Switch to offline pyttsx3 TTS for Hugging Face CPU Space

Dockerfile

@@ -1,83 +1,121 @@
+# # ================================================================
+# #   Avatar App - SadTalker + Coqui TTS (CPU version, HF-ready)
+# #   Optimized for Hugging Face Spaces Free Tier
+# # ================================================================
+
+# # ---------- Base Image ----------
+# FROM python:3.10-slim
+
+# ENV DEBIAN_FRONTEND=noninteractive
+# WORKDIR /app
+
+# # ---------- System Dependencies ----------
+# RUN apt-get update && apt-get install -y \
+#     ffmpeg git wget curl build-essential \
+#     && rm -rf /var/lib/apt/lists/*
+
+# # ---------- Environment Variables ----------
+# ENV PYTHONUNBUFFERED=1
+# ENV COQUI_TOS_AGREED=1
+# ENV CUDA_VISIBLE_DEVICES=""
+# ENV DEVICE=cpu
+# ENV PYTORCH_ENABLE_MPS_FALLBACK=1
+
+# # ---------- Install Python Dependencies (Memory-optimized order) ----------
+# RUN pip install --no-cache-dir --upgrade pip setuptools wheel
+
+# # Install NumPy with version that satisfies all dependencies
+# RUN pip install --no-cache-dir "numpy>=1.23.5,<1.24"
+
+# # CPU-only PyTorch (use version compatible with TTS library)
+# # PyTorch 2.6+ breaks TTS due to weights_only=True default, so use 2.5.1
+# RUN pip install --no-cache-dir torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cpu
+
+# # Core dependencies - install in smaller batches
+# RUN pip install --no-cache-dir scipy==1.11.3
+# RUN pip install --no-cache-dir scikit-learn==1.5.0
+# RUN pip install --no-cache-dir librosa==0.10.2.post1
+# RUN pip install --no-cache-dir Cython==3.0.10
+
+# # Streamlit and lighter dependencies
+# RUN pip install --no-cache-dir \
+#     streamlit \
+#     pydub \
+#     matplotlib \
+#     requests \
+#     python-dotenv
+
+# # ---------- Clone SadTalker ----------
+# RUN git clone --depth 1 https://github.com/OpenTalker/SadTalker.git /app/SadTalker
+
+# # Install SadTalker requirements with correct NumPy
+# RUN cd /app/SadTalker && \
+#     sed -i 's/numpy.*/numpy>=1.23.5,<1.24/' requirements.txt && \
+#     pip install --no-cache-dir -r requirements.txt
+
+# # ---------- Install Coqui TTS Stack (in stages to manage memory) ----------
+# RUN pip install --no-cache-dir transformers==4.36.2
+# RUN pip install --no-cache-dir sentencepiece==0.2.0
+# RUN pip install --no-cache-dir accelerate==0.25.0
+# RUN pip install --no-cache-dir TTS==0.22.0
+
+# # Fix any NumPy version conflicts
+# RUN pip install --no-cache-dir --force-reinstall "numpy>=1.23.5,<1.24"
+
+# # ---------- Copy Application Files ----------
+# COPY avatar_streamlit.py /app/
+
+# # ---------- Fix NumPy Compatibility in SadTalker ----------
+# RUN cd /app/SadTalker && \
+#     find . -name "*.py" -type f -exec sed -i 's/\bnp\.float\b/np.float64/g' {} + && \
+#     sed -i 's/warnings.filterwarnings("ignore", category=np.VisibleDeprecationWarning)/# warnings.filterwarnings("ignore", category=np.VisibleDeprecationWarning)/' src/face3d/util/preprocess.py && \
+#     sed -i 's/trans_params = np.array(\[w0, h0, s, t\[0\], t\[1\]\])/trans_params = np.array([w0, h0, s, float(t[0]), float(t[1])])/' src/face3d/util/preprocess.py && \
+#     echo "✅ SadTalker fixes applied!"
+
+# # ---------- Preload Coqui Model (Skip on build to save memory) ----------
+# # Model will download on first run instead
+# # RUN python3 -c "from TTS.api import TTS; TTS('tts_models/multilingual/multi-dataset/xtts_v2', gpu=False)" || true
+
+# # ---------- Expose Port and Run ----------
+# EXPOSE 8501
+# CMD ["streamlit", "run", "avatar_streamlit.py", "--server.port=8501", "--server.address=0.0.0.0"]
+
+# # FORCE_REBUILD: 2025-01-04-v5
 # ================================================================
-#   Avatar App - SadTalker + Coqui TTS (CPU version, HF-ready)
-#   Optimized for Hugging Face Spaces Free Tier
+#   Avatar App - SadTalker + pyttsx3 (Offline CPU version)
+#   Works fully offline on Hugging Face (Free CPU tier)
 # ================================================================
 
