Add Deepgram Streaming Diarization

config/benchmark_config/datasets/callhome_english.yaml

@@ -0,0 +1,6 @@
+callhome_english:
+  dataset_id: argmaxinc/callhome-english
+  split: test
+
+
+

config/benchmark_config/metrics/cpwer.yaml

@@ -0,0 +1,3 @@
+# Dummy argument to make the config file valid
+cpwer:
+  skip_overlap: false

config/pipeline_configs/DeepgramStreamingOrchestrationPipeline.yaml

@@ -0,0 +1,9 @@
+DeepgramStreamingOrchestrationPipeline:
+  pipeline_config:
+    sample_rate: 16000
+    channels: 1
+    sample_width: 2
+    realtime_resolution: 0.020
+    model_version: "nova-3"
+    enable_diarization: true
+

src/openbench/metric/word_error_metrics/word_error_metrics.py

@@ -297,7 +297,10 @@ def compute_metric(self, detail: Details) -> float:
         return (S + D + I) / N if N > 0 else 0.0
 
 
-@MetricRegistry.register_metric(PipelineType.ORCHESTRATION, MetricOptions.CPWER)
+@MetricRegistry.register_metric(
+    PipelineType.ORCHESTRATION,
+    MetricOptions.CPWER,
+)
 class ConcatenatedMinimumPermutationWER(BaseWordErrorMetric):
     """Concatenated minimum-Permutation Word Error Rate (cpWER) implementation.
 

src/openbench/pipeline/orchestration/__init__.py

@@ -3,6 +3,10 @@
 
 from .nemo import NeMoMTParakeetPipeline, NeMoMTParakeetPipelineConfig
 from .orchestration_deepgram import DeepgramOrchestrationPipeline, DeepgramOrchestrationPipelineConfig
+from .orchestration_deepgram_streaming import (
+    DeepgramStreamingOrchestrationPipeline,
+    DeepgramStreamingOrchestrationPipelineConfig,
+)
 from .orchestration_openai import OpenAIOrchestrationPipeline, OpenAIOrchestrationPipelineConfig
 from .orchestration_whisperkitpro import WhisperKitProOrchestrationConfig, WhisperKitProOrchestrationPipeline
 from .whisperx import WhisperXPipeline, WhisperXPipelineConfig
@@ -11,6 +15,8 @@
 __all__ = [
     "DeepgramOrchestrationPipeline",
     "DeepgramOrchestrationPipelineConfig",
+    "DeepgramStreamingOrchestrationPipeline",
+    "DeepgramStreamingOrchestrationPipelineConfig",
     "WhisperXPipeline",
     "WhisperXPipelineConfig",
     "WhisperKitProOrchestrationPipeline",

src/openbench/pipeline/orchestration/orchestration_deepgram_streaming.py

@@ -0,0 +1,86 @@
+# For licensing see accompanying LICENSE.md file.
+# Copyright (C) 2025 Argmax, Inc. All Rights Reserved.
+
+import numpy as np
+from pydantic import Field
+
+from ...dataset import OrchestrationSample
+from ...pipeline import Pipeline, PipelineConfig, register_pipeline
+from ...pipeline_prediction import Transcript, Word
+from ...types import PipelineType
+from ..streaming_transcription.deepgram import DeepgramApi
+from .common import OrchestrationOutput
+
+
+class DeepgramStreamingOrchestrationPipelineConfig(PipelineConfig):
+    sample_rate: int = Field(default=16000, description="Sample rate of the audio")
+    channels: int = Field(default=1, description="Number of audio channels")
+    sample_width: int = Field(default=2, description="Sample width in bytes")
+    realtime_resolution: float = Field(default=0.020, description="Real-time resolution for streaming")
+    model_version: str = Field(
+        default="nova-3", description=("The model to use for real-time transcription with diarization")
+    )
+    enable_diarization: bool = Field(default=True, description="Whether to enable speaker diarization")
+
+
+@register_pipeline
+class DeepgramStreamingOrchestrationPipeline(Pipeline):
+    _config_class = DeepgramStreamingOrchestrationPipelineConfig
+    pipeline_type = PipelineType.ORCHESTRATION
+
+    def build_pipeline(self):
+        """Build Deepgram streaming API with diarization enabled."""
+        # Create a modified config for the streaming API
+        from types import SimpleNamespace
+
+        api_config = SimpleNamespace(
+            channels=self.config.channels,
+            sample_width=self.config.sample_width,
+            sample_rate=self.config.sample_rate,
+            realtime_resolution=self.config.realtime_resolution,
+            model_version=self.config.model_version,
+            enable_diarization=self.config.enable_diarization,
+        )
+
+        pipeline = DeepgramApi(api_config)
+        return pipeline
+
+    def parse_input(self, input_sample: OrchestrationSample):
+        """Convert audio waveform to bytes for streaming."""
+        y = input_sample.waveform
+        y_int16 = (y * 32767).astype(np.int16)
+        audio_data_byte = y_int16.T.tobytes()
+        return audio_data_byte
+
+    def parse_output(self, output) -> OrchestrationOutput:
+        """Parse output to extract transcription and diarization."""
+        # Extract words with speaker info if diarization enabled
+        words = []
+
+        if "words_with_speakers" in output and output["words_with_speakers"]:
+            # This comes from diarization-enabled streaming
+            for word_info in output["words_with_speakers"]:
+                words.append(
+                    Word(
+                        word=word_info.get("word", ""),
+                        start=word_info.get("start"),
+                        end=word_info.get("end"),
+                        speaker=word_info.get("speaker"),
+                    )
+                )
+        else:
+            # Speaker labels are required for orchestration pipelines
+            raise ValueError(
+                "No speaker diarization data available. "
+                "Orchestration pipelines require speaker labels. "
+                "Ensure 'enable_diarization' is set to True in the pipeline config."
+            )
+
+        # Create final transcript with speaker-attributed words
+        transcript = Transcript(words=words)
+
+        return OrchestrationOutput(
+            prediction=transcript,
+            transcription_output=None,
+            diarization_output=None,
+        )

