509 lines
19 KiB
Python
509 lines
19 KiB
Python
import os
|
|
import io
|
|
import base64
|
|
import time
|
|
import torch
|
|
import torchaudio
|
|
import numpy as np
|
|
from flask import Flask, render_template, request
|
|
from flask_socketio import SocketIO, emit
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
from collections import deque
|
|
import requests
|
|
import huggingface_hub
|
|
from generator import load_csm_1b, Segment
|
|
import threading
|
|
import queue
|
|
from flask import stream_with_context, Response
|
|
import time
|
|
|
|
# Configure environment with longer timeouts
|
|
os.environ["HF_HUB_DOWNLOAD_TIMEOUT"] = "600" # 10 minutes timeout for downloads
|
|
requests.adapters.DEFAULT_TIMEOUT = 60 # Increase default requests timeout
|
|
|
|
# Create a models directory for caching
|
|
os.makedirs("models", exist_ok=True)
|
|
|
|
app = Flask(__name__)
|
|
app.config['SECRET_KEY'] = 'your-secret-key'
|
|
socketio = SocketIO(app, cors_allowed_origins="*")
|
|
|
|
# Explicitly check for CUDA and print more detailed info
|
|
print("\n=== CUDA Information ===")
|
|
if torch.cuda.is_available():
|
|
print(f"CUDA is available")
|
|
print(f"CUDA version: {torch.version.cuda}")
|
|
print(f"Number of GPUs: {torch.cuda.device_count()}")
|
|
for i in range(torch.cuda.device_count()):
|
|
print(f"GPU {i}: {torch.cuda.get_device_name(i)}")
|
|
else:
|
|
print("CUDA is not available")
|
|
|
|
# Check for cuDNN
|
|
try:
|
|
import ctypes
|
|
ctypes.CDLL("libcudnn_ops_infer.so.8")
|
|
print("cuDNN is available")
|
|
except:
|
|
print("cuDNN is not available (libcudnn_ops_infer.so.8 not found)")
|
|
|
|
# Check for other compute platforms
|
|
if torch.backends.mps.is_available():
|
|
print("MPS (Apple Silicon) is available")
|
|
else:
|
|
print("MPS is not available")
|
|
print("========================\n")
|
|
|
|
# Check for CUDA availability and handle potential CUDA/cuDNN issues
|
|
try:
|
|
if torch.cuda.is_available():
|
|
# Try to initialize CUDA to check if libraries are properly loaded
|
|
_ = torch.zeros(1).cuda()
|
|
device = "cuda"
|
|
whisper_compute_type = "float16"
|
|
print("🟢 CUDA is available and initialized successfully")
|
|
elif torch.backends.mps.is_available():
|
|
device = "mps"
|
|
whisper_compute_type = "float32"
|
|
print("🟢 MPS is available (Apple Silicon)")
|
|
else:
|
|
device = "cpu"
|
|
whisper_compute_type = "int8"
|
|
print("🟡 Using CPU (CUDA/MPS not available)")
|
|
except Exception as e:
|
|
print(f"🔴 Error initializing CUDA: {e}")
|
|
print("🔴 Falling back to CPU")
|
|
device = "cpu"
|
|
whisper_compute_type = "int8"
|
|
|
|
print(f"Using device: {device}")
|
|
|
|
# Initialize models with proper error handling
|
|
whisper_model = None
|
|
csm_generator = None
|
|
llm_model = None
|
|
llm_tokenizer = None
|
|
|
|
def load_models():
|
|
global whisper_model, csm_generator, llm_model, llm_tokenizer
|
|
|
|
# Initialize Faster-Whisper for transcription
|
|
try:
|
|
print("Loading Whisper model...")
|
|
# Import here to avoid immediate import errors if package is missing
|
|
from faster_whisper import WhisperModel
|
|
whisper_model = WhisperModel("base", device=device, compute_type=whisper_compute_type, download_root="./models/whisper")
|
|
print("Whisper model loaded successfully")
|
|
except Exception as e:
|
|
print(f"Error loading Whisper model: {e}")
|
|
print("Will use backup speech recognition method if available")
|
|
|
|
# Initialize CSM model for audio generation
|
|
try:
|
|
print("Loading CSM model...")
