marcosremar2's picture
Integrate AvatarAPI: prepare for direct Python calls replacing WebSocket
60e84bc
Raw
History Blame Contribute Delete
21 kB
"""
Interface Server - WebRTC Streaming com VP9
Porta: 8080
Arquitetura:
1. Cliente conecta via WebRTC (signaling por WebSocket)
2. Servidor envia stream de video VP9 + audio Opus
3. Fusao idle/lip-sync feita no backend
4. Frontend apenas renderiza o <video>
Framework: aiortc (https://github.com/aiortc/aiortc)
"""
from aiohttp import web
import aiohttp
import asyncio
import json
import base64
import os
import sys
import time
import uuid
import fractions
import numpy as np
from av import VideoFrame, AudioFrame
from aiortc import RTCPeerConnection, RTCSessionDescription, MediaStreamTrack, RTCConfiguration, RTCIceServer
from aiortc.contrib.media import MediaRelay
import cv2
import subprocess
import tempfile
from scipy import signal
# Add path to video directory for avatar_api import
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'video'))
# Import AvatarAPI
try:
from avatar_api import AvatarAPI
AVATAR_API_AVAILABLE = True
print("[Import] AvatarAPI imported successfully")
except ImportError as e:
AVATAR_API_AVAILABLE = False
print(f"[Import] WARNING: Could not import AvatarAPI: {e}")
print("[Import] Falling back to WebSocket mode")
# Configuracao
PORT = int(os.getenv("PORT", "8080"))
IDLE_VIDEO = os.path.join(os.path.dirname(__file__), "idle.mp4")
AVATAR_VIDEO = os.path.join(os.path.dirname(__file__), '..', 'video', 'avatar_parado.mp4')
WAV2LIP_CHECKPOINT = os.path.join(os.path.dirname(__file__), '..', 'video', 'checkpoints', 'wav2lip_gan.pth')
TTS_URL = os.getenv("TTS_URL", "http://localhost:8081")
# Constantes
VIDEO_FPS = 25
VIDEO_WIDTH = 854 # 480p
VIDEO_HEIGHT = 480
AUDIO_SAMPLE_RATE = 48000 # 48kHz é a frequência nativa do Opus WebRTC
ORPHEUS_SAMPLE_RATE = 24000 # Orpheus gera a 24kHz
VIDEO_TIME_BASE = fractions.Fraction(1, VIDEO_FPS)
AUDIO_TIME_BASE = fractions.Fraction(1, AUDIO_SAMPLE_RATE)
# Cache global
idle_frames_cache = []
pcs = set() # Track peer connections
relay = MediaRelay()
avatar_api = None # Global AvatarAPI instance
routes = web.RouteTableDef()
def calculate_frame_difference(frame1, frame2):
"""Calcula diferenca entre dois frames (0 = identicos, 1 = muito diferentes)."""
if frame1 is None or frame2 is None:
return 1.0
# Converter para grayscale se necessario
if len(frame1.shape) == 3:
gray1 = cv2.cvtColor(frame1, cv2.COLOR_RGB2GRAY)
else:
gray1 = frame1
if len(frame2.shape) == 3:
gray2 = cv2.cvtColor(frame2, cv2.COLOR_RGB2GRAY)
else:
gray2 = frame2
# Redimensionar para mesma resolucao se necessario
if gray1.shape != gray2.shape:
gray2 = cv2.resize(gray2, (gray1.shape[1], gray1.shape[0]))
# Calcular diferenca
diff = cv2.absdiff(gray1, gray2)
return np.mean(diff) / 255.0
def find_best_matching_idle_frame(last_speak_frame, idle_frames, sample_step=10):
"""
Encontra o frame idle mais similar ao ultimo frame de fala.
Usa amostragem para ser mais rapido.
