learning_ai_common_plat/__LOCAL_LLMs/setup-tts.sh

#!/bin/bash
# ============================================================
# TTS Setup — One-Shot Script for Fresh Laptop
#
# Sets up Orpheus TTS (via Ollama) and Qwen3-TTS (direct Python)
# on macOS (Apple Silicon) or Linux (CUDA GPU / WSL2).
#
# What this does:
#   1. Installs Python 3.12 (Homebrew on macOS, apt on Linux)
#   2. Creates Python venv with TTS packages (MPS on macOS, CUDA on Linux)
#   3. Pulls Orpheus TTS model via Ollama
#   4. Downloads SNAC audio decoder
#   5. (Optional) Downloads Qwen3-TTS 0.6B
#
# Prerequisites:
#   macOS: Homebrew + Ollama installed
#   Linux: apt + Ollama accessible at localhost:11434
#
# Usage:
#   bash setup-tts.sh
#
# After setup, test with:
#   .venv-qwen-tts/bin/python test_orpheus_tts.py
# ============================================================
set -e

SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
VENV="$SCRIPT_DIR/.venv-qwen-tts"
MODELS_DIR="$SCRIPT_DIR/models"

# HuggingFace mirror that works through corporate proxy
# On personal machines, set HF_MIRROR=https://huggingface.co to download directly
HF_MIRROR="${HF_MIRROR:-https://hf-mirror.com}"

# Detect OS
OS_TYPE="$(uname -s)"
IS_MAC=false
[ "$OS_TYPE" = "Darwin" ] && IS_MAC=true

RED='\033[0;31m'
GREEN='\033[0;32m'
YELLOW='\033[1;33m'
NC='\033[0m'

ok()   { echo -e "${GREEN}✓${NC} $1"; }
warn() { echo -e "${YELLOW}⚠${NC} $1"; }
fail() { echo -e "${RED}✗${NC} $1"; exit 1; }
step() { echo -e "\n${GREEN}=== $1 ===${NC}"; }

echo "╔══════════════════════════════════════════════╗"
echo "║       TTS Setup — Local Speech Generation    ║"
echo "║  Orpheus TTS (Ollama) + Qwen3-TTS (Python)  ║"
echo "╚══════════════════════════════════════════════╝"
echo ""

# ── 0. Check prerequisites ──────────────────────────────────
step "Checking prerequisites"

if $IS_MAC; then
    # Homebrew
    if ! command -v brew &>/dev/null; then
        fail "Homebrew not found. Install: /bin/bash -c \"\$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)\""
    fi
    ok "Homebrew"

    # Ollama (install via Homebrew if missing)
    if ! command -v ollama &>/dev/null; then
        warn "Ollama not found. Installing..."
        brew install ollama
    fi
else
    # Linux / WSL2 — Ollama should be installed on host or via install script
    if ! command -v ollama &>/dev/null; then
        # On WSL2 Ollama runs on the Windows side; check if reachable
        if ! curl -s --max-time 2 http://localhost:11434/api/tags &>/dev/null; then
            fail "Ollama not found and not reachable at localhost:11434. Install Ollama on Windows or run: curl -fsSL https://ollama.com/install.sh | sh"
        fi
        ok "Ollama reachable at localhost:11434 (Windows host)"
    fi
fi
ok "Ollama installed"

# Check if Ollama is running
if ! curl -s --max-time 2 http://localhost:11434/api/tags &>/dev/null; then
    warn "Ollama not running. Starting..."
    if command -v ollama &>/dev/null; then
        ollama serve &>/dev/null &
        sleep 3
    fi
    if ! curl -s --max-time 2 http://localhost:11434/api/tags &>/dev/null; then
        fail "Could not start Ollama. Try manually: ollama serve"
    fi
fi
ok "Ollama running on port 11434"

# GPU check
ARCH=$(uname -m)
if $IS_MAC; then
    if [ "$ARCH" != "arm64" ]; then
        warn "Not Apple Silicon ($ARCH). MPS acceleration won't be available."
    else
        ok "Apple Silicon ($ARCH) — MPS acceleration available"
    fi
else
    if command -v nvidia-smi &>/dev/null; then
        GPU_NAME=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1)
        ok "NVIDIA GPU detected: $GPU_NAME — CUDA acceleration available"
    else
        warn "nvidia-smi not found. CUDA acceleration won't be available (CPU fallback)."
    fi
fi

