This page is machine-translated from French. Read the French original.

Images, videos and voices

Beyond text, WivenLLM can produce images and short videos, read responses aloud, and transcribe speech. Image and video generation is only available if an administrator has configured it; text-to-speech and dictation functions automatically.


Image generation

Use

In a conversation:

/img un plan de coupe schématique d'une pompe à chaleur, style technique

(/image also works.) The image appears in the conversation and can be downloaded. With an image attached to the message, the command requests that the image be modified. An agent can also generate an image during a task, if the corresponding skill is enabled.

Configuration — administrator

Settings → Image Generation. Two types of suppliers:

Local — The models run on your machine; no data is sent out. This requires suitable hardware, in practice a graphics card. Supported local installations include both an integrated engine managed by WivenLLM and image generation servers that you already use.

External — The main commercial image generation services are supported. Your descriptions are then sent to the chosen provider.

Good to know

  • The generated images are attached to the conversation.
  • Describe the style as much as the subject: «technical diagram in black and white, white background» gives a very different result from «illustration».
  • An image generation model cannot reliably write readable text; do not expect visuals with accurate labels.

Video generation

Same principle, with the /video command.

/video une vue aérienne lente d'un chantier de construction

Realistic expectations: the command produces a 5-second sequence in 16:9 format. Generation is long — from several minutes to much longer depending on the hardware — and much more demanding than the image.

The configuration is done in Settings → Video Generation. Only locally running engines are available: the built-in engine, or your own ComfyUI installation. No external video generation services are offered.


Listen to the responses (text-to-speech)

A play button appears under each answer.

Available options:

  • Voice of the system (default) — uses the browser's built-in text-to-speech function. No configuration required, no data sent anywhere, quality varies depending on the operating system.
  • High-quality local voice — a downloaded engine that runs in the browser, more natural than system voice, without transmitting text; or a synthesis engine that you host.
  • External services — OpenAI, ElevenLabs or any OpenAI compatible service, for superior quality; the text of the responses is then transmitted to the provider.

Configuration in Settings → Voice and speech.


Dictating messages (voice recognition)

A microphone button in the input area transcribes your speech into text. By default, transcription is performed on the server using a local Whisper template, without an API key and without the audio leaving the server. The template is downloaded once, upon first use.

In Settings → Voice and speech, an administrator can choose another engine: a hosted service (OpenAI, Deepgram, Groq, any OpenAI compatible service), a local recognition server, or the system voice.

Important confidentiality point to be aware of: Browser-integrated speech recognition typically transmits audio to the browser publisher's servers. On a system requiring a purely local setup, do not choose the system voice option.

Outside of the desktop application, the browser only allows the microphone over HTTPS or on localhost.


Voice conversation

The button Start the voice conversation, Next to the microphone, a hands-free mode is activated: you speak, the assistant responds aloud, and listening resumes. It uses the engines configured for dictation and speech synthesis, and the same privacy rules apply.


Transcription of audio and video files

Unlike dictation: when an audio or video file is imported as a document, it is transcribed and then indexed like any other text. The duration is limited to 4 hours.

The built-in transcription engine operates locally, without a key. Its template (approximately 670 MB) is downloaded once, upon first use; subsequently, nothing is output. An external service can be substituted in Settings → Transcription Template.

Common uses: recordings of meetings, interviews, training sessions, voice messages. The transcript becomes searchable like an ordinary document — «what was decided about the budget at the meeting on the 12th?».

To record a meeting directly from the application and obtain a structured summary, see Meetings.

Quality : It depends on the clarity of the recording, the number of speakers, and any overlap in speech. A meeting recording captured by a laptop microphone will produce inconsistent results.


What comes out of the network

Function In local configuration In external configuration
Image generation Nothing The description is sent to the supplier
Video generation Nothing Not applicable: no external service
Text-to-speech Nothing (system voice: depends on the browser) The text of the response is sent to the supplier
Voice recognition Nothing with Whisper local, the default setting (system voice: depends on the browser, to be checked) The audio goes to the supplier
File transcription Nothing The file is sent to the supplier