Convert whisper-large-v3-czech-cv13 to ggml for VoiceInk

whisper-large-v3-czech-cv13.sh

#!/usr/bin/env bash
# Converts mikr/whisper-large-v3-czech-cv13 (HuggingFace Safetensors)
# to GGML .bin format compatible with VoiceInk.
#
# Requirements installed by this script:
#   - git clone of ggml-org/whisper.cpp (conversion script)
#   - git clone of openai/whisper (vocab assets needed by conversion)
#   - Python venv with: torch, transformers, numpy
#
# Usage: bash convert-to-voiceink.sh
set -euo pipefail

STAGING=~/Downloads/whisper
MODEL_ID="mikr/whisper-large-v3-czech-cv13"
OUTPUT_NAME="whisper-large-v3-czech-cv13.bin"
MODEL_DIR="model"
MODEL_FILES=(
    added_tokens.json
    config.json
    generation_config.json
    merges.txt
    model.safetensors
    normalizer.json
    preprocessor_config.json
    special_tokens_map.json
    tokenizer.json
    tokenizer_config.json
    vocab.json
)

clone_or_update_model_repo() {
    if [ ! -d "$MODEL_DIR/.git" ]; then
        rm -rf "$MODEL_DIR"
        echo "==> Cloning $MODEL_ID via Git..."
        GIT_LFS_SKIP_SMUDGE=1 git clone --depth=1 "https://huggingface.co/$MODEL_ID" "$MODEL_DIR"
    else
        echo "==> Model Git repo already exists, refreshing..."
        git -C "$MODEL_DIR" pull --ff-only
    fi

    echo "==> Downloading model.safetensors via Git LFS..."
    git -C "$MODEL_DIR" lfs pull --include=model.safetensors
    git -C "$MODEL_DIR" lfs checkout model.safetensors
}

clear_stale_hf_locks() {
    local cache_dir="$1/.cache/huggingface"

    if [ ! -d "$cache_dir" ]; then
        return
    fi

    if pgrep -f "(^|/)hf($| )|(^|/)huggingface-cli($| )" >/dev/null 2>&1; then
        return
    fi

    find "$cache_dir" -name '*.lock' -type f -delete
}

echo "==> Staging folder: $STAGING"
mkdir -p "$STAGING"
cd "$STAGING"

# ── Clone whisper.cpp (provides convert-h5-to-ggml.py) ───────────────────────
if [ ! -d "whisper.cpp" ]; then
    echo "==> Cloning ggml-org/whisper.cpp..."
    git clone --depth=1 https://github.com/ggml-org/whisper.cpp
else
    echo "==> whisper.cpp already cloned, skipping."
fi

# ── Clone openai/whisper source (vocab assets required by conversion script) ──
if [ ! -d "openai-whisper" ]; then
    echo "==> Cloning openai/whisper..."
    git clone --depth=1 https://github.com/openai/whisper openai-whisper
else
    echo "==> openai-whisper already cloned, skipping."
fi

# ── Python virtual environment ────────────────────────────────────────────────
if [ ! -d "venv" ]; then
    echo "==> Creating Python venv..."
    python3 -m venv venv
fi
# shellcheck disable=SC1091
source venv/bin/activate

echo "==> Installing Python dependencies (torch, transformers, numpy)..."
pip install --quiet --upgrade pip
pip install --quiet torch transformers numpy

# ── Download HuggingFace model ────────────────────────────────────────────────
if [ ! -f "$OUTPUT_NAME" ]; then
    echo "==> Ensuring $MODEL_ID is fully downloaded..."
    if command -v git-lfs >/dev/null 2>&1; then
        clone_or_update_model_repo
    else
        mkdir -p "$MODEL_DIR"
        clear_stale_hf_locks "$MODEL_DIR"
        hf download "$MODEL_ID" \
            --local-dir "./$MODEL_DIR" \
            --max-workers 4 \
            "${MODEL_FILES[@]}"
    fi
else
    echo "==> Output already exists, skipping model download."
fi

# ── Convert HF Safetensors → GGML ────────────────────────────────────────────
if [ ! -f "$OUTPUT_NAME" ]; then
    echo "==> Converting to GGML format..."
    rm -f ggml-model.bin ggml-model-f32.bin
    python3 ./whisper.cpp/models/convert-h5-to-ggml.py "./$MODEL_DIR/" ./openai-whisper .

    # ── Rename to final output name ───────────────────────────────────────────
    mv ggml-model.bin "$OUTPUT_NAME"
else
    echo "==> Output already exists, skipping conversion."
fi

echo ""
echo "✓ Done! Model saved at:"
echo "  $STAGING/$OUTPUT_NAME"
echo ""
echo "To import into VoiceInk:"
echo "  AI Models → Local tab → Import Local Model… → select $OUTPUT_NAME"
echo "  Then: Set as Default"