src/openbench/pipeline/pipeline_aliases.py

@@ -16,6 +16,7 @@
 )
 from .orchestration import (
     DeepgramOrchestrationPipeline,
+    DeepgramStreamingOrchestrationPipeline,
     NeMoMTParakeetPipeline,
     OpenAIOrchestrationPipeline,
     WhisperKitProOrchestrationPipeline,
@@ -171,6 +172,20 @@ def register_pipeline_aliases() -> None:
         description="Deepgram orchestration pipeline. Requires API key from https://www.deepgram.com/. Set `DEEPGRAM_API_KEY` env var.",
     )
 
+    PipelineRegistry.register_alias(
+        "deepgram-streaming-orchestration",
+        DeepgramStreamingOrchestrationPipeline,
+        default_config={
+            "sample_rate": 16000,
+            "channels": 1,
+            "sample_width": 2,
+            "realtime_resolution": 0.020,
+            "model_version": "nova-3",
+            "enable_diarization": True,
+        },
+        description=("Deepgram streaming orchestration pipeline with diarization enabled."),
+    )
+
     PipelineRegistry.register_alias(
         "whisperkitpro-orchestration-tiny",
         WhisperKitProOrchestrationPipeline,

src/openbench/pipeline/streaming_transcription/deepgram.py

@@ -26,18 +26,20 @@
 
 class DeepgramApi:
     def __init__(self, cfg) -> None:
-        self.realtime_resolution = 0.020
-        self.model_version = "nova-3"
+        self.realtime_resolution = getattr(cfg, "realtime_resolution", 0.020)
+        self.model_version = getattr(cfg, "model_version", "nova-3")
         self.api_key = os.getenv("DEEPGRAM_API_KEY")
         assert self.api_key is not None, "Please set API key in environment"
         self.channels = cfg.channels
         self.sample_width = cfg.sample_width
         self.sample_rate = cfg.sample_rate
         self.host_url = os.getenv("DEEPGRAM_HOST_URL", "wss://api.deepgram.com")
+        self.enable_diarization = getattr(cfg, "enable_diarization", False)
 
     async def run(self, data, key, channels, sample_width, sample_rate):
-        """Connect to the Deepgram real-time streaming endpoint, stream the data
-        in real-time, and print out the responses from the server.
+        """Connect to Deepgram real-time streaming endpoint.
+
+        Stream the data in real-time and print responses from server.
 
