feat(evaluation): Add Sarvam and ElevenLabs TTS

backend/app/api/docs/tts_evaluation/start_evaluation.md

@@ -9,7 +9,10 @@ Optional fields:
 
 The evaluation will:
 1. Process each text sample through the specified TTS models
-2. Generate speech audio using Gemini Batch API
-3. Store WAV audio files in S3 for human review
+2. Generate speech audio: Gemini models via the Gemini Batch API; Sarvam and ElevenLabs models via direct API calls in background workers
+3. Store 24 kHz mono WAV audio files in S3 for human review
 
-**Supported models:** `gemini-2.5-pro-preview-tts`
+Synthesis uses the dataset's language; datasets without a language are synthesized as `en-IN`.
+The response's `run_metadata.voices` lists the voice used per model.
+
+**Supported models:** `gemini-2.5-pro-preview-tts`, `gemini-3.1-flash-tts-preview`, `bulbul:v3` (Sarvam), `eleven_v3`, `eleven_v4` (ElevenLabs)

backend/app/api/routes/tts_evaluations/evaluation.py

@@ -25,6 +25,10 @@
 from app.services.tts_evaluations.constants import (
     DEFAULT_STYLE_PROMPT,
     DEFAULT_VOICE_NAME,
+    ELEVENLABS_TTS_MODELS,
+    ELEVENLABS_VOICE_ID,
+    SARVAM_SPEAKER,
+    SARVAM_TTS_MODELS,
 )
 from app.utils import APIResponse, load_description
 
@@ -114,11 +118,22 @@ def start_tts_evaluation(
             detail=f"Failed to queue batch submission: {e}",
         )
 
+    voices: dict[str, str] = {}
+    for model in run_create.models:
+        if model in SARVAM_TTS_MODELS:
+            voices[model] = SARVAM_SPEAKER
+        elif model in ELEVENLABS_TTS_MODELS:
+            voices[model] = ELEVENLABS_VOICE_ID
+        else:
+            voices[model] = DEFAULT_VOICE_NAME
+
     return APIResponse.success_response(
         data=TTSEvaluationRunPublic.from_model(
             run,
             run_metadata={
+                # Legacy key kept for existing clients; `voices` is per model.
                 "voice_name": DEFAULT_VOICE_NAME,
+                "voices": voices,
                 "style_prompt": DEFAULT_STYLE_PROMPT,
             },
         )

backend/app/celery/tasks/job_execution.py

@@ -429,6 +429,28 @@ def run_tts_result_processing(
     )
 
 
+@celery_app.task(bind=True, queue="default", priority=2)
+@gevent_timeout(settings.CELERY_TASK_SOFT_TIME_LIMIT, "run_tts_sync_generation")
+def run_tts_sync_generation(
+    self, project_id: int, job_id: str, trace_id: str, **kwargs
+):
+    from app.services.tts_evaluations.sync_generation import (
+        execute_tts_sync_generation,
+    )
+
+    _set_trace(trace_id)
+    return _run_with_otel_parent(
+        self,
+        lambda: execute_tts_sync_generation(
+            project_id=project_id,
+            job_id=job_id,
+            task_id=current_task.request.id,
+            task_instance=self,
+            **kwargs,
+        ),
+    )
+
+
 @celery_app.task(bind=True, queue="default", priority=6)
 @gevent_timeout(settings.CELERY_TASK_SOFT_TIME_LIMIT, "run_evaluation_fast_chunk")
 def run_evaluation_fast_chunk(

backend/app/celery/utils.py

@@ -270,6 +270,24 @@ def start_tts_result_processing(
     return task_id
 
 
+def start_tts_sync_generation(
+    project_id: int, job_id: str, trace_id: str = "N/A", **kwargs: Any
+) -> str:
+    from app.celery.tasks.job_execution import run_tts_sync_generation
+
+    task_id = _enqueue_with_trace_context(
+        run_tts_sync_generation,
+        project_id=project_id,
+        job_id=job_id,
+        trace_id=trace_id,
+        **kwargs,
+    )
+    logger.info(
+        f"[start_tts_sync_generation] Started job {job_id} with Celery task {task_id}"
+    )
+    return task_id
+
+
 def start_fast_evaluation_chunk(
     eval_run_id: int, chunk_index: int, trace_id: str = "N/A"
 ) -> str:

backend/app/core/batch/operations.py

@@ -49,6 +49,11 @@ def start_batch_job(
 
     batch_job = create_batch_job(session=session, batch_job_create=batch_job_create)
 
+    # create_batch_job's refresh() re-opens a transaction; end it so no pooled
+    # connection sits idle-in-transaction during the (slow) upload + create.
+    # Nothing is pending at this point, so this only releases the connection.
+    session.commit()
+
     try:
         batch_result = provider.create_batch(jsonl_data=jsonl_data, config=config)