Voice: dictation and call transcription

The module AI Voice adds speech to text to eYssen. You can dictate messages instead of typing them, and you can have a Discuss call transcribed and summarised, with action items. It uses a speech-to-text model that an administrator configures once.

Set up speech to text

1. Add a provider that offers the OpenAI-compatible transcription route, for example OpenAI, or Groq or a self-hosted Whisper server using the provider type OpenAI with its own base URL (see Providers and models). Anthropic and OpenRouter have no transcription route and cannot be used for speech to text.

2. Refresh the models of the provider. Ids that contain whisper or transcribe are classified as Speech to text models automatically.

3. Open AI ‣ Configuration ‣ Settings, find the block Voice and choose the Speech-to-text model. Dictation and call transcription stay hidden until a model is chosen.

The other settings of the block are:

• Spoken language: an optional two-letter language code such as hu or es. Leave it empty to detect the language automatically.

• Max recording size (MB): the limit per uploaded recording or call piece (10 MB by default, never more than 25 MB).

• Max dictation length (seconds): a dictation stops by itself after this time (120 seconds by default).

• Call chunk length (seconds): call audio is uploaded and transcribed in pieces of this length (30 seconds by default).

• Max transcribed minutes per call and person (180 by default).

• Keep call audio (days): how long the call recordings are kept (7 days by default). With 0 the audio is not stored at all, only the text. The text always stays.

• Summarize calls automatically: when a transcribed call ends, a summary with action items is posted in the channel (on by default).

Note

Speech to text uses the AI capability Ask, the same rate limits as other AI requests and the daily call and cost caps. Each request is written to the AI log, with its cost estimated from the audio length or the token usage.

Dictate instead of typing

A microphone button (Dictate) appears in the message box of the chatter and of Discuss, also when you edit a message, and next to the input of the AI panel. It is shown only to AI users, once a speech-to-text model is configured and when your browser can record.

1. Click the microphone and allow the browser to use it when asked.

2. Speak. Click the button again (Stop dictation) when you are done, or wait for the maximum length.

3. The recognised text is inserted at the cursor. It is never sent automatically: read it, correct it and send it with the normal button.

The dictation audio is sent to the provider for transcription and is never stored in eYssen.

Transcribe a call

In a Discuss call, every participant sees the action Transcribe in the call controls (or Join transcription once someone has started it, and Stop transcription while you take part). While a call is being transcribed, the action is highlighted for everyone.

How consent works

Transcription is opt-in for each person:

• Turning it on asks you to confirm that your microphone will be recorded and transcribed.

• When the first participant starts the transcription, a notice is posted in the channel. It says who started it, that only people who turn it on themselves are recorded, each from their own microphone, and how long the audio is kept. Everyone who joins later is announced the same way.

• Each browser records only its own microphone, and only while that person has it turned on and is not muted. The server refuses audio from people who did not opt in and from people outside the channel.

• Turning it off, leaving the call or the end of the call stops your recording. Guests who are not users cannot take part.

Important

Check the privacy and employment rules that apply to you before you use call transcription, and tell the people on your calls. The audio is sent to the speech-to-text provider you configured.

Transcript, summary and action items

When the last person turns transcription off, or nobody is left in the call, the transcript ends. If Summarize calls automatically is on, a summary is posted in the channel about a minute later, as a comment from the person who started the transcription. It contains the summary, the action items with a suggested person and a link to the transcript.

Open all transcripts in AI ‣ Call transcripts. Members of the channel see their own calls; AI administrators see all of them. On a transcript you have the tabs Summary and Transcript, the list of Action items and the buttons Summarize or Summarize again if you want a summary on demand (not while the call is still being recorded). Filter the list with I took part and Not summarized.

The AI only suggests an assignee when the name spoken in the call matches exactly one internal member of the channel. Nothing is scheduled automatically: review the items on the transcript, edit the task, person or due date, tick the ones you want and click Create activities. Your click is the confirmation. Every item with an assignee gets a To-Do activity; the assignee must be a channel member.

Limits

• Words spoken across the boundary of two pieces of audio can be split, and a single browser records one speaker only.

• Very long calls are summarised from the last 60,000 characters of the transcript.

• The push-to-talk and voice-activation settings of the call are not mirrored; only the mute state is respected.