v1.8.4 Released: System Audio, Watch Together, Voice Improvements + More

Hi everyone,

v1.8.4 has been released:

System Audio and Watch Together Improvements

Fixed false system-audio detections caused by internal noise or processes on silent computers, reducing comments about music or other sounds when no actual audio is playing.

Improved silence detection, noise filtering, blank-audio handling, stalled-capture recovery, and sound-resume behavior for more dependable listening across devices.

Improved handling of music and dialogue together for more relevant mixed-media reactions.

Standalone System Audio Listen now includes adjustable response frequency while continuing to listen and transcribe, allowing more natural pauses without losing audio awareness. You can change the frequency within the Settings and System Audio Listen section.

Voice and Microphone Improvements

Improved sentence completion timing, pause handling, and recovery when sending is interrupted for more natural spoken conversations.

Improved speaker-echo reduction to help prevent Ciel’s voice from being transcribed as user speech.

Microphone transcription and automatic sending now pause safely while Ciel is thinking or speaking, reducing accidental interruptions and echoed responses.

Conversation Context Improvements

Added optional Conversation Context expansion controls in Models & Performance, allowing supported models to use larger verified context limits while preserving the existing defaults.

Automatic capacity checks and conservative fallback keep expanded context within each model and device’s available limits, while context errors safely reduce the expansion level.

(Longer context helps Ciel retain more recent dialogue and relevant history during complex conversations, improving continuity in follow-up replies.)

Context and Memory Improvements

System Audio and Watch Together now retain recent media context more naturally, keeping follow-up replies better connected to what was just prior heard or seen.

Improved memory protection so media, audio, vision, and ambient observations are not treated as personal user memories.

Refined multilingual identity, profile, recall, and conversation handling for more consistent responses across supported languages.

Cloud and Local AI Improvements

Improved OpenAI, Anthropic, and OpenRouter context handling for responses, tools, goals, retries, and vision.

Improved handling of failed, truncated, interrupted, and incomplete provider responses to reduce confusing or unfinished results.

Interface and Settings Improvements

Added localized System Audio response-frequency titles and descriptions for clearer controls across supported languages.

Updated High Quality STT descriptions for voice, Watch Together, and System Audio use.

Improved floating chatbox opacity for clearer and more consistent readability.

Additional Improvements and Bug Fixes

Various audio, microphone, Watch Together, UI, backend, localization, startup, and response-flow bugs were fixed, along with the removal of several deprecated settings.

If you encounter any issues please post them in the discussion boards so they can be investigated.

Update will be live on all platforms and compact branch soon.

Thanks, everyone!