feat: hybrid pipeline — gpt-realtime-whisper + gpt-5.4-nano + TTS
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
This commit is contained in:
+53
-92
@@ -1,8 +1,6 @@
|
||||
"""Kira — AI body double backend
|
||||
|
||||
OpenAI Realtime API pipeline:
|
||||
mic audio → [built-in STT → GPT-4o-mini → built-in TTS] → speaker audio
|
||||
Single WebSocket, ~300-800ms latency
|
||||
Hybrid pipeline: gpt-realtime-whisper (streaming STT) → gpt-5.4-nano (LLM) → OpenAI TTS
|
||||
"""
|
||||
|
||||
import json
|
||||
@@ -15,7 +13,7 @@ from fastapi import FastAPI, WebSocket, WebSocketDisconnect
|
||||
from fastapi.middleware.cors import CORSMiddleware
|
||||
|
||||
from config import settings
|
||||
from services.realtime import RealtimeRelay
|
||||
from services.hybrid import HybridPipeline
|
||||
from services.memory import kira_memory
|
||||
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
@@ -54,102 +52,92 @@ async def conversation_ws(websocket: WebSocket):
|
||||
identified = False
|
||||
logger.info(f"[{session_id}] WebSocket connected")
|
||||
|
||||
# Track conversation for Honcho
|
||||
pending_transcripts: list[tuple[str, str]] = []
|
||||
|
||||
# Will be set when Realtime relay is ready
|
||||
relay_ready = asyncio.Event()
|
||||
relay: RealtimeRelay | None = None
|
||||
relay_task: asyncio.Task | None = None
|
||||
pending_transcripts: list[str] = []
|
||||
pipeline: HybridPipeline | None = None
|
||||
pipeline_task: asyncio.Task | None = None
|
||||
pipeline_ready = asyncio.Event()
|
||||
audio_queue: asyncio.Queue[bytes] = asyncio.Queue()
|
||||
text_queue: asyncio.Queue[str] = asyncio.Queue()
|
||||
|
||||
memory_suffix = ""
|
||||
|
||||
async def on_ready():
|
||||
relay_ready.set()
|
||||
logger.info(f"[{session_id}] Realtime relay ready")
|
||||
pipeline_ready.set()
|
||||
logger.info(f"[{session_id}] Pipeline ready")
|
||||
|
||||
async def on_transcript_delta(delta: str):
|
||||
"""Streaming partial transcript."""
|
||||
await websocket.send_json({"type": "transcript_delta", "text": delta})
|
||||
|
||||
async def on_transcript_done(full: str):
|
||||
"""Full utterance received."""
|
||||
await websocket.send_json({"type": "transcript", "role": "user", "text": full})
|
||||
|
||||
async def on_audio_delta(audio_bytes: bytes):
|
||||
"""Forward audio chunks from OpenAI to the client."""
|
||||
"""Forward TTS audio to client."""
|
||||
try:
|
||||
audio_b64 = base64.b64encode(audio_bytes).decode("utf-8")
|
||||
await websocket.send_json({
|
||||
"type": "audio",
|
||||
"data": audio_b64,
|
||||
})
|
||||
await websocket.send_json({"type": "audio", "data": audio_b64})
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
async def on_transcript(text: str):
|
||||
"""Store transcripts for Honcho."""
|
||||
pending_transcripts.append(("transcript", text))
|
||||
role, content = text.split(": ", 1)
|
||||
logger.info(f"[{session_id}] {role}: {content}")
|
||||
await websocket.send_json({
|
||||
"type": "transcript",
|
||||
"role": role,
|
||||
"text": content,
|
||||
})
|
||||
|
||||
async def on_speech_started():
|
||||
"""Kira started speaking."""
|
||||
async def on_speech_start():
|
||||
await websocket.send_json({"type": "speaking_start"})
|
||||
|
||||
async def on_speech_stopped():
|
||||
"""Kira finished speaking."""
|
||||
async def on_speech_end():
|
||||
await websocket.send_json({"type": "speaking_end"})
|
||||
|
||||
async def on_interruption():
|
||||
"""User interrupted — Kira stops speaking."""
