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

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


This site uses Just the Docs, a documentation theme for Jekyll.