Add ElevenLabs dialogue generation pipeline with supporting changes

src/openbench/dataset/dataset_aliases.py

@@ -660,6 +660,7 @@ def register_dataset_aliases() -> None:
                 PipelineType.DIARIZATION,
                 PipelineType.STREAMING_TRANSCRIPTION,
                 PipelineType.ORCHESTRATION,
+                PipelineType.SPEECH_GENERATION,
             },
             description="Local dataset for testing. To use this dataset you need to set the `LOCAL_DATASET_PATH` and `LOCAL_DATASET_SPLIT` environment variables.",
         )

src/openbench/dataset/dataset_speech_generation.py

@@ -13,6 +13,7 @@ class SpeechGenerationExtraInfo(TypedDict, total=False):
     """Extra info for speech generation samples."""
 
     language: str
+    dialogue: list[dict]
 
 
 class SpeechGenerationRow(TypedDict):
@@ -95,5 +96,7 @@ def prepare_sample(self, row: SpeechGenerationRow) -> tuple[Transcript, SpeechGe
         extra_info: SpeechGenerationExtraInfo = {}
         if "language" in row:
             extra_info["language"] = row["language"]
+        if "dialogue" in row and row["dialogue"]:
+            extra_info["dialogue"] = row["dialogue"]
 
         return reference, extra_info

src/openbench/metric/keyword_boosting_metrics/boosting_metrics.py

@@ -26,6 +26,9 @@ def compute_keyword_stats(
     ) -> dict[str, Any]:
         """Compute keyword statistics between reference and hypothesis."""
 
+        if not dictionary:
+            return {"true_positives": 0, "ground_truth": 0, "false_positives": 0, "keyword_stats": {}}
+
         # Convert transcripts to text
         ref_text = reference.get_transcript_string()
         hyp_text = hypothesis.get_transcript_string()

src/openbench/pipeline/pipeline_aliases.py

@@ -26,6 +26,7 @@
 )
 from .pipeline_registry import PipelineRegistry
 from .speech_generation import (
+    ElevenLabsDialogueGenerationPipeline,
     WhisperKitSpeechGenerationPipeline,
 )
 from .streaming_transcription import (
@@ -668,6 +669,31 @@ def register_pipeline_aliases() -> None:
         "Requires `WHISPERKIT_CLI_PATH` env var pointing to the whisperkit-cli binary.",
     )
 
+    PipelineRegistry.register_alias(
+        "elevenlabs-dialogue-generation",
+        ElevenLabsDialogueGenerationPipeline,
+        default_config={
+            "out_dir": "./speech_generation_results",
+            "model_id": "eleven_v3",
+            "speaker_voice_map": {
+                "doctor": "9BWtsMINqrJLrRacOk9x",
+                "patient": "IKne3meq5aSn9XLyUdCD",
+                "assistant": "pFZP5JQG7iQjIQuC4Bku",
+            },
+            "default_voice_id": "9BWtsMINqrJLrRacOk9x",
+            "max_chars_per_chunk": 4500,
+            "chunk_silence_duration": 0.75,
+            "transcription_cli_path": os.getenv("WHISPERKITPRO_CLI_PATH"),
+            "transcription_repo_id": "argmaxinc/parakeetkit-pro",
+            "transcription_model_variant": "nvidia_parakeet-v2_476MB",
+            "keep_generated_audio": False,
+        },
+        description="ElevenLabs dialogue generation pipeline. Generates multi-speaker conversational audio "
+        "from dialogue turns using ElevenLabs text_to_dialogue API, then transcribes the generated "
+        "audio to compute WER against the original dialogue text. "
+        "Requires `ELEVENLABS_API_KEY` and `WHISPERKITPRO_CLI_PATH` env vars.",
+    )
+
     ################# STREAMING TRANSCRIPTION PIPELINES #################
 
     PipelineRegistry.register_alias(

src/openbench/pipeline/speech_generation/__init__.py

@@ -2,10 +2,16 @@
 # Copyright (C) 2025 Argmax, Inc. All Rights Reserved.
 
 from .common import SpeechGenerationConfig, SpeechGenerationOutput
+from .speech_generation_elevenlabs_dialogue import (
+    ElevenLabsDialogueGenerationConfig,
+    ElevenLabsDialogueGenerationPipeline,
+)
 from .speech_generation_wkp import WhisperKitSpeechGenerationPipeline
 
 
 __all__ = [
+    "ElevenLabsDialogueGenerationConfig",
+    "ElevenLabsDialogueGenerationPipeline",
     "SpeechGenerationConfig",
     "SpeechGenerationOutput",
     "WhisperKitSpeechGenerationPipeline",