voice clone config and improved installation/launch

.gitignore

@@ -96,5 +96,5 @@ data/voices/private
 # Keep Example Configs (negation patterns)
 # =============================================================================
 !data/config.example.yml
-!.env.example
+!.env.example\
 publish.sh

CLAUDE.md

@@ -11,6 +11,9 @@ Fulloch (the **Full**y **Loc**al **H**ome voice assistant) is a fully local, pri
 ### Development
 ```bash
 pip install -r requirements.txt
+# Install special packages (see requirements.txt for details)
+pip install --no-deps git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215
+# GPU only: pip install --no-build-isolation --no-deps git+https://github.com/Dao-AILab/flash-attention.git@ef9e6a6
 pip install -e ".[dev]"  # Install with dev dependencies
 python app.py
 ```
@@ -86,9 +89,10 @@ SILENCE_THRESHOLD = 0.001    # RMS threshold (lower = more sensitive)
 ```
 
 ### Config Files (not in git)
-- `data/config.yml`: Service endpoints, wakeword, integration settings
+- `data/config.yml`: Service endpoints, wakeword, voice_clone, integration settings
 - `.env`: Credentials (Spotify, Google, etc.)
 - `data/models/`: Local model cache (~4-5GB)
+- `data/voices/`: Voice clone reference files (wav/txt pairs)
 
 ### Example Config Files (in git)
 - `data/config.example.yml`: Template with all settings documented

CONTRIBUTING.md

@@ -11,6 +11,9 @@ Thank you for your interest in contributing to Fulloch! This document provides g
    python -m venv .venv
    source .venv/bin/activate  # On Windows: .venv\Scripts\activate
    pip install -r requirements.txt
+   # Install special packages (see requirements.txt for details)
+   pip install --no-deps git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215
+   # GPU only: pip install --no-build-isolation --no-deps git+https://github.com/Dao-AILab/flash-attention.git@ef9e6a6
    pip install -e ".[dev]"  # Install dev dependencies
    ```
 4. Copy configuration files:

Dockerfile

@@ -2,16 +2,25 @@ FROM python:3.12-slim
 
 WORKDIR /app
 
-# Install dependencies first to leverage Docker cache
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    sox \
+    libsox-dev \
+    libsox-fmt-all \
+    ffmpeg \
+    git \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Python dependencies
 COPY requirements.txt .
-RUN apt install sox libsox-dev libsox-fmt-all ffmpeg
-RUN pip install --no-cache-dir -r requirements.txt
+RUN pip install --no-cache-dir -r requirements.txt && \
+    pip install --no-deps git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215
 
 # Copy application code
 COPY app.py .
+COPY core/ core/
 COPY tools/ tools/
 COPY utils/ utils/
-COPY wav/ wav/
 COPY audio/ audio/
 
 # Run the app

Dockerfile_gpu

@@ -1,45 +1,35 @@
-# Use NVIDIA CUDA base image (includes nvcc compiler for building)
-FROM nvidia/cuda:12.4.1-devel-ubuntu22.04
+# Use PyTorch image with CUDA support (includes Python 3.11, PyTorch, CUDA)
+FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel
 
-# Avoid interactive prompts during package installation
+# Avoid interactive prompts
 ENV DEBIAN_FRONTEND=noninteractive
 
-# Install Python 3.12 and build tools
+# Install system dependencies
 RUN apt-get update && apt-get install -y \
-    software-properties-common \
-    && add-apt-repository ppa:deadsnakes/ppa \
-    && apt-get update && apt-get install -y \
-    python3.12 \
-    python3.12-venv \
-    python3.12-dev \
-    python3-pip \
     git \
-    build-essential \
-    cmake \
+    sox \
+    libsox-dev \
+    libsox-fmt-all \
+    ffmpeg \
     && rm -rf /var/lib/apt/lists/*
 
-# Set Python 3.12 as default
-RUN ln -s /usr/bin/python3.12 /usr/bin/python
-
 WORKDIR /app
 
-# 2. Set Environment Variables to force CUDA build
-# -DGGML_CUDA=on is the flag for recent llama-cpp-python versions (0.3.x)
+# Set Environment Variables for llama-cpp-python CUDA build
 ENV CMAKE_ARGS="-DGGML_CUDA=on"
 ENV FORCE_CMAKE=1
 
+# Install Python dependencies
 COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt && \
+    pip install --no-deps git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215 && \
+    pip install flash-attn --no-build-isolation
 
-# 3. Install dependencies
-# This will now compile llama-cpp-python with CUDA support
-RUN python3.12 -m pip install --upgrade pip && \
-    python3.12 -m pip install --no-cache-dir -r requirements.txt
-
-# Copy application files (ignoring __pycache__ via .dockerignore)
+# Copy application code
 COPY app.py .