|
|
csm_generator = load_csm_1b(device=device)
|
|
print("CSM model loaded successfully")
|
|
except Exception as e:
|
|
print(f"Error loading CSM model: {e}")
|
|
print("Audio generation will not be available")
|
|
|
|
# Initialize Llama 3.2 model for response generation
|
|
try:
|
|
print("Loading Llama 3.2 model...")
|
|
llm_model_id = "meta-llama/Llama-3.2-1B" # Choose appropriate size based on resources
|
|
llm_tokenizer = AutoTokenizer.from_pretrained(llm_model_id, cache_dir="./models/llama")
|
|
# Use the right data type based on device
|
|
dtype = torch.bfloat16 if device != "cpu" else torch.float32
|
|
llm_model = AutoModelForCausalLM.from_pretrained(
|
|
llm_model_id,
|
|
torch_dtype=dtype,
|
|
device_map=device,
|
|
cache_dir="./models/llama",
|
|
low_cpu_mem_usage=True
|
|
)
|
|
print("Llama 3.2 model loaded successfully")
|
|
except Exception as e:
|
|
print(f"Error loading Llama 3.2 model: {e}")
|
|
print("Will use a fallback response generation method")
|
|
|
|
# Store conversation context
|
|
conversation_context = {} # session_id -> context
|
|
CHUNK_SIZE = 24000 # Number of audio samples per chunk (1 second at 24kHz)
|
|
audio_stream_queues = {} # session_id -> queue for audio chunks
|
|
|
|
@app.route('/')
|
|
def index():
|
|
return render_template('index.html')
|
|
|
|
@socketio.on('connect')
|
|
def handle_connect():
|
|
print(f"Client connected: {request.sid}")
|
|
conversation_context[request.sid] = {
|
|
'segments': [],
|
|
'speakers': [0, 1], # 0 = user, 1 = bot
|
|
'audio_buffer': deque(maxlen=10), # Store recent audio chunks
|
|
'is_speaking': False,
|
|
'silence_start': None
|
|
}
|
|
emit('ready', {'message': 'Connection established'})
|
|
|
|
@socketio.on('disconnect')
|
|
def handle_disconnect():
|
|
print(f"Client disconnected: {request.sid}")
|
|
session_id = request.sid
|
|
|
|
# Clean up resources
|
|
if session_id in conversation_context:
|
|
del conversation_context[session_id]
|
|
|
|
if session_id in audio_stream_queues:
|
|
del audio_stream_queues[session_id]
|
|
|
|
@socketio.on('start_speaking')
|
|
def handle_start_speaking():
|
|
if request.sid in conversation_context:
|
|
conversation_context[request.sid]['is_speaking'] = True
|
|
conversation_context[request.sid]['audio_buffer'].clear()
|
|
print(f"User {request.sid} started speaking")
|
|
|
|
@socketio.on('audio_chunk')
|
|
def handle_audio_chunk(data):
|
|
if request.sid not in conversation_context:
|
|
return
|
|
|
|
context = conversation_context[request.sid]
|
|
|
|
# Decode audio data
|
|
audio_data = base64.b64decode(data['audio'])
|
|
audio_numpy = np.frombuffer(audio_data, dtype=np.float32)
|
|
audio_tensor = torch.tensor(audio_numpy)
|
|
|
|
# Add to buffer
|
|
context['audio_buffer'].append(audio_tensor)
|
|
|
|
# Check for silence to detect end of speech
|
|
if context['is_speaking'] and is_silence(audio_tensor):
|
|
if context['silence_start'] is None:
|
|
context['silence_start'] = time.time()
|
|
elif time.time() - context['silence_start'] > 1.0: # 1 second of silence
|
|
# Process the complete utterance
|
|
process_user_utterance(request.sid)
|
|
else:
|
|
context['silence_start'] = None
|
|
|
|
@socketio.on('stop_speaking')
|
|
def handle_stop_speaking():
|
|
if request.sid in conversation_context:
|
|
conversation_context[request.sid]['is_speaking'] = False
