RTVIObserver drops upstream-only TranscriptionFrame from Gemini Live (and other realtime LLMs)

[acrdlph]

Environment

• pipecat version: 0.0.102 (and current main at commit c682a44)
• Python version: 3.13.5
• Operating System: macOS 15.4 (Darwin 25.2.0)

Issue description

Commit 5f64dae0c ("Filter RTVIObserver to downstream frames only") re-added a direction != FrameDirection.DOWNSTREAM: return guard to RTVIObserver.on_push_frame(). This was intended to prevent duplicate RTVI messages for frames that are broadcast in both directions (e.g. UserStartedSpeakingFrame, FunctionCallResultFrame).

However, GeminiLiveLLMService pushes TranscriptionFrame only upstream (FrameDirection.UPSTREAM) — it is never broadcast downstream. As a result, the RTVIObserver silently drops every user transcription, and RTVI clients never receive user-transcription messages.

The same issue likely affects OpenAIRealtimeLLMService and any other realtime/multimodal LLM service that pushes transcription frames upstream only.

Relevant code:

• RTVIObserver.on_push_frame direction guard: src/pipecat/processors/frameworks/rtvi.py (line ~1226)
• GeminiLiveLLMService._push_user_transcription pushing upstream: src/pipecat/services/google/gemini_live/llm.py (line ~1615)

# RTVIObserver — drops ALL upstream frames including TranscriptionFrame
if direction != FrameDirection.DOWNSTREAM:
    return

# GeminiLiveLLMService — pushes TranscriptionFrame upstream only
await self.push_frame(
    TranscriptionFrame(text=text, user_id="", timestamp=time_now_iso8601(), result=result),
    FrameDirection.UPSTREAM,
)

Reproduction steps

1. Create a bot using GeminiLiveLLMService with any pipeline (e.g. the 26b-gemini-live-function-calling.py example)
2. Connect an RTVI client (e.g. @pipecat-ai/client-react)
3. Speak to the bot
4. Observe that the bot responds with audio, but the RTVI client never receives user-transcription messages

Expected behavior

RTVI clients should receive user-transcription messages containing the user's speech transcribed by the Gemini Live API (via input_audio_transcription).

This worked correctly before commit 5f64dae0c when the direction filter was not present in RTVIObserver.

Actual behavior

RTVIObserver.on_push_frame() returns early for all upstream frames. Since TranscriptionFrame from Gemini Live is only pushed upstream, it is never processed, and no user-transcription RTVI message is ever sent to the client.

The Gemini Live service does receive and process transcriptions from the API (visible in logs as [Transcription:user] [...] at DEBUG level), but the RTVIObserver discards them before they can be forwarded.

Logs

# Transcription IS received from Gemini API:
2026-02-14 19:06:28.685 | DEBUG | pipecat.services.google.gemini_live.llm:_handle_msg_input_transcription:1683 - [Transcription:user] [maybe problem solve neighbor yeah.]

# But NO corresponding RTVI user-transcription message is ever sent to the client

Suggested fix

Either:

1. Broadcast TranscriptionFrame in both directions from GeminiLiveLLMService (and other realtime LLM services), so the downstream copy reaches RTVIObserver — OR
2. Allowlist TranscriptionFrame and InterimTranscriptionFrame in the RTVIObserver's direction check, so upstream-only transcription frames are still processed