+COPY core/ core/
 COPY tools/ tools/
 COPY utils/ utils/
-COPY wav/ wav/
 COPY audio/ audio/
 
-CMD ["python3.12", "app.py"]
+CMD ["python", "app.py"]

SECURITY.md

@@ -4,9 +4,9 @@
 
 Fulloch is designed with privacy as a core principle. All processing happens locally on your device:
 
-- **Speech Recognition**: Moonshine ASR runs entirely on-device
-- **Text-to-Speech**: Kokoro TTS runs entirely on-device
-- **Language Model**: Qwen runs entirely on-device via llama.cpp
+- **Speech Recognition**: Qwen3 ASR runs entirely on-device (or Moonshine Tiny for edge devices)
+- **Text-to-Speech**: Qwen3 TTS with voice cloning runs entirely on-device (or Kokoro for edge devices)
+- **Language Model**: Qwen 3 4B runs entirely on-device via llama.cpp
 - **No Cloud Dependencies**: No data is sent to external servers for AI processing
 
 ## Reporting a Vulnerability

app.py

@@ -31,7 +31,10 @@ os.environ["VLLM_NO_USAGE_STATS"] = "1"
 
 import logging
 
-logging.basicConfig(level=logging.INFO)
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s"
+)
 logger = logging.getLogger(__name__)
 
 from core.assistant import Assistant
@@ -41,10 +44,11 @@ WAKEWORD = config['general']['wakeword']
 USE_AI = config['general']['use_ai']
 USE_TINY_ASR = config['general'].get('use_tiny_asr', False)
 USE_TINY_TTS = config['general'].get('use_tiny_tts', False)
+VOICE_CLONE = config['general'].get('voice_clone', None)
 
 def main():
     """Main entry point for the voice assistant."""
-    assistant = Assistant(wakeword=WAKEWORD, use_ai=USE_AI, use_tiny_asr=USE_TINY_ASR, use_tiny_tts=USE_TINY_TTS)
+    assistant = Assistant(wakeword=WAKEWORD, use_ai=USE_AI, use_tiny_asr=USE_TINY_ASR, use_tiny_tts=USE_TINY_TTS, voice_clone=VOICE_CLONE)
     assistant.run()
 
 

core/assistant.py

@@ -31,7 +31,7 @@ class Assistant:
         grammar: JSON grammar for structured output
     """
 
-    def __init__(self, wakeword: str, use_ai: bool, use_tiny_asr: bool = False, use_tiny_tts: bool = False):
+    def __init__(self, wakeword: str, use_ai: bool, use_tiny_asr: bool = False, use_tiny_tts: bool = False, voice_clone: str = "cori"):
         """
         Initialize the assistant.
 
@@ -40,11 +40,13 @@ class Assistant:
             use_ai: Whether to use the SLM for intent detection
             use_tiny_asr: Whether to use Moonshine Tiny ASR instead of Qwen ASR
             use_tiny_tts: Whether to use Kokoro TTS instead of Qwen TTS
+            voice_clone: Name of the voice clone wav/txt file to use for TTS
         """
         self.wakeword = wakeword.lower()
         self.use_ai = use_ai
         self.use_tiny_asr = use_tiny_asr
         self.use_tiny_tts = use_tiny_tts
+        self.voice_clone = voice_clone
         self.audio_capture = AudioCapture()
 
         # Models loaded lazily in transcriber thread
@@ -54,6 +56,7 @@ class Assistant:
         self.grammar = None
         self.intent_prompt = None
         self.chat_prompt = None
+        self.voice_prompt = None
 
     def _load_models(self):
         """Load ASR, TTS and SLM models."""
