Whisper.cpp
TAS Score: S3/3 — D5/5 — A3 / T2
Brief Description
C/C++ port of OpenAI’s Whisper speech-to-text model. Transcribes audio locally on CPU — no cloud, no API keys, no data leaves your machine. Supports 99 languages. From the same author as llama.cpp (Georgi Gerganov).
Architectural Role
Compute/inference layer: local speech-to-text. Replaces cloud transcription services (Google Speech-to-Text, Otter.ai, OpenAI Whisper API). Pairs with Piper (text-to-speech) for a complete local voice pipeline. Can be integrated into Home Assistant, Jellyfin, and other self-hosted tools.
Technical Autonomy
- Works completely offline (after model download)
- Stores data locally — audio never leaves the machine
- Does not require external accounts
- Allows data export — standard text/SRT/VTT output
- Compiles from source with no external dependencies
Philosophical Assessment (whose.world criteria)
| Criterion | Status | Comments |
|---|---|---|
| Pause | ✅ | Stop the process. Models stay on disk. |
| Exit | ✅ | Models are standard GGML files. Output is plain text. |
| Recoverability | ✅ | Recompile, re-download model. Nothing to lose. |
| Visibility | ✅ | MIT license. Full source on GitHub. |
| External Dependencies | ✅ | None. Compiles standalone. Models downloadable from HuggingFace. |
Configuration (Minimal)
# Build from source
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && make
# Download a model
bash models/download-ggml-model.sh base.en
# Transcribe an audio file
./main -m models/ggml-base.en.bin -f audio.wav
# Real-time microphone transcription
./stream -m models/ggml-base.en.bin
Related Recipes
- Home Office — local transcription for meetings
Alternatives
| Alternative | Autonomy | Notes |
|---|---|---|
| OpenAI Whisper API | A0 / T0 | Cloud-only. Audio sent to OpenAI servers. Pay per minute. |
| Google Speech-to-Text | A0 / T0 | Cloud-only. Audio sent to Google. |
| Otter.ai | A0 / T0 | Cloud-only. Requires account. Stores all transcriptions on their servers. |
Trajectory
Direction: opening
MIT licensed, same author as llama.cpp, active community. Expanding hardware support (CoreML, CUDA, Vulkan). GGML model format widely adopted. No commercial entity behind it — pure open-source.
Signal assessment:
| Signal | Status | Evidence |
|---|---|---|
| License | ✅ | MIT. No changes. |
| Feature gating | ✅ | No paid tier. Everything free and open. |
| Self-hosting | ✅ | Single compilation. Runs on anything from Raspberry Pi to server GPUs. |
| Governance | ✅ | Community-driven. Healthy contributor base. |
Signal key: ✅ opening · ➖ neutral · ⚠️ closing
Sources
- Website: github.com/ggerganov/whisper.cpp
- Repository: github.com/ggerganov/whisper.cpp
- Original model: github.com/openai/whisper