|
|
process_user_utterance(request.sid)
|
|
print(f"User {request.sid} stopped speaking")
|
|
|
|
def is_silence(audio_tensor, threshold=0.02):
|
|
"""Check if an audio chunk is silence based on amplitude threshold"""
|
|
return torch.mean(torch.abs(audio_tensor)) < threshold
|
|
|
|
def process_user_utterance(session_id):
|
|
"""Process completed user utterance, generate response and stream audio back"""
|
|
context = conversation_context[session_id]
|
|
|
|
if not context['audio_buffer']:
|
|
return
|
|
|
|
# Combine audio chunks
|
|
full_audio = torch.cat(list(context['audio_buffer']), dim=0)
|
|
context['audio_buffer'].clear()
|
|
context['is_speaking'] = False
|
|
context['silence_start'] = None
|
|
|
|
# Save audio to temporary WAV file for transcription
|
|
temp_audio_path = f"temp_audio_{session_id}.wav"
|
|
torchaudio.save(
|
|
temp_audio_path,
|
|
full_audio.unsqueeze(0),
|
|
44100 # Assuming 44.1kHz from client
|
|
)
|
|
|
|
try:
|
|
# Try using Whisper first if available
|
|
if whisper_model is not None:
|
|
user_text = transcribe_with_whisper(temp_audio_path)
|
|
else:
|
|
# Fallback to Google's speech recognition
|
|
user_text = transcribe_with_google(temp_audio_path)
|
|
|
|
if not user_text:
|
|
print("No speech detected.")
|
|
emit('error', {'message': 'No speech detected. Please try again.'}, room=session_id)
|
|
return
|
|
|
|
print(f"Transcribed: {user_text}")
|
|
|
|
# Add to conversation segments
|
|
user_segment = Segment(
|
|
text=user_text,
|
|
speaker=0, # User is speaker 0
|
|
audio=full_audio
|
|
)
|
|
context['segments'].append(user_segment)
|
|
|
|
# Generate bot response text
|
|
bot_response = generate_llm_response(user_text, context['segments'])
|
|
print(f"Bot response: {bot_response}")
|
|
|
|
# Send transcribed text to client
|
|
emit('transcription', {'text': user_text}, room=session_id)
|
|
|
|
# Generate and stream audio response if CSM is available
|
|
if csm_generator is not None:
|
|
# Set up streaming queue for this session
|
|
if session_id not in audio_stream_queues:
|
|
audio_stream_queues[session_id] = queue.Queue()
|
|
else:
|
|
# Clear any existing items in the queue
|
|
while not audio_stream_queues[session_id].empty():
|
|
audio_stream_queues[session_id].get()
|
|
|
|
# Start audio generation in a separate thread to not block the server
|
|
threading.Thread(
|
|
target=generate_and_stream_audio,
|
|
args=(bot_response, context['segments'], session_id),
|
|
daemon=True
|
|
).start()
|
|
|
|
# Initial response with text
|
|
emit('start_streaming_response', {'text': bot_response}, room=session_id)
|
|
else:
|
|
# Send text-only response if audio generation isn't available
|
|
emit('text_response', {'text': bot_response}, room=session_id)
|
|
|
|
# Add text-only bot response to conversation history
|
|
bot_segment = Segment(
|
|
text=bot_response,
|
|
speaker=1, # Bot is speaker 1
|
|
audio=torch.zeros(1) # Placeholder empty audio
|
|
)
|
|
context['segments'].append(bot_segment)
|
|
|
|
except Exception as e:
|
|
print(f"Error processing speech: {e}")
|
|
emit('error', {'message': f'Error processing speech: {str(e)}'}, room=session_id)
|
|
finally:
|
|
# Cleanup temp file
|
|
if os.path.exists(temp_audio_path):
|
|
os.remove(temp_audio_path)
|
|
|
|
def transcribe_with_whisper(audio_path):
|
|