@@ -71,8 +74,10 @@ class Assistant:
             from .tts_tiny import speak_stream, remove_emoji
             logger.info("Using Kokoro TTS")
         else:
-            from .tts import speak_stream, remove_emoji
-            logger.info("Using Qwen TTS")
+            from .tts import speak_stream, remove_emoji, set_voice, warmup_model
+            self.voice_prompt = set_voice(self.voice_clone)
+            logger.info(f"Using Qwen TTS with voice clone: {self.voice_clone}")
+            warmup_model(self.voice_prompt)
         
         self.remove_emoji = remove_emoji
         self.speak_stream = speak_stream
@@ -142,7 +147,7 @@ class Assistant:
                     "Just a second.",
                     "Got it, let me think.",
                     "Let's see."
-                ]))
+                ]), self.voice_prompt)
 
                 logger.info(f"AI chat query: {user_prompt}")
                 answer = generate_slm(
@@ -198,7 +203,7 @@ class Assistant:
                 # Process and respond
                 answer = self._handle_wakeword(user_prompt)
                 cleaned = self.remove_emoji(answer.replace('"', '').replace('*', ''))
-                self.speak_stream(cleaned)
+                self.speak_stream(cleaned, self.voice_prompt)
 
                 # Resume transcription
                 self.audio_capture.transcribing = True

core/tts.py

@@ -8,13 +8,16 @@ import queue
 import re
 import threading
 import torch
+
+# Set precision before any CUDA operations
+torch.set_float32_matmul_precision('high')
+
 import sounddevice as sd
 from qwen_tts import Qwen3TTSModel
 
 logger = logging.getLogger(__name__)
 
-REF_AUDIO = "./data/voices/cori.wav"
-REF_TEXT = "./data/voices/cori.txt"
+VOICES_DIR = "./data/voices"
 
 # Device configuration
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
@@ -43,10 +46,7 @@ def remove_emoji(text: str, rem_think: bool = True) -> str:
 
     return EMOJI_PATTERN.sub("", text)
 
-# Get reference audio text
-with open(REF_TEXT) as f:
-    ref_text = f.read()
- 
+
 # Load model
 model = Qwen3TTSModel.from_pretrained(
     "Qwen/Qwen3-TTS-12Hz-1.7B-Base",
@@ -62,35 +62,52 @@ model.enable_streaming_optimizations(
     compile_mode="reduce-overhead",  # Includes CUDA graphs automatically
 )
 
-# Create voice clone prompt from reference audio
-prompt = model.create_voice_clone_prompt(
-    ref_audio=REF_AUDIO,
-    ref_text=ref_text,
-)
+
+def set_voice(voice_name: str):
+    """
+    Load voice clone prompt from audio/text files to use for TTS.
+
+    Args:
+        voice_name: Name of the voice (matches wav/txt files in data/voices)
+    """
+    ref_audio = f"{VOICES_DIR}/{voice_name}.wav"
+    ref_text_path = f"{VOICES_DIR}/{voice_name}.txt"
+
+    with open(ref_text_path) as f:
+        ref_text = f.read()
+
+    logger.info(f"Setting voice clone to: {voice_name}")
+    return model.create_voice_clone_prompt(
+        ref_audio=ref_audio,
+        ref_text=ref_text,
+    )
+
 
 # Warmup: run dummy generation to initialize torch.compile and CUDA graphs
-logger.info("Warming up TTS model...")
-for _ in model.stream_generate_voice_clone(
-    # Reference text must be longer to properly initialise model
-    text="A rainbow is a meteorological phenomenon that is caused by reflection, refraction and dispersion of light in water droplets resulting in a spectrum of light appearing in the sky.",
-    language="english",
-    voice_clone_prompt=prompt,
-    overlap_samples=512,
-    emit_every_frames=12,
-    decode_window_frames=80,
-    first_chunk_emit_every=5,
-    first_chunk_decode_window=48,
-    first_chunk_frames=48,
-):
-    pass  # Discard output
-logger.info("TTS model ready")
+def warmup_model(prompt):
+    logger.info("Warming up TTS model...")
