Skip to content

v0.2.0

Choose a tag to compare

View all tags
@achetronic achetronic released this 25 Mar 02:56
· 14 commits to master since this release
v0.2.0
73e0251

Performance

  • 3x faster inference The decoder model was being loaded from disk ~150 times per audio request. Now it loads once at startup and stays in memory.
  • Worker pool New -workers flag (default: 4) controls how many requests can run in parallel. Each worker keeps its own decoder session ready to go.
  • Minor optimizations Pre-computed Hann window, pre-compiled regex, encoder path resolved once at startup instead of per request.

New features

  • API key — Set PARAKEET_API_KEY env var and all /v1/* endpoints will require Authorization: Bearer <key>. Don't set it, nothing changes. /health is always open.
  • Structured logging — Switched to Go's slog. Two new flags:
    • -log-format json for production log pipelines
    • -log-level debug|info|warn|error to control verbosity
  • Graceful shutdown — Ctrl+C or SIGTERM now waits for running requests to finish (up to 30s) before cleaning up. No more crashes on shutdown.

Breaking changes

  • -debug flag is gone. Use -log-level debug instead.

New flags & env vars

Flag/Variable What it does Default
-log-level Log verbosity info
-log-format text or json output text
-workers Parallel inference slots 4
PARAKEET_API_KEY API key (empty = no auth) empty
Assets 5
Loading