-# ---------- Base Image ----------
 FROM python:3.10-slim
 
 ENV DEBIAN_FRONTEND=noninteractive
 WORKDIR /app
 
-# ---------- System Dependencies ----------
+# ---------- System dependencies ----------
 RUN apt-get update && apt-get install -y \
-    ffmpeg git wget curl build-essential \
+    ffmpeg git wget curl build-essential espeak \
     && rm -rf /var/lib/apt/lists/*
 
-# ---------- Environment Variables ----------
+# ---------- Copy source files ----------
+COPY . /app
+
+# ---------- Environment ----------
 ENV PYTHONUNBUFFERED=1
-ENV COQUI_TOS_AGREED=1
 ENV CUDA_VISIBLE_DEVICES=""
 ENV DEVICE=cpu
-ENV PYTORCH_ENABLE_MPS_FALLBACK=1
-
-# ---------- Install Python Dependencies (Memory-optimized order) ----------
-RUN pip install --no-cache-dir --upgrade pip setuptools wheel
-
-# Install NumPy with version that satisfies all dependencies
-RUN pip install --no-cache-dir "numpy>=1.23.5,<1.24"
 
-# CPU-only PyTorch (use version compatible with TTS library)
-# PyTorch 2.6+ breaks TTS due to weights_only=True default, so use 2.5.1
-RUN pip install --no-cache-dir torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cpu
-
-# Core dependencies - install in smaller batches
-RUN pip install --no-cache-dir scipy==1.11.3
-RUN pip install --no-cache-dir scikit-learn==1.5.0
-RUN pip install --no-cache-dir librosa==0.10.2.post1
-RUN pip install --no-cache-dir Cython==3.0.10
-
-# Streamlit and lighter dependencies
-RUN pip install --no-cache-dir \
-    streamlit \
-    pydub \
-    matplotlib \
-    requests \
-    python-dotenv
+# ---------- Python dependencies ----------
+RUN pip install --no-cache-dir --upgrade pip setuptools wheel \
+    && pip install --no-cache-dir -r requirements.txt
 
 # ---------- Clone SadTalker ----------
-RUN git clone --depth 1 https://github.com/OpenTalker/SadTalker.git /app/SadTalker
-
-# Install SadTalker requirements with correct NumPy
-RUN cd /app/SadTalker && \
-    sed -i 's/numpy.*/numpy>=1.23.5,<1.24/' requirements.txt && \
-    pip install --no-cache-dir -r requirements.txt
-
-# ---------- Install Coqui TTS Stack (in stages to manage memory) ----------
-RUN pip install --no-cache-dir transformers==4.36.2
-RUN pip install --no-cache-dir sentencepiece==0.2.0
-RUN pip install --no-cache-dir accelerate==0.25.0
-RUN pip install --no-cache-dir TTS==0.22.0
-
-# Fix any NumPy version conflicts
-RUN pip install --no-cache-dir --force-reinstall "numpy>=1.23.5,<1.24"
-
-# ---------- Copy Application Files ----------
-COPY avatar_streamlit.py /app/
-
-# ---------- Fix NumPy Compatibility in SadTalker ----------
-RUN cd /app/SadTalker && \
-    find . -name "*.py" -type f -exec sed -i 's/\bnp\.float\b/np.float64/g' {} + && \
-    sed -i 's/warnings.filterwarnings("ignore", category=np.VisibleDeprecationWarning)/# warnings.filterwarnings("ignore", category=np.VisibleDeprecationWarning)/' src/face3d/util/preprocess.py && \
-    sed -i 's/trans_params = np.array(\[w0, h0, s, t\[0\], t\[1\]\])/trans_params = np.array([w0, h0, s, float(t[0]), float(t[1])])/' src/face3d/util/preprocess.py && \
-    echo "✅ SadTalker fixes applied!"
-
-# ---------- Preload Coqui Model (Skip on build to save memory) ----------
-# Model will download on first run instead
-# RUN python3 -c "from TTS.api import TTS; TTS('tts_models/multilingual/multi-dataset/xtts_v2', gpu=False)" || true
+RUN git clone https://github.com/OpenTalker/SadTalker.git /app/SadTalker \
+    && pip install --no-cache-dir -r /app/SadTalker/requirements.txt \
+    && pip uninstall -y torch torchvision torchaudio \
+    && pip install --no-cache-dir torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
 