|
||||
await websocket.send_json({"type": "interruption"})
|
||||
|
||||
async def on_error(msg: str):
|
||||
await websocket.send_json({"type": "error", "message": msg})
|
||||
|
||||
# ── Create and start the Realtime relay ──
|
||||
relay = RealtimeRelay(
|
||||
# Create pipeline
|
||||
pipeline = HybridPipeline(
|
||||
on_transcript_delta=on_transcript_delta,
|
||||
on_transcript_done=on_transcript_done,
|
||||
on_audio_delta=on_audio_delta,
|
||||
on_transcript=on_transcript,
|
||||
on_speech_started=on_speech_started,
|
||||
on_speech_stopped=on_speech_stopped,
|
||||
on_interruption=on_interruption,
|
||||
on_error=on_error,
|
||||
on_speech_start=on_speech_start,
|
||||
on_speech_end=on_speech_end,
|
||||
on_ready=on_ready,
|
||||
on_error=on_error,
|
||||
memory_suffix=memory_suffix,
|
||||
)
|
||||
|
||||
relay_task = asyncio.create_task(relay.connect())
|
||||
pipeline_task = asyncio.create_task(pipeline.connect())
|
||||
|
||||
# Wait for relay to be ready
|
||||
try:
|
||||
await asyncio.wait_for(relay_ready.wait(), timeout=15)
|
||||
await asyncio.wait_for(pipeline_ready.wait(), timeout=15)
|
||||
except asyncio.TimeoutError:
|
||||
logger.error(f"[{session_id}] Realtime relay failed to connect")
|
||||
logger.error(f"[{session_id}] Pipeline failed to connect")
|
||||
await websocket.send_json({"type": "error", "message": "Failed to connect to AI"})
|
||||
relay_task.cancel()
|
||||
pipeline_task.cancel()
|
||||
return
|
||||
|
||||
# ── Forward audio/text from client to relay ──
|
||||
# Forward audio/text from client to pipeline
|
||||
async def forward_audio():
|
||||
while relay and relay._connected:
|
||||
while pipeline and pipeline._connected:
|
||||
try:
|
||||
pcm16 = await asyncio.wait_for(audio_queue.get(), timeout=1)
|
||||
await relay.send_audio(pcm16)
|
||||
await pipeline.send_audio(pcm16)
|
||||
except asyncio.TimeoutError:
|
||||
continue
|
||||
except Exception:
|
||||
break
|
||||
|
||||
async def forward_text():
|
||||
while relay and relay._connected:
|
||||
while pipeline and pipeline._connected:
|
||||
try:
|
||||
text = await asyncio.wait_for(text_queue.get(), timeout=1)
|
||||
await relay.send_text(text)
|
||||
await pipeline.send_text(text)
|
||||
# Store in Honcho
|
||||
if kira_memory.enabled and identified:
|
||||
kira_memory.store_user_message(text)
|
||||
except asyncio.TimeoutError:
|
||||
continue
|
||||
except Exception:
|
||||
break
|
||||
|
||||
fwd_audio_task = asyncio.create_task(forward_audio())
|
||||
fwd_text_task = asyncio.create_task(forward_text())
|
||||
fwd_audio = asyncio.create_task(forward_audio())
|
||||
fwd_text = asyncio.create_task(forward_text())
|
||||
|
||||
try:
|
||||
while True:
|
||||
@@ -171,30 +159,16 @@ async def conversation_ws(websocket: WebSocket):
|
||||
kira_memory.ensure_peers(user_id)
|
||||
kira_memory.ensure_session(session_id)
|
||||
|
||||
# Inject Honcho context into the Realtime session instructions
|
||||
memory_suffix = ""
|
||||
# Build memory context and update pipeline
|
||||
if kira_memory.enabled:
|
||||
try:
|
||||
ctx = kira_memory.build_system_prompt_suffix()
|
||||
if ctx:
|
||||
pipeline._memory_suffix = ctx
|
||||
memory_suffix = ctx
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
if relay and relay._connected and memory_suffix:
|
||||
await relay._send({
|
||||
"type": "session.update",
|
||||
"session": {
|
||||
"instructions": (
|
||||
"You are Kira, a warm, kind, and encouraging AI body double. "
|
||||
"Speak in a friendly, girly-pop tone. Help someone with ADHD "
|
||||
"stay focused. Keep responses short and supportive. "
|
||||
"Check in, remind breaks, celebrate wins. Never judgmental."
|
||||
+ memory_suffix
|
||||
),
|
||||
},
|
||||
})
|
||||
|
||||
await websocket.send_json({
|
||||
"type": "identified",
|
||||
"user_id": user_id,
|
||||
@@ -210,7 +184,7 @@ async def conversation_ws(websocket: WebSocket):
|
||||
kira_memory.set_user_preference(user_id, key, value)
|
||||
continue
|
||||
|
||||
# ── Audio from frontend (PCM16) ──
|
||||
# ── Audio (PCM16) ──
|
||||
if msg_type == "audio":
|
||||
audio_b64 = msg.get("data", "")
|
||||
if audio_b64:
|
||||
@@ -223,9 +197,6 @@ async def conversation_ws(websocket: WebSocket):
|
||||
text = msg.get("text", "").strip()
|
||||
if text:
|
||||
await text_queue.put(text)
|
||||
# Also store in Honcho immediately
|
||||
if kira_memory.enabled and identified:
|
||||
kira_memory.store_user_message(text)
|
||||
continue
|
||||
|
||||
if msg_type == "ping":
|
||||
@@ -236,19 +207,9 @@ async def conversation_ws(websocket: WebSocket):
|
||||
except Exception as e:
|
||||
logger.error(f"[{session_id}] Error: {e}")
|
||||
finally:
|
||||
# Store pending transcripts in Honcho
|
||||
if kira_memory.enabled and identified:
|
||||
for _, transcript_text in pending_transcripts:
|
||||
if transcript_text.startswith("user: "):
|
||||
content = transcript_text[6:]
|
||||
kira_memory.store_user_message(content)
|
||||
elif transcript_text.startswith("assistant: "):
|
||||
content = transcript_text[11:]
|
||||
kira_memory.store_kira_message(content)
|
||||
|
||||
fwd_audio_task.cancel()
|
||||
fwd_text_task.cancel()
|
||||
if relay:
|
||||
await relay.disconnect()
|
||||
if relay_task:
|
||||
relay_task.cancel()
|
||||
fwd_audio.cancel()
|
||||
fwd_text.cancel()
|
||||
if pipeline:
|
||||
await pipeline.disconnect()
|
||||
if pipeline_task:
|
||||
pipeline_task.cancel()
|
||||
|
||||
Reference in New Issue
Block a user