Skip to content

Releases

v1.1.1

Choose a tag to compare

@github-actions github-actions released this 26 Oct 16:56

Fixes

  • Spanish transcription on the FluidAudio engine (macOS) - long stretches
    of speech are cut every ~3 s for low latency, and Parakeet v3 drifted toward
    English at each cut, so Spanish came out as look-alike nonsense ("equipo de
    ciencia" for "equipo de cinco"). Each cut is now decoded with the preceding
    audio as context and stitched by word timing, so transcripts match an uncut
    decode while finals still arrive every ~3 s for interview mode.

Docs

  • README section for Interview & Consultant mode: setup, roles, and how the
    quick and depth answers work.

v1.1.0

Choose a tag to compare

@github-actions github-actions released this 20 Oct 17:27

This release turns ContextCall into a live interview copilot powered by
Claude, and adds a native, Neural Engine speech engine on Apple Silicon Macs.

Claude brain

  • Claude CLI provider - new "Claude CLI" AI mode that runs your own local
    claude login (ContextCall never reads or stores its credentials). Live
    suggestions share one warm conversation per recording session, so each one
    builds on the previous ones instead of starting from zero. Notes and
    auto-summaries stream through Claude too.
  • Anthropic API provider - "Anthropic API" AI mode using a Console API key
    (stored in the OS keychain) with a configurable base URL and auth header for
    compatible gateways; streaming, prompt caching and multi-turn history.
  • Shared conversation interface - ConversationSession abstracts
    persistent multi-turn backends, so interview mode can run over the Claude
    CLI or the Anthropic API (Settings → Interview → Connection).

Interview mode

  • Live interview copilot - detects interviewer questions (question marks,
    "walk me through…", "tell me about…", Spanish openers too) the moment they
    are transcribed and streams two answers: a quick first-person answer to say
    right away (first words in ~1 s) and a deeper technical answer from Opus.
  • Grounded in your resume - PDF, Markdown or text resume plus job
    description and notes are loaded once at session start; PDFs go to Claude
    as documents.
  • Keeps context - both tiers remember the whole interview, so follow-ups
    make sense; a new question interrupts an answer still streaming; small talk
    is ignored.
  • Ask anytime - type a question (answered directly, even without
    recording), or press Answer for what the interviewer just said.
  • Consultant role, configurable quick/depth models (Haiku default for the
    quick tier), optional thinking and web search for the depth tier, English
    pinning when the language is set to English.
  • Interview workspace - answers beside the live transcript.

Speech-to-text

  • whisper-rs is the default engine - in-process, no Python or server, small
    model download.
  • FluidAudio voice engine (macOS, Apple Silicon) - Parakeet TDT running on
    the Apple Neural Engine through a bundled Swift helper, so transcription is
    fast and on-device with no Python or server and little CPU/GPU load.
    • Multilingual v3 model by default, plus an English-only v2 model.
    • Speaker labels from on-device diarization.
    • Honors the transcription language, pinning English output for en-*
      locales so v3 doesn't drift between languages.
    • Models download on first use; select it in Settings → Engines.

Fixes

  • Notes and auto-summaries no longer fail with 401 Unauthorized when a CLI
    provider is selected.
  • tauri dev no longer rebuilds in a loop while the FluidAudio helper compiles
    (.taurignore).