-# ---------- Expose Port and Run ----------
+# ---------- Expose and run ----------
 EXPOSE 8501
 CMD ["streamlit", "run", "avatar_streamlit.py", "--server.port=8501", "--server.address=0.0.0.0"]
 
-# FORCE_REBUILD: 2025-01-04-v5
+# FORCE_REBUILD: 2025-10-30_21-00

avatar_streamlit.py

@@ -58,17 +58,48 @@ def ask_ollama(question: str) -> str:
 #     return mp3_path
 
 
+# def synthesize_speech(text, lang="he"):
+#     """Offline GPU TTS using Coqui XTTS v2."""
+#     print("🗣️ Generating speech with Coqui TTS...")
+#     model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
+#     tts = TTS(model_name, gpu=False)
+#     fd, wav_path = tempfile.mkstemp(suffix=".wav")
+#     os.close(fd)
+#     tts.tts_to_file(text=text, file_path=wav_path, language=lang)
+#     print("✅ Saved audio to:", wav_path)
+#     return wav_path
+
+import pyttsx3
+import tempfile
+import os
+
 def synthesize_speech(text, lang="he"):
-    """Offline GPU TTS using Coqui XTTS v2."""
-    print("🗣️ Generating speech with Coqui TTS...")
-    model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
-    tts = TTS(model_name, gpu=False)
+    """Offline speech synthesis using pyttsx3 (no internet)."""
+    print("🗣️ Generating speech locally with pyttsx3...")
     fd, wav_path = tempfile.mkstemp(suffix=".wav")
     os.close(fd)
-    tts.tts_to_file(text=text, file_path=wav_path, language=lang)
-    print("✅ Saved audio to:", wav_path)
+
+    engine = pyttsx3.init()
+    voices = engine.getProperty("voices")
+
+    # Try to use Hebrew if available
+    selected = False
+    for voice in voices:
+        if "he" in voice.id.lower() or "hebrew" in voice.name.lower():
+            engine.setProperty("voice", voice.id)
+            selected = True
+            print(f"✅ Using Hebrew voice: {voice.name}")
+            break
+
+    if not selected:
+        print("⚠️ No Hebrew voice found, using default voice.")
+
+    engine.save_to_file(text, wav_path)
+    engine.runAndWait()
+    print("✅ Speech saved to:", wav_path)
     return wav_path
 
+
 # --- Check SadTalker availability ---
 SADTALKER_AVAILABLE = True  # set False if not installed
 # ========== SadTalker Integration ==========

requirements.txt

@@ -1,13 +1,23 @@
+# numpy==1.23.5
+# scipy==1.11.3
+# scikit-learn==1.5.0
+# librosa==0.10.2.post1
+# streamlit
+# TTS==0.22.0
+# transformers==4.36.2
+# sentencepiece==0.2.0
+# accelerate==0.25.0
+# pydub
+# matplotlib
+# requests
+# python-dotenv
+streamlit
+pyttsx3
+pydub
+matplotlib
 numpy==1.23.5
 scipy==1.11.3
 scikit-learn==1.5.0
 librosa==0.10.2.post1
-streamlit
-TTS==0.22.0
-transformers==4.36.2
-sentencepiece==0.2.0
-accelerate==0.25.0
-pydub
-matplotlib
 requests
-python-dotenv
+python-dotenv