# ── 1. Install Python 3.12 ──────────────────────────────────
step "Python 3.12"

PYTHON_CMD=""
# Check various Python 3.12 locations
for cmd in python3.12 /opt/homebrew/bin/python3.12 /usr/local/bin/python3.12 python3; do
    if command -v "$cmd" &>/dev/null; then
        PYTHON_CMD="$cmd"
        break
    fi
done

if [ -z "$PYTHON_CMD" ]; then
    if $IS_MAC; then
        warn "Python 3.12 not found. Installing via Homebrew..."
        brew install python@3.12
        PYTHON_CMD="/opt/homebrew/bin/python3.12"
    else
        warn "Python 3.12 not found. Installing via apt..."
        sudo apt update && sudo apt install -y python3.12 python3.12-venv python3-pip
        PYTHON_CMD="python3.12"
    fi
fi

PYTHON_VER=$("$PYTHON_CMD" --version 2>&1)
ok "$PYTHON_VER at $PYTHON_CMD"

# ── 2. Create venv ──────────────────────────────────────────
step "Python virtual environment"

if [ -f "$VENV/bin/python" ]; then
    ok "Venv exists at $VENV"
else
    echo "Creating venv..."
    "$PYTHON_CMD" -m venv "$VENV"
    ok "Venv created at $VENV"
fi

# ── 3. Install Python packages ──────────────────────────────
step "Python packages"

# Check if snac is installed (quick proxy for all packages)
if "$VENV/bin/python" -c "import snac" &>/dev/null; then
    ok "Packages already installed (snac, torch, etc.)"
else
    echo "Installing packages (this may take a few minutes)..."
    "$VENV/bin/pip" install -U pip --quiet
    if $IS_MAC; then
        # macOS: default PyTorch includes MPS support
        "$VENV/bin/pip" install -U snac qwen-tts --quiet
    else
        # Linux: install PyTorch with CUDA first, then snac/qwen-tts
        echo "Installing PyTorch with CUDA support..."
        "$VENV/bin/pip" install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124 --quiet
        "$VENV/bin/pip" install -U snac qwen-tts --quiet
    fi
    ok "Packages installed"
fi

# ── 4. Pull Orpheus TTS model ───────────────────────────────
step "Orpheus TTS model (Ollama)"

if ollama list 2>/dev/null | grep -q "orpheus"; then
    ok "Orpheus TTS already downloaded"
else
    echo "Pulling sematre/orpheus:en (4 GB)..."
    NO_PROXY="ollama.com,registry.ollama.ai" ollama pull sematre/orpheus:en
    ok "Orpheus TTS downloaded"
fi

# ── 5. Download SNAC decoder ────────────────────────────────
step "SNAC 24kHz audio decoder (~76 MB)"

mkdir -p "$MODELS_DIR/snac_24khz"

if [ -f "$MODELS_DIR/snac_24khz/pytorch_model.bin" ]; then
    if $IS_MAC; then
        SIZE=$(stat -f%z "$MODELS_DIR/snac_24khz/pytorch_model.bin" 2>/dev/null)
    else
        SIZE=$(stat -c%s "$MODELS_DIR/snac_24khz/pytorch_model.bin" 2>/dev/null)
    fi
    if [ "$SIZE" -gt 1000000 ]; then
        ok "SNAC decoder already downloaded ($(echo "scale=1; $SIZE/1048576" | bc) MB)"
    else
        warn "SNAC file looks corrupted (${SIZE} bytes). Re-downloading..."
        rm -f "$MODELS_DIR/snac_24khz/pytorch_model.bin"
    fi
fi

if [ ! -f "$MODELS_DIR/snac_24khz/pytorch_model.bin" ]; then
    echo "Downloading config.json..."
    curl -k -sL -o "$MODELS_DIR/snac_24khz/config.json" \
        "$HF_MIRROR/hubertsiuzdak/snac_24khz/raw/main/config.json"

    # Verify config is JSON (not an HTML block page)
    if ! python3 -c "import json; json.load(open('$MODELS_DIR/snac_24khz/config.json'))" &>/dev/null; then
        fail "Downloaded config.json is not valid JSON. The mirror may be blocked. Try from home network."
    fi
    ok "config.json downloaded"

    echo "Downloading pytorch_model.bin (~76 MB)..."
    curl -k -L --progress-bar -o "$MODELS_DIR/snac_24khz/pytorch_model.bin" \
        "$HF_MIRROR/hubertsiuzdak/snac_24khz/resolve/main/pytorch_model.bin"

    # Verify it's a real model file (zip/pytorch format), not HTML
    FILE_TYPE=$(file -b "$MODELS_DIR/snac_24khz/pytorch_model.bin" | head -c 20)
    if echo "$FILE_TYPE" | grep -qi "html"; then
        rm -f "$MODELS_DIR/snac_24khz/pytorch_model.bin"
        fail "Downloaded model is HTML (proxy block page). Try from home network."
    fi
    ok "SNAC decoder downloaded"
fi