"""
if not idle_frames or last_speak_frame is None:
return 0, float('inf')
best_idx = 0
best_diff = float('inf')
# Primeira passada: amostragem grosseira
for i in range(0, len(idle_frames), sample_step):
diff = calculate_frame_difference(last_speak_frame, idle_frames[i])
if diff < best_diff:
best_diff = diff
best_idx = i
# Segunda passada: refinamento na regiao
start = max(0, best_idx - sample_step)
end = min(len(idle_frames), best_idx + sample_step)
for i in range(start, end):
diff = calculate_frame_difference(last_speak_frame, idle_frames[i])
if diff < best_diff:
best_diff = diff
best_idx = i
return best_idx, best_diff
def trim_high_motion_frames(frames, threshold_multiplier=1.0, max_trim=20):
"""
Remove frames do final que tem movimento muito alto (saltos).
"""
if len(frames) < 20:
return frames, None
# Calcular diferencas entre frames consecutivos (ultimos 20)
last_n = min(20, len(frames) - 1)
differences = []
for i in range(len(frames) - last_n, len(frames)):
if i > 0:
diff = calculate_frame_difference(frames[i-1], frames[i])
differences.append((i, diff))
if not differences:
return frames, frames[-1] if frames else None
# Calcular media e desvio padrao
diffs = [d[1] for d in differences]
mean_diff = np.mean(diffs)
std_diff = np.std(diffs)
# Threshold agressivo
threshold = mean_diff + threshold_multiplier * std_diff
min_threshold = 0.7
if threshold > min_threshold:
threshold = min_threshold
# Encontrar onde comecam os frames problematicos
trim_from = len(frames)
frames_removed = 0
for i in range(len(differences) - 1, -1, -1):
idx, diff = differences[i]
if diff > threshold:
trim_from = idx
frames_removed += 1
if frames_removed >= max_trim:
break
else:
break
frames_to_trim = len(frames) - trim_from
if frames_to_trim > 0 and frames_to_trim <= max_trim:
print(f"[Trim] Removendo {frames_to_trim} frames problematicos")
trimmed_frames = frames[:trim_from]
return trimmed_frames, trimmed_frames[-1] if trimmed_frames else None
return frames, frames[-1] if frames else None
def load_idle_frames():
"""Carrega frames do idle.mp4 como arrays numpy, redimensionados para 480p."""
global idle_frames_cache
if idle_frames_cache:
return idle_frames_cache
if not os.path.exists(IDLE_VIDEO):
print(f"[Idle] Arquivo nao encontrado: {IDLE_VIDEO}")
return []
print(f"[Idle] Carregando frames de {IDLE_VIDEO}...")
cap = cv2.VideoCapture(IDLE_VIDEO)
# Detectar resolução original
original_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
original_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
print(f"[Idle] Resolucao original: {original_width}x{original_height}")
print(f"[Idle] Redimensionando para: {VIDEO_WIDTH}x{VIDEO_HEIGHT} (480p)")
while True:
ret, frame = cap.read()
if not ret:
break
# Converter BGR para RGB
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
# Redimensionar para 480p se necessario
if original_width != VIDEO_WIDTH or original_height != VIDEO_HEIGHT:
frame_rgb = cv2.resize(frame_rgb, (VIDEO_WIDTH, VIDEO_HEIGHT),
interpolation=cv2.INTER_LANCZOS4)
idle_frames_cache.append(frame_rgb)
cap.release()
print(f"[Idle] Carregados {len(idle_frames_cache)} frames @ {VIDEO_WIDTH}x{VIDEO_HEIGHT}")
return idle_frames_cache
class AvatarVideoTrack(MediaStreamTrack):
"""
Track de video que envia frames do avatar.
Alterna entre idle e lip-sync conforme necessario.
"""
kind = "video"
def __init__(self):
super().__init__()
self.idle_frames = load_idle_frames()
self.current_idx = 0
self.frame_count = 0
self.start_time = None
# Estado
self.is_speaking = False
self.speaking_frames = []
self.speaking_idx = 0
self.best_idle_idx = None # Frame idle para transicao suave
# Dimensoes do video (480p fixo)
self.width = VIDEO_WIDTH
self.height = VIDEO_HEIGHT
print(f"[VideoTrack] Inicializado: {self.width}x{self.height} @ {VIDEO_FPS}fps (480p fixo)")
async def recv(self):
"""Retorna o proximo frame de video."""