         This uses a pre-recorded file as an example. It mimics a real-time
         connection by sending `REALTIME_RESOLUTION` seconds of audio every
@@ -62,9 +64,23 @@ async def run(self, data, key, channels, sample_width, sample_rate):
         confirmed_interim_transcripts = []
         model_timestamps_hypothesis = []
         model_timestamps_confirmed = []
-        # Connect to the real-time streaming endpoint, attaching our API key.
+        words_with_speakers = []
+
+        # Build connection URL with optional diarization
+        url = (
+            f"{self.host_url}/v1/listen?"
+            f"model={self.model_version}&"
+            f"channels={channels}&"
+            f"sample_rate={sample_rate}&"
+            f"encoding=linear16&"
+            f"interim_results=true"
+        )
+        if self.enable_diarization:
+            url += "&diarize=true"
+
+        # Connect to the real-time streaming endpoint
         async with websockets.connect(
-            f"{self.host_url}/v1/listen?model={self.model_version}&channels={channels}&sample_rate={sample_rate}&encoding=linear16&interim_results=true",
+            url,
             additional_headers={
                 "Authorization": "Token {}".format(key),
             },
@@ -91,8 +107,8 @@ async def sender(ws):
                 await ws.send(json.dumps({"type": "CloseStream"}))
 
             async def receiver(ws):
-                """Print out the messages received from the server."""
-                nonlocal audio_cursor
+                """Print out messages received from the server."""
+                nonlocal audio_cursor, words_with_speakers
                 global transcript
                 global interim_transcripts
                 global audio_cursor_l
@@ -105,26 +121,37 @@ async def receiver(ws):
                 async for msg in ws:
                     msg = json.loads(msg)
                     if "request_id" in msg:
-                        # This is the final metadata message. It gets sent as the
-                        # very last message by Deepgram during a clean shutdown.
+                        # This is the final metadata message.
                         # There is no transcript in it.
                         continue
-                    if msg["channel"]["alternatives"][0]["transcript"] != "":
+                    alternatives = msg["channel"]["alternatives"][0]
+                    if alternatives["transcript"] != "":
                         if not msg["is_final"]:
                             audio_cursor_l.append(audio_cursor)
-                            model_timestamps_hypothesis.append(msg["channel"]["alternatives"][0]["words"])
-                            interim_transcripts.append(
-                                transcript + " " + msg["channel"]["alternatives"][0]["transcript"]
-                            )
-                            logger.debug(
-                                "\n" + "Transcription: " + transcript + msg["channel"]["alternatives"][0]["transcript"]
-                            )
+                            model_timestamps_hypothesis.append(alternatives["words"])
+                            interim_transcripts.append(transcript + " " + alternatives["transcript"])
+                            logger.debug("\n" + "Transcription: " + transcript + alternatives["transcript"])
 
                         elif msg["is_final"]:
                             confirmed_audio_cursor_l.append(audio_cursor)
-                            transcript = transcript + " " + msg["channel"]["alternatives"][0]["transcript"]
+                            transcript = transcript + " " + alternatives["transcript"]
                             confirmed_interim_transcripts.append(transcript)
-                            model_timestamps_confirmed.append(msg["channel"]["alternatives"][0]["words"])
+                            words = alternatives["words"]
+                            model_timestamps_confirmed.append(words)
+
+                            # Collect speaker info if diarization enabled
+                            if self.enable_diarization:
+                                for word_info in words:
+                                    if "speaker" in word_info:
+                                        speaker_label = f"SPEAKER_{word_info['speaker']}"
+                                        words_with_speakers.append(
+                                            {
+                                                "word": word_info.get("word", ""),
+                                                "speaker": speaker_label,
+                                                "start": word_info.get("start", 0),
+                                                "end": word_info.get("end", 0),
+                                            }
+                                        )
 
             await asyncio.gather(sender(ws), receiver(ws))
             return (
@@ -135,6 +162,7 @@ async def receiver(ws):
                 confirmed_audio_cursor_l,
                 model_timestamps_hypothesis,
                 model_timestamps_confirmed,
+                words_with_speakers,
             )
 
     def __call__(self, sample):
@@ -147,17 +175,19 @@ def __call__(self, sample):
             confirmed_audio_cursor_l,
             model_timestamps_hypothesis,
             model_timestamps_confirmed,
+            words_with_speakers,
         ) = asyncio.get_event_loop().run_until_complete(
             self.run(sample, self.api_key, self.channels, self.sample_width, self.sample_rate)
         )
         return {
             "transcript": transcript,
             "interim_transcripts": interim_transcripts,
             "audio_cursor": audio_cursor_l,
-            "confirmed_interim_transcripts": confirmed_interim_transcripts,
+            "confirmed_interim_transcripts": (confirmed_interim_transcripts),
             "confirmed_audio_cursor": confirmed_audio_cursor_l,
-            "model_timestamps_hypothesis": model_timestamps_hypothesis,
-            "model_timestamps_confirmed": model_timestamps_confirmed,
+            "model_timestamps_hypothesis": (model_timestamps_hypothesis),
+            "model_timestamps_confirmed": (model_timestamps_confirmed),
+            "words_with_speakers": words_with_speakers,
         }