Skip to content

transcription.md

Latest commit

 

History

History
27 lines (24 loc) · 1.54 KB

transcription.md

File metadata and controls

27 lines (24 loc) · 1.54 KB
Raw

Transcription engines

camkit transcribe resolves an engine by precedence (highest wins): an explicit --engine openai|replicate|whisper-cpp flag, then environment, then the auto default. auto picks:

  1. OPENAI_API_KEY set → OpenAI whisper-1 (best quality). Note: this is pinned to whisper-1, not a "newer" model — the gpt-4o-transcribe models don't return the word-level timestamps the rebuild step needs.
  2. Else REPLICATE_API_TOKEN set → Replicate hosted vaibhavs10/incredibly-fast-whisper (word-level timestamps). Beats local whisper-cpp whenever the token is present — use --engine whisper-cpp to force local. Model version is pinned in code; refresh via Replicate's /v1/models/.../versions API if predictions start failing.
  3. Else whisper-cli on PATH → local whisper.cpp. By default it reuses the ggml model Camtasia downloads to Camtasia.app/Contents/Resources/models/speechToText/ (tiny/quantized — fast, lower fidelity). Override with CAMKIT_WHISPER_MODEL (path to a larger ggml-*.bin) or CAMKIT_WHISPER_BIN.
  4. None → an error telling you to set OPENAI_API_KEY or REPLICATE_API_TOKEN, or run brew install whisper-cpp. camkit never auto-installs (no silent brew).

camkit reuses Camtasia's model file but not its bundled libwhisper.dylib (private, code-signed, undocumented ABI) — you bring your own whisper-cli runner. The tiny local model has coarser word timestamps, so cross-checking with camkit silences matters even more on the local path.