if self.start_time is None:
self.start_time = time.time()
# Calcular pts baseado no tempo
pts = int(self.frame_count * VIDEO_TIME_BASE.denominator / VIDEO_FPS)
self.frame_count += 1
# Escolher frame: speaking ou idle
if self.is_speaking and self.speaking_frames:
if self.speaking_idx < len(self.speaking_frames):
frame_data = self.speaking_frames[self.speaking_idx]
self.speaking_idx += 1
else:
# Acabou a fala, voltar ao idle no best_idle_idx
self.is_speaking = False
self.speaking_frames = []
self.speaking_idx = 0
# Usar o frame idle pre-calculado para transicao suave
if self.best_idle_idx is not None and self.idle_frames:
self.current_idx = self.best_idle_idx
self.best_idle_idx = None
print(f"[VideoTrack] Transicao suave -> idle frame {self.current_idx}")
frame_data = self.idle_frames[self.current_idx % len(self.idle_frames)]
self.current_idx += 1
elif self.idle_frames:
frame_data = self.idle_frames[self.current_idx % len(self.idle_frames)]
self.current_idx += 1
else:
# Fallback: frame preto
frame_data = np.zeros((self.height, self.width, 3), dtype=np.uint8)
# Criar VideoFrame
frame = VideoFrame.from_ndarray(frame_data, format="rgb24")
frame.pts = pts
frame.time_base = VIDEO_TIME_BASE
# Manter timing de 25fps
elapsed = time.time() - self.start_time
expected = self.frame_count / VIDEO_FPS
if expected > elapsed:
await asyncio.sleep(expected - elapsed)
return frame
def set_speaking_frames(self, frames):
"""Define frames de lip-sync para reproduzir com transicao suave."""
# Aplicar trim de frames problematicos
trimmed_frames, last_frame = trim_high_motion_frames(frames)
# Encontrar o melhor frame idle para transicao suave
if last_frame is not None and self.idle_frames:
best_idx, best_diff = find_best_matching_idle_frame(
last_frame, self.idle_frames, sample_step=10
)
self.best_idle_idx = best_idx
print(f"[VideoTrack] Best match: idle frame {best_idx} (diff: {best_diff:.2f})")
else:
self.best_idle_idx = None
self.speaking_frames = trimmed_frames
self.speaking_idx = 0
self.is_speaking = True
print(f"[VideoTrack] Speaking: {len(trimmed_frames)} frames (original: {len(frames)})")
class AvatarAudioTrack(MediaStreamTrack):
"""
Track de audio que envia silencio ou audio do Orpheus.
"""
kind = "audio"
def __init__(self):
super().__init__()
self.sample_rate = AUDIO_SAMPLE_RATE
self.samples_per_frame = 1920 # 40ms @ 48kHz (era 960 @ 24kHz)
self.frame_count = 0
self.start_time = None
# Buffer de audio
self.audio_buffer = []
self.buffer_idx = 0
print(f"[AudioTrack] Inicializado: {self.sample_rate}Hz, {self.samples_per_frame} samples/frame (40ms)")
async def recv(self):
"""Retorna o proximo frame de audio."""
if self.start_time is None:
self.start_time = time.time()
pts = self.frame_count * self.samples_per_frame
self.frame_count += 1
# Pegar audio do buffer ou silencio
if self.audio_buffer and self.buffer_idx < len(self.audio_buffer):
samples = self.audio_buffer[self.buffer_idx]
self.buffer_idx += 1
else:
# Silencio
samples = np.zeros(self.samples_per_frame, dtype=np.int16)
# Criar AudioFrame
frame = AudioFrame(format="s16", layout="mono", samples=len(samples))
frame.sample_rate = self.sample_rate
frame.pts = pts
frame.time_base = AUDIO_TIME_BASE
# Copiar samples
frame.planes[0].update(samples.tobytes())
# Manter timing
elapsed = time.time() - self.start_time
expected = self.frame_count * self.samples_per_frame / self.sample_rate
if expected > elapsed:
await asyncio.sleep(expected - elapsed)
return frame
def set_audio(self, pcm_data):
"""Define audio PCM para reproduzir."""
