Speech to text with AI: transcription, minutes and translation
Meetings, customer calls, dictation and voice messages hold information that rarely gets written down in full. Speech recognition with AI turns them into text, detects the spoken language, translates if needed and summarises. Sortia Labs builds this into your workflows and also uses it in its own app Loma.
- Speech to text
- AI transcription
- Speech recognition
- Whisper
- Dictation with AI
- Meeting minutes
Where speech recognition saves time
Meetings and minutes
A meeting recording becomes a transcript, a summary and a list of agreed tasks. Your team reviews and adds to it instead of taking notes during the meeting.
Phone and customer calls
A call recording becomes a note for the CRM or ticket system, with the request, the agreed next steps and any callback. This requires the consent of the people on the call.
Dictate instead of type
Reports, e-mails, assessments or notes are spoken and delivered as text, corrected if you like and rewritten in another tone, such as friendly, professional or concise.
Translation in the same step
The system detects the spoken language and delivers original and translation together, in many languages. That helps with international teams, suppliers or customers who do not speak German.
Technical terms and names
Product names, abbreviations and the vocabulary of your industry are given to the model as hints or corrected in a second pass. That way the minutes also get right the terms a general model does not know.
How we put speech recognition to work
- 1
Sources and target
Where does the audio come from (phone system, video conference, app or dictation device) and where should the text go? That decides whether real time is needed or processing afterwards is enough.
- 2
Choosing the model
Whisper on your own server, a service with an EU data centre or a provider such as OpenAI or Google: we compare the options on your own recordings, including dialect and background noise.
- 3
Integration
Transcript and summary land where they are needed: in the CRM, the ticket system, the project tool or as a document in your filing, with approval before anything is sent.
Software we build and run ourselves.
Questions about Speech to text
Are we allowed to record and transcribe conversations?
As a rule this needs the consent of everyone involved, for phone calls as well as meetings. We build the notice and the consent into the workflow and agree with you how long audio and text are kept. The legal assessment of each case is for your data protection officer or legal adviser.
How well does the AI handle dialect and technical terms?
That depends on the microphone, the way people speak and the subject. Current models usually recognise everyday language well; technical terms, product names and personal names need help. So we give the model your vocabulary and test with your own recordings before you decide.
Can speech recognition run on our own server?
Yes. The open Whisper model can run on your own hardware, with a graphics card even for larger volumes. We run it ourselves for the Loma web app on our server in Germany. Recordings then do not need to leave your network, and instead of per-minute fees you pay for operation and hardware.
Does transcription work in real time?
Yes, with streaming models the text appears while people speak, for example for live captions or notes during a call. For minutes, processing after the conversation is often enough and tends to be more accurate and cheaper.
Read on.
What should become text in your company?
Describe your recordings and where the text should go. The prototype shows you transcript and summary from your own material.