"""Transcribe audio using Faster-Whisper"""
|
|
segments, info = whisper_model.transcribe(audio_path, beam_size=5)
|
|
|
|
# Collect all text from segments
|
|
user_text = ""
|
|
for segment in segments:
|
|
segment_text = segment.text.strip()
|
|
print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment_text}")
|
|
user_text += segment_text + " "
|
|
|
|
print(f"Transcribed text: {user_text.strip()}")
|
|
|
|
return user_text.strip()
|
|
|
|
def transcribe_with_google(audio_path):
|
|
"""Fallback transcription using Google's speech recognition"""
|
|
import speech_recognition as sr
|
|
recognizer = sr.Recognizer()
|
|
|
|
with sr.AudioFile(audio_path) as source:
|
|
audio = recognizer.record(source)
|
|
try:
|
|
text = recognizer.recognize_google(audio)
|
|
return text
|
|
except sr.UnknownValueError:
|
|
return ""
|
|
except sr.RequestError:
|
|
# If Google API fails, try a basic energy-based VAD approach
|
|
# This is a very basic fallback and won't give good results
|
|
return "[Speech detected but transcription failed]"
|
|
|
|
def generate_llm_response(user_text, conversation_segments):
|
|
"""Generate text response using available model"""
|
|
if llm_model is not None and llm_tokenizer is not None:
|
|
# Format conversation history for the LLM
|
|
conversation_history = ""
|
|
for segment in conversation_segments[-5:]: # Use last 5 utterances for context
|
|
speaker_name = "User" if segment.speaker == 0 else "Assistant"
|
|
conversation_history += f"{speaker_name}: {segment.text}\n"
|
|
|
|
# Add the current user query
|
|
conversation_history += f"User: {user_text}\nAssistant:"
|
|
|
|
try:
|
|
# Generate response
|
|
inputs = llm_tokenizer(conversation_history, return_tensors="pt").to(device)
|
|
output = llm_model.generate(
|
|
inputs.input_ids,
|
|
max_new_tokens=150,
|
|
temperature=0.7,
|
|
top_p=0.9,
|
|
do_sample=True
|
|
)
|
|
|
|
response = llm_tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
|
|
return response.strip()
|
|
except Exception as e:
|
|
print(f"Error generating response with LLM: {e}")
|
|
return fallback_response(user_text)
|
|
else:
|
|
return fallback_response(user_text)
|
|
|
|
def fallback_response(user_text):
|
|
"""Generate a simple fallback response when LLM is not available"""
|
|
# Simple rule-based responses
|
|
user_text_lower = user_text.lower()
|
|
|
|
if "hello" in user_text_lower or "hi" in user_text_lower:
|
|
return "Hello! I'm a simple fallback assistant. The main language model couldn't be loaded, so I have limited capabilities."
|
|
|
|
elif "how are you" in user_text_lower:
|
|
return "I'm functioning within my limited capabilities. How can I assist you today?"
|
|
|
|
elif "thank" in user_text_lower:
|
|
return "You're welcome! Let me know if there's anything else I can help with."
|
|
|
|
elif "bye" in user_text_lower or "goodbye" in user_text_lower:
|
|
return "Goodbye! Have a great day!"
|
|
|
|
elif any(q in user_text_lower for q in ["what", "who", "where", "when", "why", "how"]):
|
|
return "I'm running in fallback mode and can't answer complex questions. Please try again when the main language model is available."
|
|
|
|
else:
|
|
return "I understand you said something about that. Unfortunately, I'm running in fallback mode with limited capabilities. Please try again later when the main model is available."