# Converter bytes para numpy array
samples = np.frombuffer(pcm_data, dtype=np.int16)
print(f"[AudioTrack] Recebido: {len(samples)} samples @ {ORPHEUS_SAMPLE_RATE}Hz")
# RESAMPLE: Orpheus gera 24kHz, mas Opus WebRTC funciona melhor em 48kHz (nativo)
# Upsample 24kHz -> 48kHz (2x) usando scipy high-quality resampling
if ORPHEUS_SAMPLE_RATE != AUDIO_SAMPLE_RATE:
num_samples_48k = int(len(samples) * AUDIO_SAMPLE_RATE / ORPHEUS_SAMPLE_RATE)
samples_48k = signal.resample(samples, num_samples_48k).astype(np.int16)
print(f"[AudioTrack] Upsampled: {len(samples_48k)} samples @ {AUDIO_SAMPLE_RATE}Hz (scipy high-quality)")
samples = samples_48k
# Dividir em frames de 40ms (1920 samples @ 48kHz)
self.audio_buffer = []
for i in range(0, len(samples), self.samples_per_frame):
chunk = samples[i:i + self.samples_per_frame]
if len(chunk) < self.samples_per_frame:
# Padding com zeros
chunk = np.pad(chunk, (0, self.samples_per_frame - len(chunk)))
self.audio_buffer.append(chunk)
self.buffer_idx = 0
print(f"[AudioTrack] Buffer: {len(self.audio_buffer)} frames x {self.samples_per_frame} samples (40ms cada)")
class AvatarSession:
"""Gerencia uma sessao WebRTC com o cliente."""
def __init__(self, pc, video_track, audio_track):
self.pc = pc
self.video_track = video_track
self.audio_track = audio_track
self.wav2lip_ws = None
self.wav2lip_session = None
async def generate(self, text: str, voice: str):
"""Gera fala com lip-sync via Wav2Lip."""
print(f"[Session] Gerando: '{text[:50]}...'")
try:
# Conectar ao Wav2Lip
self.wav2lip_session = aiohttp.ClientSession()
self.wav2lip_ws = await self.wav2lip_session.ws_connect(
WAV2LIP_WS,
timeout=aiohttp.ClientWSTimeout(ws_close=120)
)
# Enviar requisicao
await self.wav2lip_ws.send_json({
"action": "generate",
"text": text,
"voice": voice
})
speaking_frames = []
audio_data = b''
# Receber frames e audio
async for msg in self.wav2lip_ws:
if msg.type == aiohttp.WSMsgType.TEXT:
data = json.loads(msg.data)
msg_type = data.get("type", "")
if msg_type == "frame":
frame_b64 = data.get("frame", "")
if frame_b64:
# Decodificar JPEG para numpy
jpeg_data = base64.b64decode(frame_b64)
nparr = np.frombuffer(jpeg_data, np.uint8)
frame = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
speaking_frames.append(frame_rgb)
elif msg_type == "full_audio":
audio_b64 = data.get("audio", "")
if audio_b64:
audio_data = base64.b64decode(audio_b64)
elif msg_type == "done":
break
elif msg_type == "error":
print(f"[Session] Erro Wav2Lip: {data.get('message')}")
break
elif msg.type in (aiohttp.WSMsgType.CLOSED, aiohttp.WSMsgType.ERROR):
break
await self.wav2lip_ws.close()
await self.wav2lip_session.close()
# Aplicar frames e audio aos tracks
if speaking_frames:
self.video_track.set_speaking_frames(speaking_frames)
if audio_data:
self.audio_track.set_audio(audio_data)
print(f"[Session] Gerado: {len(speaking_frames)} frames, {len(audio_data)} bytes audio")
except Exception as e:
print(f"[Session] Erro: {e}")
import traceback
traceback.print_exc()
async def close(self):
"""Fecha a sessao."""
if self.wav2lip_ws and not self.wav2lip_ws.closed:
await self.wav2lip_ws.close()
if self.wav2lip_session:
await self.wav2lip_session.close()
# Armazenar sessoes ativas
sessions = {}
@routes.post("/offer")
async def offer(request):
"""Recebe offer SDP do cliente e retorna answer."""
