test

README.md

@@ -155,6 +155,40 @@ The C3PO model supports all XTTS-v2 languages:
 
 ## Setup
 
+### CPU-Only Installation (Recommended for most users)
+
+For CPU-only usage (no GPU required):
+```bash
+# Ubuntu/Debian
+sudo apt-get install espeak-ng ffmpeg git git-lfs
+
+# macOS
+brew install espeak ffmpeg git git-lfs
+```
+
+2. **Install CPU-only PyTorch and dependencies:**
+```bash
+# Option 1: Use the provided script
+chmod +x install_cpu.sh
+./install_cpu.sh
+
+# Option 2: Manual installation
+pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu
+pip install -r requirements.txt
+python -m unidic download
+```
+
+3. **Set CPU-only environment variables:**
+```bash
+export FORCE_CPU=true
+export CUDA_VISIBLE_DEVICES=""
+```
+
+4. **Run the API:**
+```bash
+uvicorn app:app --host 0.0.0.0 --port 7860
+```
+
 ### Hugging Face Spaces Deployment
 
 This API is optimized for Hugging Face Spaces with:
@@ -255,6 +289,13 @@ Automatically configured:
 
 ## Troubleshooting
 
+### CPU Performance
+When running on CPU:
+- Speech generation will be slower than GPU (30-60 seconds vs 3-5 seconds)
+- Memory usage is lower (2-4GB RAM vs 4-8GB VRAM)
+- No CUDA installation required
+- Works on any system with sufficient RAM
+
 ### PyTorch Loading Issues
 The API includes fixes for PyTorch 2.6's `weights_only=True` default. If you encounter loading issues, ensure the compatibility fix is applied.
 
@@ -270,9 +311,16 @@ If the C3PO model fails to download:
 - Ensure reference audio is 3-10 seconds long
 
 ### Memory Issues
-- Use CPU mode for lower memory usage: set `CUDA_VISIBLE_DEVICES=""`
+- **CPU Mode**: Requires 2-4GB RAM, works on most modern computers
+- **GPU Mode**: Requires 4GB+ VRAM for optimal performance
 - Reduce text length for batch processing
-- Consider using GPU with sufficient VRAM (4GB+ recommended)
+- Use CPU mode with `FORCE_CPU=true` environment variable
+
+### CPU-Only Installation Issues
+If you encounter GPU-related errors:
+1. Set environment variables: `export FORCE_CPU=true CUDA_VISIBLE_DEVICES=""`
+2. Install CPU-only PyTorch: `pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu`
+3. Restart the API after setting environment variables
 
 ## License
 

requirements.txt

@@ -7,11 +7,5 @@ mecab-python3==1.0.6
 unidic-lite==1.0.8
 unidic==1.1.0
 langid
-pydub
-fastapi
-uvicorn[standard]
-torch
-torchaudio
-soundfile
-scipy
-numpy
+uvicorn
+pydub

test.py

@@ -3,15 +3,17 @@ import torch
 import torchaudio
 import subprocess
 
+# Set environment variables for CPU-only usage
+os.environ['COQUI_TOS_AGREED'] = '1'
+os.environ['NUMBA_DISABLE_JIT'] = '1'
+os.environ['FORCE_CPU'] = 'true'
+os.environ['CUDA_VISIBLE_DEVICES'] = ''
+
 # Fix PyTorch weights_only issue for XTTS
 import torch.serialization
 from TTS.tts.configs.xtts_config import XttsConfig
 torch.serialization.add_safe_globals([XttsConfig])
 
-# Set environment variables
-os.environ['COQUI_TOS_AGREED'] = '1'
-os.environ['NUMBA_DISABLE_JIT'] = '1'
-
 from TTS.api import TTS
 from TTS.tts.configs.xtts_config import XttsConfig
 from TTS.tts.models.xtts import Xtts
@@ -50,11 +52,8 @@ model.load_checkpoint(
     eval=True,
 )
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-if device == "cuda":
-    model.cuda()
-
-print(f"C3PO model loaded on {device}")
+device = "cpu"  # Force CPU usage
+print(f"C3PO model loaded on {device} (forced CPU mode)")
 
 # Text to convert to speech
 text = "Hello there! I am C-3PO, human-cyborg relations. How may I assist you today?"