+    for _ in model.stream_generate_voice_clone(
+        # Reference text must be longer to properly initialise model
+        text="A rainbow is a meteorological phenomenon that is caused by reflection, refraction and dispersion of light in water droplets.",
+        language="english",
+        voice_clone_prompt=prompt,
+        overlap_samples=512,
+        emit_every_frames=12,
+        decode_window_frames=80,
+        first_chunk_emit_every=5,
+        first_chunk_decode_window=48,
+        first_chunk_frames=48,
+    ):
+        pass  # Discard output
+    logger.info("TTS model ready")
 
-def speak_stream(text: str, voice: str = "cori", speed: float = 1.0):
+def speak_stream(text: str, prompt, voice: str = "cori", speed: float = 1.0):
     """
     Generate speech from text using stream optimised Qwen3 TTS from rekuenkdr
 
     Args:
         text: Text to synthesize
+        prompt: Prepared voice cloning prompt
         voice: Not used
         speed: Not used
     """

core/tts_tiny.py

@@ -53,12 +53,13 @@ def remove_emoji(text: str, rem_think: bool = True) -> str:
     return EMOJI_PATTERN.sub("", text)
 
 
-def speak_stream(text: str, voice: str = "af_bella", speed: float = 1.2):
+def speak_stream(text: str, prompt=None, voice: str = "af_bella", speed: float = 1.2):
     """
     Generate speech from text using Kokoro and stream to speakers.
 
     Args:
         text: Text to synthesize
+        prompt: Not used
         voice: Voice model to use (default: af_bella)
         speed: Speech speed multiplier (default: 1.2)
     """

data/config.example.yml

@@ -15,6 +15,8 @@ general:
   use_tiny_asr: False
   # Use the Kokoro TTS for faster voice streaming (voice cloning not available)
   use_tiny_tts: False
+  # Set name of wav/txt file if using voice cloning
+  voice_clone: "cori"
 
 # =============================================================================
 # Spotify Integration

launch.sh

@@ -1,6 +1,14 @@
 #!/bin/bash
 set -e
 
+# Helper function to ask user for confirmation
+ask_download() {
+    local model_name="$1"
+    read -p "Download $model_name? (y/n): " response
+    response=${response,,}
+    [[ "$response" == "y" || "$response" == "yes" ]]
+}
+
 # Directory definitions
 BASE_DIR="$(pwd)/data/models"
 GRAMMAR_DIR="$BASE_DIR/grammars"
@@ -26,74 +34,115 @@ fi
 echo "📂 Checking directory structure..."
 mkdir -p "$HUB_DIR"
 
+# 2a. Check for config.yml
+CONFIG_FILE="$(pwd)/data/config.yml"
+CONFIG_EXAMPLE="$(pwd)/data/config.example.yml"
+
+if [ ! -f "$CONFIG_FILE" ]; then
+    echo "📝 config.yml not found. Creating from template..."
+    cp "$CONFIG_EXAMPLE" "$CONFIG_FILE"
+    echo ""
+    echo "⚠️  Please edit data/config.yml with your settings before continuing."
+    echo "   See data/config.example.yml for documentation on each option."
+    echo ""
+    echo "   Run ./launch.sh again when ready."
+    exit 0
+else
+    echo "✅ config.yml exists."
+fi
+
 # 3. Check and Download json.gbnf
 if [ ! -f "$GRAMMAR_DIR/json.gbnf" ]; then
-echo "⬇️ Downloading json.gbnf..."
-wget -q --show-progress -O "$GRAMMAR_DIR/json.gbnf" \
-"https://raw.githubusercontent.com/ggml-org/llama.cpp/master/grammars/json.gbnf"
+    if ask_download "json.gbnf (grammar file)"; then
+        echo "⬇️ Downloading json.gbnf..."
+        wget -q --show-progress -O "$GRAMMAR_DIR/json.gbnf" \
+        "https://raw.githubusercontent.com/ggml-org/llama.cpp/master/grammars/json.gbnf"
+    else
+        echo "⏭️ Skipping json.gbnf"
+    fi
 else
-echo "✅ json.gbnf exists."
+    echo "✅ json.gbnf exists."
 fi
 
 # 4. Check and Download Qwen3 GGUF
 if [ ! -f "$BASE_DIR/$QWEN_FILE" ]; then
-    echo "⬇️ Downloading $QWEN_FILE..."