|
|
|
|
def generate_audio_response(text, conversation_segments):
|
|
"""Generate audio response using CSM"""
|
|
try:
|
|
# Use the last few conversation segments as context
|
|
context_segments = conversation_segments[-4:] if len(conversation_segments) > 4 else conversation_segments
|
|
|
|
# Generate audio for bot response
|
|
audio = csm_generator.generate(
|
|
text=text,
|
|
speaker=1, # Bot is speaker 1
|
|
context=context_segments,
|
|
max_audio_length_ms=10000, # 10 seconds max
|
|
temperature=0.9,
|
|
topk=50
|
|
)
|
|
|
|
return audio
|
|
except Exception as e:
|
|
print(f"Error generating audio: {e}")
|
|
# Return silence as fallback
|
|
return torch.zeros(csm_generator.sample_rate * 3) # 3 seconds of silence
|
|
|
|
def generate_and_stream_audio(text, conversation_segments, session_id):
|
|
"""Generate audio response using CSM and stream it in chunks"""
|
|
try:
|
|
# Use the last few conversation segments as context
|
|
context_segments = conversation_segments[-4:] if len(conversation_segments) > 4 else conversation_segments
|
|
|
|
# Generate full audio for bot response
|
|
audio = csm_generator.generate(
|
|
text=text,
|
|
speaker=1, # Bot is speaker 1
|
|
context=context_segments,
|
|
max_audio_length_ms=10000, # 10 seconds max
|
|
temperature=0.9,
|
|
topk=50
|
|
)
|
|
|
|
# Store the full audio for conversation history
|
|
bot_segment = Segment(
|
|
text=text,
|
|
speaker=1, # Bot is speaker 1
|
|
audio=audio
|
|
)
|
|
if session_id in conversation_context:
|
|
conversation_context[session_id]['segments'].append(bot_segment)
|
|
|
|
# Split audio into chunks for streaming
|
|
chunk_size = CHUNK_SIZE
|
|
for i in range(0, len(audio), chunk_size):
|
|
chunk = audio[i:i+chunk_size]
|
|
|
|
# Convert audio chunk to base64 for streaming
|
|
audio_bytes = io.BytesIO()
|
|
torchaudio.save(audio_bytes, chunk.unsqueeze(0).cpu(), csm_generator.sample_rate, format="wav")
|
|
audio_bytes.seek(0)
|
|
audio_b64 = base64.b64encode(audio_bytes.read()).decode('utf-8')
|
|
|
|
# Send the chunk to the client
|
|
if session_id in audio_stream_queues:
|
|
audio_stream_queues[session_id].put({
|
|
'audio': audio_b64,
|
|
'is_last': i + chunk_size >= len(audio)
|
|
})
|
|
else:
|
|
# Session was disconnected before we finished generating
|
|
break
|
|
|
|
# Signal the end of streaming if queue still exists
|
|
if session_id in audio_stream_queues:
|
|
# Add an empty chunk as a sentinel to signal end of streaming
|
|
audio_stream_queues[session_id].put(None)
|
|
|
|
except Exception as e:
|
|
print(f"Error generating or streaming audio: {e}")
|
|
# Send error message to client
|
|
if session_id in conversation_context:
|
|
socketio.emit('error', {
|
|
'message': f'Error generating audio: {str(e)}'
|
|
}, room=session_id)
|
|
|
|
# Send a final message to unblock the client
|
|
if session_id in audio_stream_queues:
|
|
audio_stream_queues[session_id].put(None)
|
|
|
|
@socketio.on('request_audio_chunk')
|
|
def handle_request_audio_chunk():
|
|
"""Send the next audio chunk in the queue to the client"""
|
|
session_id = request.sid
|
|
|
|
if session_id not in audio_stream_queues:
|
|
emit('error', {'message': 'No audio stream available'})
|
|
return
|
|
|
|
# Get the next chunk or wait for it to be available
|
|
try:
|
|
if not audio_stream_queues[session_id].empty():
|
|
chunk = audio_stream_queues[session_id].get(block=False)
|
|
|
|
# If chunk is None, we're done streaming
|
|
if chunk is None:
|
|
emit('end_streaming')
|
|
# Clean up the queue
|
|
if session_id in audio_stream_queues:
|
|
del audio_stream_queues[session_id]
|
|
else:
|
|
emit('audio_chunk', chunk)
|
|
else:
|
|
# If the queue is empty but we're still generating, tell client to wait
|
|
emit('wait_for_chunk')
|
|
except Exception as e:
|
|
print(f"Error sending audio chunk: {e}")
|
|
emit('error', {'message': f'Error streaming audio: {str(e)}'})
|
|
|
|
if __name__ == '__main__':
|
|
# Ensure the existing index.html file is in the correct location
|
|
if not os.path.exists('templates'):
|
|
os.makedirs('templates')
|
|
|
|
if os.path.exists('index.html') and not os.path.exists('templates/index.html'):
|
|
os.rename('index.html', 'templates/index.html')
|
|
|
|
# Load models asynchronously before starting the server
|
|
print("Starting model loading...")
|
|
load_models()
|
|
|
|
# Start the server
|
|
print("Starting Flask SocketIO server...")
|
|
socketio.run(app, host='0.0.0.0', port=5000, debug=False) |