Video and speech

Creative tools include image generation/editing, video generation and speech. Enable VideoGenerationTool, VideoStatusTool and SpeakTool in the tool picker.

Video

Configure Settings → Model Roles → Video generation with a LiteLLM video model. Refresh model capabilities to discover models whose upstream mode is video_generation, or enter a custom model ID. This role does not inherit the chat model. Model-specific duration, size and reference-image limits still apply.

generate_video(prompt, seconds?, size?, image_path?) submits a paid provider job and returns a job_id. Reference images are local files of up to 20 MB. check_video(job_id) queries progress and downloads a completed result as MP4. Wait at least 10 seconds between checks. Job records persist in the project videos/ directory and are scoped to the originating chat. A new process can resume checks; there is no background polling while Suzent is closed. Provider retention limits still apply. Retry check_video, not generation, after a transient status/download error. After an ambiguous submission timeout, inspect provider jobs before resubmitting to avoid duplicate charges.

Example: “Generate an 8-second landscape video of mist moving through a forest. Check its progress, then show the finished video.” For image-to-video, attach an image and ask to use that image as the reference.

Completed videos appear outside the activity rail at the point they were retrieved, with playback controls. Later tools start another rail segment.

Speech

Settings → Voice & audio saves the API speech model and global defaults together. Model Roles shows a TTS summary linking to this page. Gemini and OpenAI models provide built-in voice suggestions and model-specific controls; unknown/custom providers retain a manual voice ID field. Changing models resets the voice and incompatible options before Save. Each speak call may override engine, voice, speed, volume, language, pitch, output format, and prompt (tone/style instructions).

• System: no API or TTS model required. Uses a local voice installed on the device viewing the message. Supports voice, language, speed, pitch and volume. Install a system voice if none are available. This mode does not export audio or interpret style instructions. Voice availability depends on the device/webview.
• API: uses the TTS model role and provider credentials; saves audio in the project's audio/ directory for replay. Supports provider voice IDs, speed, format and style instructions. Pitch/language overrides apply to system speech only. Volume controls playback. Auto format uses WAV for Gemini and MP3 for other providers. Gemini style instructions are included in its text prompt; numeric speed and formats other than WAV are rejected before generation because this adapter supports WAV and prompt-based pacing for Gemini. Gemini audio requests carry the configured credentials directly, bypassing LiteLLM’s speech bridge. Other models may impose their own limits.

New speech results in the current chat play automatically by default. System and API speech share a queue and provide Stop and Replay controls. Disable Automatically play new tool speech in Voice settings for manual playback. History loading, refresh, chat switches and repeated stream snapshots do not replay earlier results. Switching chats stops playback and clears the queue. If device/browser autoplay is blocked, click Play to retry.

An existing TTS role defaults to API speech until voice settings are saved; otherwise the default is system speech. System and API voices use different IDs.

Examples:

• “Read this in Chinese using system speech, speed 0.9, pitch 1.1.”
• “Generate API speech with voice alloy, as a WAV file, in a calm tone.”

The speak tool now returns playable results in chat instead of opening the backend server's speaker device. Hardware voice-node playback is unchanged.

Specialist role recommendations require a configured provider (credentials for API services, explicitly enabled models for keyless local services). Unconfigured providers are excluded without deleting existing role assignments. This checks configuration, not account/model access, and does not make verification requests.