-    huggingface-cli download "$QWEN_REPO" "$QWEN_FILE" \
-        --local-dir "$BASE_DIR" \
-        --local-dir-use-symlinks False
+    if ask_download "Qwen3 4B SLM (2.5GB)"; then
+        echo "⬇️ Downloading $QWEN_FILE..."
+        huggingface-cli download "$QWEN_REPO" "$QWEN_FILE" \
+            --local-dir "$BASE_DIR" \
+            --local-dir-use-symlinks False
+    else
+        echo "⏭️ Skipping Qwen3 SLM"
+    fi
 else
     echo "✅ $QWEN_FILE exists."
 fi
 
 # 5. Check and Download Qwen3 TTS model
 if [ ! -d "$TTS_DIR" ]; then
-echo "⬇️ Downloading Qwen3 TTS..."
-huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-Base \
---cache-dir "$HUB_DIR"
+    if ask_download "Qwen3 TTS (3.4GB)"; then
+        echo "⬇️ Downloading Qwen3 TTS..."
+        huggingface-cli download Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+            --cache-dir "$HUB_DIR"
+    else
+        echo "⏭️ Skipping Qwen3 TTS"
+    fi
 else
-echo "✅ Qwen3 TTS model exists."
+    echo "✅ Qwen3 TTS model exists."
 fi
 
 # 5a. Check and Download Kokoro-82M (TTS)
 if [ ! -d "$TTS_TINY_DIR" ]; then
-echo "⬇️ Downloading Kokoro-82M..."
-huggingface-cli download hexgrad/Kokoro-82M \
---cache-dir "$HUB_DIR"
+    if ask_download "Kokoro-82M TTS Tiny (200MB)"; then
+        echo "⬇️ Downloading Kokoro-82M..."
+        huggingface-cli download hexgrad/Kokoro-82M \
+            --cache-dir "$HUB_DIR"
+    else
+        echo "⏭️ Skipping Kokoro-82M"
+    fi
 else
-echo "✅ Kokoro-82M exists."
+    echo "✅ Kokoro-82M exists."
 fi
 
 # 6. Check and Download Qwen3 ASR model
 if [ ! -d "$ASR_DIR" ]; then
-    echo "⬇️ Downloading Qwen3 ASR..."
-    huggingface-cli download Qwen/Qwen3-ASR-1.7B \
-        --cache-dir "$HUB_DIR"
+    if ask_download "Qwen3 ASR (3.4GB)"; then
+        echo "⬇️ Downloading Qwen3 ASR..."
+        huggingface-cli download Qwen/Qwen3-ASR-1.7B \
+            --cache-dir "$HUB_DIR"
+    else
+        echo "⏭️ Skipping Qwen3 ASR"
+    fi
 else
     echo "✅ Qwen3 ASR model exists."
 fi
 
 # 6a. Check and Download Moonshine Tiny ASR model
 if [ ! -d "$ASR_TINY_DIR" ]; then
-echo "⬇️ Downloading Moonshine Tiny..."
-huggingface-cli download UsefulSensors/moonshine-tiny \
---cache-dir "$HUB_DIR"
+    if ask_download "Moonshine Tiny ASR (60MB)"; then
+        echo "⬇️ Downloading Moonshine Tiny..."
+        huggingface-cli download UsefulSensors/moonshine-tiny \
+            --cache-dir "$HUB_DIR"
+    else
+        echo "⏭️ Skipping Moonshine Tiny"
+    fi
 else
-echo "✅ Moonshine-tiny exists."
+    echo "✅ Moonshine-tiny exists."
 fi
 
-# # 7. Prompt the user
-# read -p "Are you using a GPU? (y/n): " response
-# response=${response,,}
-# if [[ "$response" == "y" || "$response" == "yes" ]]; then
-#     mv Dockerfile Dockerfile_cpu
-#     mv Dockerfile_gpu Dockerfile
-#     mv compose.yml compose_cpu.yml
-#     mv compose_gpu.yml compose.yml
-#     echo "✅ Using GPU enabled containers"
-# else
-#     echo "✅ Using default containers"
-# fi
+# 7. Prompt the user
+read -p "Are you using a GPU? (y/n): " response
+response=${response,,}
+if [[ "$response" == "y" || "$response" == "yes" ]]; then
+    mv Dockerfile Dockerfile_cpu
+    mv Dockerfile_gpu Dockerfile
+    mv compose.yml compose_cpu.yml
+    mv compose_gpu.yml compose.yml
+    echo "✅ Using GPU enabled containers"
+else
+    echo "✅ Using default containers"
+fi
 
-# # 8. Launch Docker Compose
-# echo "🚀 All files checked. Starting services..."