# Verify SNAC loads in Python
echo "Verifying SNAC decoder loads..."
if "$VENV/bin/python" -c "
import snac, torch
model = snac.SNAC.from_pretrained('$MODELS_DIR/snac_24khz')
print(f'SNAC: {sum(p.numel() for p in model.parameters())/1e6:.1f}M parameters')
" 2>/dev/null; then
    ok "SNAC decoder verified"
else
    fail "SNAC decoder failed to load. Delete models/snac_24khz/ and re-run."
fi

# ── 6. (Optional) Download Qwen3-TTS ────────────────────────
step "Qwen3-TTS 0.6B (optional, ~1.7 GB total)"

QWEN_TOKENIZER_DIR="$MODELS_DIR/Qwen3-TTS-Tokenizer-12Hz"
QWEN_MODEL_DIR="$MODELS_DIR/Qwen3-TTS-12Hz-0.6B-CustomVoice"

if [ -d "$QWEN_MODEL_DIR" ] && [ -f "$QWEN_MODEL_DIR/config.json" ]; then
    ok "Qwen3-TTS already downloaded"
else
    echo "Qwen3-TTS 0.6B requires ~1.7 GB download (tokenizer + model)."
    echo "This is optional — Orpheus TTS (above) works without it."
    read -p "Download Qwen3-TTS? [y/N] " -n 1 -r
    echo
    if [[ $REPLY =~ ^[Yy]$ ]]; then
        # Tokenizer (~650 MB)
        echo "Downloading Qwen3-TTS Tokenizer (~650 MB)..."
        mkdir -p "$QWEN_TOKENIZER_DIR"
        for f in config.json configuration.json preprocessor_config.json; do
            curl -k -sL -o "$QWEN_TOKENIZER_DIR/$f" \
                "$HF_MIRROR/Qwen/Qwen3-TTS-Tokenizer-12Hz/resolve/main/$f" 2>/dev/null || true
        done
        curl -k -L --progress-bar -o "$QWEN_TOKENIZER_DIR/model.safetensors" \
            "$HF_MIRROR/Qwen/Qwen3-TTS-Tokenizer-12Hz/resolve/main/model.safetensors"
        ok "Tokenizer downloaded"

        # Model
        echo "Downloading Qwen3-TTS 0.6B (~1.2 GB)..."
        mkdir -p "$QWEN_MODEL_DIR"
        for f in config.json generation_config.json; do
            curl -k -sL -o "$QWEN_MODEL_DIR/$f" \
                "$HF_MIRROR/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice/resolve/main/$f" 2>/dev/null || true
        done
        curl -k -L --progress-bar -o "$QWEN_MODEL_DIR/model.safetensors" \
            "$HF_MIRROR/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice/resolve/main/model.safetensors"
        ok "Qwen3-TTS 0.6B downloaded"
    else
        warn "Skipped. You can re-run this script later to download."
    fi
fi

# ── Summary ──────────────────────────────────────────────────
step "Setup Complete"

echo ""
echo "Installed components:"
echo "  Orpheus TTS (Ollama):  $(ollama list 2>/dev/null | grep orpheus | awk '{print $NF}' || echo 'ready')"
echo "  SNAC decoder:          $MODELS_DIR/snac_24khz/"
if [ -d "$QWEN_MODEL_DIR" ] && [ -f "$QWEN_MODEL_DIR/config.json" ]; then
    echo "  Qwen3-TTS 0.6B:       $QWEN_MODEL_DIR/"
else
    echo "  Qwen3-TTS 0.6B:       (not installed — re-run setup to add)"
fi
echo ""
echo "Disk usage:"
du -sh "$MODELS_DIR"/* 2>/dev/null | sed 's/^/  /'
echo ""
echo "Test commands:"
echo "  $VENV/bin/python $SCRIPT_DIR/test_orpheus_tts.py"
if $IS_MAC; then
    echo "  afplay test_orpheus_tara.wav"
else
    echo "  aplay test_orpheus_tara.wav  (or: ffplay -nodisp -autoexit test_orpheus_tara.wav)"
fi
if [ -d "$QWEN_MODEL_DIR" ]; then
    echo "  $VENV/bin/python $SCRIPT_DIR/test_qwen_tts.py"
fi
echo ""
echo "Voices: tara, leah, jess, leo, dan, mia, zac, zoe"
echo "Emotion: <laugh>, <chuckle>, <sigh>, <cough>, <groan>, <yawn>, <gasp>"