Support for Non-Streaming Mode with Language Detection in Deepgram STT

[ringofhealth]

Problem Description

The current Deepgram STT implementation in LiveKit Agents Python SDK has a limitation where language detection (detect_language=True) cannot be used with the standard STT class because:

1. The Deepgram STT class always advertises streaming capabilities in its constructor
2. When detect_language=True is set, the stream() method raises an error stating that language detection is not supported in streaming mode
3. There's no way to use language detection without creating a custom subclass

Current Behavior

        super().__init__(
            capabilities=stt.STTCapabilities(streaming=True, interim_results=interim_results)
        )

Expected Behavior

The STT class should automatically disable streaming capabilities when detect_language=True is set, allowing it to work seamlessly in non-streaming mode without requiring custom implementations.

Proposed Solution

Modify the Deepgram STT __init__ method to conditionally set streaming capabilities based on the detect_language parameter:

def __init__(
    self,
    *,
    model: DeepgramModels | str = "nova-3",
    language: DeepgramLanguages | str = "en-US", 
    detect_language: bool = False,
    # ... other parameters
) -> None:
    # Conditionally disable streaming when language detection is enabled
    super().__init__(
        capabilities=stt.STTCapabilities(
            streaming=not detect_language,  # Disable streaming if detect_language=True
            interim_results=interim_results
        )
    )
    # ... rest of initialization

Use Case

This is particularly important for voice agent applications that need to:

1. Detect the caller's language automatically at the beginning of a conversation
2. Use non-streaming recognition for the initial language detection phase
3. Switch to language-specific agents after detection

Current Workaround

We currently have to create a custom subclass that overrides the initialization:

class NonStreamingDeepgramSTT(deepgram.STT):
    def __init__(self, *args, **kwargs):
        detect_language = kwargs.get("detect_language", False)
        interim_results = kwargs.get("interim_results", True)
        
        # Initialize parent stt.STT with appropriate capabilities
        stt.STT.__init__(
            self,
            capabilities=stt.STTCapabilities(
                streaming=not detect_language,
                interim_results=interim_results,
            ),
        )
        # ... manually set up all Deepgram attributes

Impact

This change would:

• Make the Deepgram STT implementation more intuitive
• Remove the need for custom subclasses when using language detection
• Align with Deepgram's actual API capabilities (which supports language detection in non-streaming mode)

Version Information

• LiveKit Agents Python SDK version: Latest
• Deepgram plugin version: Latest

Additional Context

The Deepgram API itself supports language detection in non-streaming mode (prerecorded API), so this is purely a limitation of the LiveKit SDK wrapper, not the underlying service.

[longcw]

You can wrap the Deepgram STT with stt.StreamAdapter to force it works in non-streaming mode

    from livekit.agents import stt

    session = AgentSession(
        ...
        stt=stt.StreamAdapter(
            stt=deepgram.STT(model="nova-3", language="multi", detect_language=True),
            vad=ctx.proc.userdata["vad"],
        ),
    )

[ringofhealth]

this works. thank!