.py`
-- `mappers.py` is the **only** param mapper in the codebase; the assessment fork was merged back into it. Callers: `crud/evaluations/{batch,fast,judge}.py`, `crud/assessment/batch.py`, `services/assessment/api/batch.py`, `services/assessment/prefilter/request_builder.py`. Structured output is keyed `output_schema` for every provider (OpenAI `text.format`, Anthropic `output_config.format`, Gemini `output_schema`), and Anthropic reads `effort` into `output_config.effort`. Every param that only assessment set (`top_p`, `max_output_tokens`, `thinking_level`, `thinking`, `output_schema`) defaults to `None` on `TextLLMParams` and is dropped by `ParamSerialization._dump_compact`, so a config that does not set it maps exactly as before. Anthropic never receives `temperature`/`top_p`: the Messages API returns 400 for a non-default sampling value on every Claude model Kaapi serves, so the mapper drops both, warning only when the value was not Anthropic's own default of 1.0. Gemini likewise never receives `temperature`/`top_p` for text or STT: newer models return 400 for them and they have been no-ops since 3.6 Flash, so the mapper does not send them. Gemini TTS `language` is passed through as the full BCP-47 locale (`en-IN`, not `en`) so the accent survives. `normalize_llm_text` no longer lives here; it moved to `services/assessment/validators.py`, the only place that calls it.
+- `mappers.py` is the **only** param mapper in the codebase; the assessment fork was merged back into it. Callers: `crud/evaluations/{batch,fast,judge}.py`, `crud/assessment/batch.py`, `services/assessment/api/batch.py`, `services/assessment/prefilter/request_builder.py`. Structured output is keyed `output_schema` for every provider (OpenAI `text.format`, Anthropic `output_config.format`, Gemini `output_schema`), and Anthropic reads `effort` into `output_config.effort`. Every param that only assessment set (`top_p`, `max_output_tokens`, `thinking_level`, `thinking`, `output_schema`) defaults to `None` on `TextLLMParams` and is dropped by `ParamSerialization._dump_compact`, so a config that does not set it maps exactly as before. Anthropic never receives `temperature`/`top_p`: the Messages API returns 400 for a non-default sampling value on every Claude model Kaapi serves, so the mapper drops both, warning only when the value was not Anthropic's own default of 1.0. Gemini likewise never receives `temperature`/`top_p` for text or STT: newer models return 400 for them and they have been no-ops since 3.6 Flash, so the mapper does not send them. Gemini TTS `language` maps through `BCP47_LOCALE_TO_GEMINI_LANG` to a full locale (`en-IN`, not `en`, so the accent survives; Kaapi's `od-IN` → `or-IN`); an unlisted tag (e.g. `as-IN`) is dropped with a warning and Gemini auto-detects. `normalize_llm_text` no longer lives here; it moved to `services/assessment/validators.py`, the only place that calls it.
 - `services/guardrails/` — validator execution
 - `crud/llm.py`, `crud/llm_chain.py`, `crud/config/` — persistence