Callesnikov Speech

Use this skill to process long local audio incrementally through an offline-friendly ffmpeg and whisper.cpp pipeline.

Only perform the stage the user asked for. Do not advance to later stages unless the user explicitly asks to continue.

Core Rules

• Keep every runtime asset self-contained in this skill directory: C:\Users\User333\.codex\skills\callesnikov-speech.
• Put whisper.cpp, build outputs, model files, helper downloads, temporary files, and generated skill-owned metadata under this skill folder only. Do not install models or binaries globally.
• Prefer the model at runtime/whisper.cpp/models/ggml-large-v3-turbo.bin when it exists.
• Use official sources when checking freshness or downloading runtime assets: https://github.com/ggml-org/whisper.cpp for runtime, https://huggingface.co/ggerganov/whisper.cpp/tree/main for pre-converted GGML model files, and official OpenAI Whisper references for model lineage.
• If a needed codec, audio format dependency, or system package is outside the bundled skill runtime, ask the user where to install or download it.
• Work in the current task directory for audio chunks and transcript outputs.
• Use the scripts in scripts/ for repeatable stages when possible.
• Treat runtime/whisper.cpp/models/ as the only model directory unless the user explicitly provides another model path for a one-off run.
• Never accept a generic main executable from PATH on Windows; use a local whisper-cli.exe, local main.exe, or an explicit user-provided binary path.

Runtime And Model Sources

Expected local layout:

C:\Users\User333\.codex\skills\callesnikov-speech\
  runtime\
    whisper.cpp\
      build\
      models\
        ggml-large-v3-turbo.bin

Use these upstream sources when bootstrapping or updating:

• Runtime repository: https://github.com/ggml-org/whisper.cpp
• GGML model repository: https://huggingface.co/ggerganov/whisper.cpp/tree/main
• Default model file: ggml-large-v3-turbo.bin
• whisper.cpp model downloader after cloning/building: runtime\whisper.cpp\models\download-ggml-model.cmd large-v3-turbo

When assets are missing, explain the missing pieces and offer a local skill-folder bootstrap. Do not download the large model or clone/build whisper.cpp unless the user explicitly asks.

Bootstrap command, only after explicit user approval:

C:\Users\User333\.codex\skills\callesnikov-speech\scripts\bootstrap_whisper_cpp.ps1 -Model large-v3-turbo

This script clones/updates whisper.cpp, builds it with CMake, downloads the selected GGML model, and then runs preflight. It writes only under C:\Users\User333\.codex\skills\callesnikov-speech\runtime.

Stage 0: Bootstrap And Preflight

On the first user message in the current chat, run one preflight for local tooling and the default model:

python "C:\Users\User333\.codex\skills\callesnikov-speech\scripts\preflight.py"

Check ffmpeg, python, cmake, a runnable local whisper.cpp binary, and at least one usable ggml model under runtime/whisper.cpp/models/.

If files are missing, explain exactly what is absent. Downloads and builds must be placed inside this skill folder. Do not start transcription until the user explicitly asks.

Stage 1: Split Source Audio

When the user provides a source audio file, split it into short chunks:

python "C:\Users\User333\.codex\skills\callesnikov-speech\scripts\split_audio.py" "C:\path\to\audio.mp3"

Default to 60 second chunks unless the user asks for another value in the 50-70 range. Write chunks into splitted_records as audioChunk_XXX.wav.

After success, tell the user the source audio was split and point to splitted_records.

Stage 2: Prepare Output Directories

After splitting, create transcript output directories next to splitted_records:

python "C:\Users\User333\.codex\skills\callesnikov-speech\scripts\prepare_dirs.py"

This creates transferred and low_merged. Continue to stage 3 only when the user asks.

Stage 3: Transcribe Audio Chunks

Transcribe prepared chunks from splitted_records, not the original long recording:

python "C:\Users\User333\.codex\skills\callesnikov-speech\scripts\transcribe_chunks.py"

Use the model requested by the user, or default to runtime/whisper.cpp/models/ggml-large-v3-turbo.bin. If that model is missing, stop and request permission to bootstrap/download it into the skill folder. Save json, srt, and txt artifacts for every processed chunk into transferred.

After success, tell the user each chunk was converted and the results are in transferred.

Stage 4: Merge TXT Chunks

Merge only the txt outputs from transferred:

python "C:\Users\User333\.codex\skills\callesnikov-speech\scripts\merge_txt_chunks.py"

Default to 20 text chunks per merged file unless the user asks for another value in the 20-30 range. Write merged files to low_merged using mergedFromXToY.txt, where X and Y are the included chunk numbers.