Skip to content

AnswerCue v2.8.6

Choose a tag to compare

View all tags
@devasheeshG devasheeshG released this 08 Oct 16:16
· 15 commits to main since this release
v2.8.6
1179013

Release date: 2026-10-08

Summary

AnswerCue v2.8.6 replaces local speech transcription with selectable OpenAI and ElevenLabs real-time transcription. LLM settings, transcription keys, and the active transcription model are managed in separate sections of AI Providers.

What's New

  • Transcription providers: Save an OpenAI or ElevenLabs API key, test session access, and choose an active transcription model in one searchable custom dropdown.
  • OpenAI GPT Live Transcribe: Streams incremental text with the minimal delay setting and commits captured speech turns for final transcripts.
  • ElevenLabs Scribe v2 Realtime: Streams partial and committed transcripts through its real-time WebSocket API.
  • Independent audio channels: Microphone and meeting audio use separate sessions with the selected provider/model so user and interviewer roles remain distinct.

Improvements

  • Renamed the provider-card section to LLM providers while retaining the AI Providers settings tab. The LLM model and OpenAI response tier remain independent of transcription.
  • OpenAI transcription uses its separate saved key, or reuses the saved OpenAI LLM key when no separate key is supplied.
  • Existing local-only configurations choose an available cloud provider from saved keys. Explicitly selected transcription models persist across restarts and are not silently switched when their key is removed.
  • First-run setup now configures cloud transcription instead of downloading a local model. Audio settings link to the transcription section while retaining device controls.
  • Connection tests wait for an accepted provider session without recording audio. Transcription changes are disabled during an active interview; stop first, then configure and restart.

Fixes

  • Removed forced Moonshine routing and the local model download/preload path.
  • Startup audio buffers until the provider accepts the session; buffers are bounded and discarded audio is reported.
  • Captured PCM is mixed to mono and converted to the provider sample rate. Small capture chunks are grouped into approximately 100 ms packets.
  • OpenAI final results are reconciled by turn identity and emitted in committed-turn order, even when completion events arrive out of order.
  • Stop flushes remaining audio and waits up to two seconds for trailing transcript results before saving the interview. Connection failures, rejected credentials, quota errors, and drain timeouts are surfaced.
  • Stop cancels reconnect/setup/keepalive timers; stale sockets cannot insert text into a restarted interview. Connection tests clean up immediately after session acceptance.

Technical

  • Bumped package version to 2.8.6. Includes the full-file instruction context, Persona, searchable LLM chooser, restartable interviews, DMG signing repairs, and Dock fixes from previous releases.
  • Removed local speech model code, the speech worker, model downloader, local-model UI/IPC, Hugging Face Transformers, ONNX Runtime, and unused ElevenLabs SDK/direct gRPC dependencies. No local speech model is bundled or downloaded. Existing user model caches are not automatically deleted.
  • Legacy Google/Deepgram SDKs are development dependencies for historical adapter checks and are excluded from production dependencies.
  • Full automated suite: 1,053 passed, zero failed, 41 skipped. Renderer type checking and production builds pass. Electron type checking retains the unchanged fork's 37 diagnostics, with no new diagnostics.
  • Protocol tests cover accepted sessions, startup buffering, PCM conversion, partial/final results, out-of-order completions, explicit commits, trailing results, bounded buffers, stale sockets, fatal provider errors, and timer cleanup. Credential tests cover upgrade behavior, key reuse, selected-model persistence, and missing keys.
  • Browser checks exercise provider/model search, key saving, connection success/error UI, cloud-only onboarding, and start/stop/restart with the same interview identity. Browser checks use simulated Electron IPC; protocol tests use simulated provider sockets.
  • Live API audio remains unverified: this environment could not securely pass user API keys into its test runner. Use Test Connection and an installed-Mac audio check to confirm your account's model access, network, transcription accuracy, and latency. No fixed end-to-end latency is promised.
  • Audio now streams to the selected cloud transcription provider during interviews. Internet, provider API access, and usage charges are required. Two audio streams may incur separate usage. Updated privacy policy and in-app help reflect this behavior.
  • macOS CI verifies app bundles, mounted DMGs, and extracted updater ZIPs before uploading release assets. Native permissions, audio capture, and Dock behavior still require installed-Mac confirmation.

Platform Downloads

macOS

  • Apple Silicon: AnswerCue-2.8.6-arm64.dmg or AnswerCue-2.8.6-arm64-mac.zip.
  • Intel: AnswerCue-2.8.6-x64.dmg or AnswerCue-2.8.6-mac.zip.
  • Auto-update metadata: latest-mac.yml and ZIP blockmaps.

Ad-hoc signed and unnotarized. Replace the previous app in Applications; settings and interview history use the existing data directory. If Gatekeeper blocks this trusted fork build, remove its quarantine flag:

xattr -dr com.apple.quarantine /Applications/AnswerCue.app

Windows

No Windows installer is included in this release.

Linux

No Linux package is included in this release.

Assets 9
Loading