-# docker compose up -d
+# 8. Launch Docker Compose
+echo "🚀 All files checked. Starting services..."
+docker compose up -d

pyproject.toml

@@ -37,20 +37,23 @@ classifiers = [
 ]
 
 # Core dependencies required for basic operation
+# Note: qwen-tts and flash-attn require special install flags, see requirements.txt
 dependencies = [
-    "numpy>=1.22.0",
+    "numpy>=1.24.0",
     "setuptools>=80.0.0",
     "pyyaml>=6.0",
     "python-dotenv>=1.0.0",
     # Audio
     "sounddevice>=0.5.0",
     "soundfile>=0.13.0",
+    "torchaudio>=2.0.0",
     # AI/ML Core
     "torch>=2.0.0",
     "transformers>=4.40.0",
     "kokoro>=0.9.0",
     "accelerate>=1.0.0",
     "llama-cpp-python>=0.3.0",
+    "qwen-asr>=0.0.6",
 ]
 
 [project.optional-dependencies]

readme.md

@@ -71,6 +71,9 @@ A privacy-focused voice assistant that runs speech recognition, text-to-speech,
 git clone https://github.com/yourusername/fulloch.git
 cd fulloch
 pip install -r requirements.txt
+# Install special packages (see requirements.txt for details)
+pip install --no-deps git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215
+# GPU only: pip install --no-build-isolation --no-deps git+https://github.com/Dao-AILab/flash-attention.git@ef9e6a6
 ```
 
 ### 2. Configure
@@ -123,6 +126,7 @@ general:
   use_ai: true               # Enable SLM for intent detection
   use_tiny_asr: false        # Use Moonshine Tiny ASR for edge devices
   use_tiny_tts: false        # Use Kokoro TTS for edge devices
+  voice_clone: "cori"        # Voice clone name for Qwen3 TTS
 ```
 
 **ASR Options:**
@@ -133,6 +137,9 @@ general:
 - `use_tiny_tts: false` (default) — Uses Qwen3-TTS with voice cloning for natural speech
 - `use_tiny_tts: true` — Uses Kokoro TTS for faster synthesis on low-resource edge devices
 
+**Voice Cloning:**
+- `voice_clone: "name"` — Specifies which voice to clone for Qwen3 TTS. Place your reference audio (`name.wav`) and transcript (`name.txt`) in `data/voices/`. Only used when `use_tiny_tts: false`.
+
 ### Spotify
 
 1. Create an app at [Spotify Developer Dashboard](https://developer.spotify.com/dashboard)

requirements.txt

@@ -1,24 +1,26 @@
 # Core dependencies
-numpy==1.22.0
+numpy>=1.24.0
 setuptools==80.9.0
 xmltodict==1.0.2
 word2number==1.1
 beautifulsoup4==4.14.2
 
 # Audio processing
-sounddevice==0.5.3
-soundfile==0.13.1
-torchaudio==2.9.1
+sounddevice>=0.5.0
+soundfile>=0.13.0
 
 # AI and ML
-torch==2.9.1
-transformers==4.57.6
-kokoro==0.9.4
+torch>=2.4.0
+torchaudio>=2.4.0
+transformers==4.57.3
+kokoro>=0.9.4
 accelerate==1.12.0
-llama_cpp_python==0.3.16
+llama_cpp_python>=0.3.16
 qwen-asr==0.0.6
-git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215#egg=qwen-tts --no-deps
-git+https://github.com/Dao-AILab/flash-attention.git@ef9e6a6#egg=flash-attn --no-build-isolation --no-deps
+
+# These packages require special install flags - see Dockerfile
+# qwen-tts: pip install --no-deps git+https://github.com/rekuenkdr/Qwen3-TTS-streaming.git@97da215
+# flash-attn: pip install --no-build-isolation --no-deps git+https://github.com/Dao-AILab/flash-attention.git@ef9e6a6
 
 # Smart home integration
 phue==1.1.0