params = await request.json()
offer = RTCSessionDescription(sdp=params["sdp"], type=params["type"])
# Configurar ICE servers (STUN + TURN publicos)
ice_servers = [
RTCIceServer(urls=["stun:stun.l.google.com:19302"]),
# Servidores TURN com múltiplas URLs
RTCIceServer(
urls=[
"turn:openrelay.metered.ca:80",
"turn:openrelay.metered.ca:443",
"turn:openrelay.metered.ca:443?transport=tcp"
],
username="openrelayproject",
credential="openrelayproject"
),
# TURN alternativo (Twilio)
RTCIceServer(
urls=["turn:global.turn.twilio.com:3478?transport=udp"],
username="f4b4035eaa76f4a55de5f4351567653ee4ff6fa97b50b6b334fcc1be9c27212d",
credential="w1uxM55V9yVoqyVFjt+mxDBV0F87AUCemaYVQGxsPLw="
),
]
config = RTCConfiguration(iceServers=ice_servers)
pc = RTCPeerConnection(configuration=config)
pc_id = str(uuid.uuid4())
pcs.add(pc)
print(f"[WebRTC] Nova conexao: {pc_id}")
# Criar tracks
video_track = AvatarVideoTrack()
audio_track = AvatarAudioTrack()
# Adicionar tracks ao peer connection
pc.addTrack(video_track)
pc.addTrack(audio_track)
# Criar sessao
session = AvatarSession(pc, video_track, audio_track)
sessions[pc_id] = session
@pc.on("iceconnectionstatechange")
async def on_ice_state():
print(f"[ICE] Estado: {pc.iceConnectionState}")
@pc.on("icegatheringstatechange")
async def on_ice_gathering():
print(f"[ICE] Gathering: {pc.iceGatheringState}")
@pc.on("connectionstatechange")
async def on_connectionstatechange():
print(f"[WebRTC] Estado: {pc.connectionState}")
if pc.connectionState == "failed" or pc.connectionState == "closed":
await pc.close()
pcs.discard(pc)
if pc_id in sessions:
await sessions[pc_id].close()
del sessions[pc_id]
# Processar offer e criar answer
await pc.setRemoteDescription(offer)
answer = await pc.createAnswer()
await pc.setLocalDescription(answer)
return web.json_response({
"sdp": pc.localDescription.sdp,
"type": pc.localDescription.type,
"session_id": pc_id
})
@routes.post("/generate")
async def generate(request):
"""Gera fala com lip-sync."""
params = await request.json()
session_id = params.get("session_id")
text = params.get("text", "").strip()
voice = params.get("voice", "tara")
if not session_id or session_id not in sessions:
return web.json_response({"error": "Sessao invalida"}, status=400)
if not text:
return web.json_response({"error": "Texto obrigatorio"}, status=400)
session = sessions[session_id]
asyncio.create_task(session.generate(text, voice))
return web.json_response({"status": "generating"})
@routes.get("/")
async def index(request):
return web.FileResponse(os.path.join(os.path.dirname(__file__), "index.html"))
@routes.get("/{filename}")
async def static_file(request):
filename = request.match_info["filename"]
filepath = os.path.join(os.path.dirname(__file__), filename)
if os.path.exists(filepath):
return web.FileResponse(filepath)
return web.Response(status=404)
@routes.get("/health")
async def health(request):
return web.json_response({
"status": "ok",
"mode": "webrtc",
"connections": len(pcs)
})
async def on_shutdown(app):
"""Fecha todas as conexoes ao desligar."""
coros = [pc.close() for pc in pcs]
await asyncio.gather(*coros)
pcs.clear()
app = web.Application()
app.add_routes(routes)
app.on_shutdown.append(on_shutdown)
if __name__ == "__main__":
print("=" * 50)
print("Interface Server - WebRTC VP9 Streaming")
print("=" * 50)
print(f"Porta: {PORT}")
print(f"Idle Video: {IDLE_VIDEO}")
print(f"Wav2Lip: {WAV2LIP_WS}")
print("=" * 50)
print("Endpoints:")
print(" POST /offer - WebRTC signaling")
print(" POST /generate - Gerar fala")
print("=" * 50)
# Pre-carregar idle frames
print("Carregando idle frames...")
load_idle_frames()
print("=" * 50)
web.run_app(app, host="0.0.0.0", port=PORT)