The REST STT revert commit (0e74a16) deleted lines 78-262 including the message loop, identify handler, audio handler, text handler, and disconnect cleanup. This caused the WS to accept then immediately close, triggering a reconnect loop.
Refactored for clarity: transcribe_audio(), get_kira_response(), stream_tts() as standalone async helpers. Full pipeline restored.
- Backend: added transcribe_audio (gpt-4o-transcribe), switched audio handler to full blob -> REST -> LLM -> streaming TTS
- Frontend: MediaRecorder (webm/opus) full recording sent on stop (one blob per utterance)
- Removed dead WhisperStream callbacks and pending_transcript/lock
- This unblocks voice per AUDIT item 1 (Option B fallback). Deltas will come in later item.
- Also preps for deprecation fix (MediaRecorder is the good path).
- URL now uses ?model=gpt-realtime-whisper (was invalid gpt-4o-mini-realtime-preview)
- Cleaned session.update (removed modalities that may not apply)
- Expanded _handle to catch input_audio_transcription.delta and .completed events
- on_error now forwards transcription errors to frontend client
- Per AUDIT + PLAN item 1
Replaces REST-based transcription (gpt-4o-transcribe) with WebSocket
streaming via gpt-realtime-whisper. Frontend captures PCM16 audio and
streams it through the backend to a Realtime transcription session.
- Server-side VAD detects utterance boundaries automatically
- Word-level transcript deltas stream to the client in real-time
- On utterance end, gpt-5.4-nano generates a response
- TTS streams back via with_streaming_response
- Total pipeline: PCM16 → Realtime WS → LLM → streaming TTS
Replaces gpt-4o-mini-transcribe (/usr/bin/bash.003/min) with gpt-realtime-whisper
(/usr/bin/bash.017/min). Expected to reduce transcription latency from ~2.6s
to ~1s due to the model's realtime optimization.
Replaced synchronous TTS (waiting for full audio at 5.9s) with
streaming TTS that sends audio chunks as they arrive. Backend now
accumulates chunks in audioBufferRef and plays the complete stream
on speaking_end. Reduces TTS latency from ~6s to ~1s first byte.
The memory context was being rebuilt on every conversation turn via
build_system_prompt(), which calls Honcho's dialectic reasoning API
twice (get_user_context + get_kira_context). Each call takes 5-15s.
Now the memory suffix is computed ONCE during identify and cached in
a memory_suffix variable for the session duration. Per-turn latency
drops from ~37s to ~3s.
Also removed duplicated _pcm16_to_wav and cleaned up orphaned code.
PCM16 capture via AudioContext was streaming raw audio continuously,
causing massive accumulated buffers that took ~20s to transcribe.
Replaced with MediaRecorder which records compressed Opus/webm and
sends a single blob on release — much smaller, faster to transcribe.
Also removed all unused PCM16/WAV helper functions from both frontend
and backend.
Frontend captures PCM16 mono 24kHz audio. The transcription API
expects a proper audio container format (wav, webm, etc.), not raw
PCM16 data. Added _pcm16_to_wav() to wrap the raw bytes in a WAV
header before sending to gpt-4o-mini-transcribe.
The openai library's beta.realtime.connect() hardcodes the obsolete
'OpenAI-Beta: realtime=v1' header which the GA API rejects. Connecting
directly via the websockets library with only the Authorization header
resolves the 'beta_api_shape_disabled' error.
Hybrid approach gives streaming STT at ~/usr/bin/bash.017/min + cheap brain
at ~/usr/bin/bash.001/min + TTS at ~/usr/bin/bash.015/min = ~/usr/bin/bash.033/min total.
- gpt-realtime-whisper handles streaming transcription with VAD
- gpt-5.4-nano handles response generation (chat completions)
- OpenAI TTS (nova) for voice output
- Server VAD detects utterance boundaries
- Honcho memory context injected into system prompt
- Removed old full Realtime relay service
The old OpenAI-Beta: realtime=v1 header is rejected by the GA API.
Removing it via extra_headers override. Using gpt-realtime-2 which
is the current production Realtime model.
Replaced the 3-step sequential pipeline (Whisper STT → DeepSeek LLM
→ OpenAI TTS) with a single OpenAI Realtime API WebSocket using
gpt-4o-mini-realtime-preview.
- ~300-800ms latency vs 1-3s
- Server VAD for automatic turn detection
- Streaming audio chunks during playback
- Interruptions: user can speak over Kira mid-response
- Honcho memory still injected into session instructions
- Frontend captures PCM16 mono 24kHz via AudioContext
- Backend relays client ↔ OpenAI Realtime API
- Supports both voice (PCM16) and text input
- WelcomeScreen: first-time name entry with cute onboarding
- identify WS message: sets user_id, loads saved prefs from Honcho
- set_preference WS message: saves scene/outfit/accessory to Honcho metadata
- Preferences auto-load on return visits via localStorage + Honcho peer meta
- Kira uses the user's name in greeting and prompts
- Backend: get/set preference methods in KiraMemory service
- Frontend: optimistic preference updates, synced to backend on change
hobokenchicken