Speech input

With speech input you dictate into any text field in Moliri instead of typing, such as the condition report in the store. Each person sets it up for themselves in the profile under Speech input and chooses the service that recognises the speech. The sound goes straight to that service; with In the browser it stays in the browser. It never passes through the Moliri server.

Setting up

When it is switched off, the page only shows the switch Speech input on. When it is on, the service, its fields and Test connection appear below, then the triggers under Dictation. Save applies everything; only then do text fields offer dictation. Until a service is filled in completely, the text next to Save reads: The microphone is offered once a service is set up.

In a new setup, Transcribe by Moliri is preselected. If you have already saved a service, you keep it. Test connection queries the chosen service and reports The service responds. or the reason why not. With In the browser, Download or Test model takes its place.

Under Dictation, In the browser, ElevenLabs and your own server have the Spoken language (German by default, also Detect automatically). For Transcribe it is missing: Transcribe recognises the language itself.

Transcribe by Moliri

Transcribe is a Windows app from Moliri that turns speech into text on your own PC. Moliri connects to its live transcription. The default address is ws://localhost:8765/api/live; Transcribe shows the valid address on its Live page under Programme verbinden (connect programs).

While you speak, the recognised text stands in grey in the field. Each finished sentence is inserted straight away.

Allowing Moliri in Transcribe

Websites can only connect to Transcribe once you allow them there. On your first dictation, Transcribe therefore turns Moliri away. Moliri reports Transcribe refuses the connection., below it the message from Transcribe with the address of your Moliri instance, and offers Try again.

1. Open Transcribe, the Live page, Programme verbinden. Under Wollten sich verbinden (wanted to connect) you see the address of your Moliri instance.

2. Click Erlauben (allow) next to the address. It then appears under Wer sich verbinden darf (who may connect), where Nicht mehr erlauben (no longer allow) blocks it again.

3. In Moliri, click Try again. The dictation starts.

If Transcribe is not running or the address is wrong, Moliri reports Transcribe does not answer. with the hint Start Transcribe and check the address (“Live” page). The first time, Chrome may ask whether the page may access devices on the local network; allow it, because Transcribe runs on the same PC.

In the browser

Here Whisper recognises the speech in the browser, on your own computer. For this the browser downloads the model once from Hugging Face. The files then stay in this browser, speech input works without the internet as well, and your speech does not leave the computer. Each browser and each computer needs the download for itself.

Next to each model is the size of the download, roughly 80 to 600 MB. It depends on the model and on whether the browser can use the graphics card: on the processor it loads the smaller files.

Download loads the chosen model and shows Loading … with a percentage while it does. Afterwards the page says Ready, running on the graphics card. or Ready, running on the processor (slower than on a graphics card). If the model is already stored in the browser, the button is called Test model, and next to it Remove from browser deletes the files again. If you dictate without downloading first, Moliri loads the model on the first dictation and shows Loading speech model … at the field. If it is already stored, Moliri starts it in the background as soon as a text field gets the cursor.

The text comes after you stop speaking, after a few seconds on an ordinary laptop, faster with a graphics card than on the processor. Grey interim text while you speak only appears on fast devices. Even More accurate does not always get rare words and names right, so it is worth glancing over the inserted text. If the browser cannot run a model, the page says This browser cannot run a model.

Dictating

You can dictate into any text field of the interface: single-line fields, text areas and fields with formatting, when entering data as well as in dialogs and settings. Password fields are excluded, as are fields that are locked at the moment. There are two triggers, and you can switch off each of them:

Esc cancels a running dictation without inserting anything. While recording, a hint at the field shows what is happening (Connecting …, Listening …, Transcribing …).

The text lands at the cursor position and behaves like typed text: it is saved when you leave the field, and Ctrl + Z undoes it. After saving, the profile page shows the field Try it, where you can test the setup.

Messages

Further reading

• Profile : the other pages of the profile
• Edit an object : entering and saving values