The words a recogniser gets wrong in a service or a council meeting are the ones already on paper. Hand it the agenda, take the audio from the mixing desk rather than the room, and a weekly event costs less than the coffee afterwards.
A Sunday service, a Tuesday lecture and a monthly council meeting have more in common than they look. Each happens again and again in the same room, each is run by a volunteer or a clerk rather than an AV company, and each serves people who would follow better in another language, people who cannot hear well and people who could not be there at all. The budget is small and the people running it change. The person who set it up last year has usually moved to the flower rota.
This article covers setting VocalStack up once for a room like that, so that every week after is a matter of pressing start. It also covers the handful of things that decide whether the words on the screen are right. For the one-off version, a conference with several stages, see Running a multilingual conference with live captions. Start with the thing you will print.
A broadcast is one public link and a QR code. Anyone can open the link in a browser, pick a language and read as the speaker talks. There is nothing to install and no account to create. One broadcast can be fed by as many live transcriptions as you like, so this Sunday's recording and the next both appear at the same address.
Print the QR code once and put it on the notice board, the order of service, the lecture slides or the public gallery wall. Listeners scan it to open the live translation and pick their language. If you ever rename the link, the old one keeps opening the same broadcast, so last year's leaflets still work.
A listener who opens the link early sees that the broadcast has not started yet. Someone who joins late sees everything said before they arrived, then follows live. A listener whose connection drops catches up on their own.
A broadcast is public by default, which is right for a service or a council session and wrong for a closed meeting or a seminar. For those, switch it to private and set a password: only people with the password can open the link, and they are asked for it before the transcript appears. Either way, say somewhere visible that the room is being transcribed and where the text goes. People praying aloud, or speaking for three minutes in public comment, did not necessarily come expecting a transcript. A line saying so in the order of service or on the agenda is cheap. Create the broadcast at {{app}}/broadcasts/new.
A link is only half of it; something has to put words on it.
Before any software, the audio. Rooms built for congregations and councils are built for reverberation: stone, high ceilings, wooden pews. A phone or laptop on the lectern hears the speaker once directly and then several more times off the walls. Distance plus echo is the single biggest cause of recognition errors, well ahead of accents. If the room has a mixing desk, take the feed from there (an aux or record output), so the recogniser hears what the microphones hear. Anyone who moves around, such as the preacher who leaves the pulpit or the lecturer at the whiteboard, should wear a lapel or headset mic rather than rely on one on a stand. Fix this and most of the rest is tuning.
Most churches, universities and councils already put that desk audio somewhere, and a live transcription can take it from there rather than from a device in the room.
For a stream, VocalStack pulls the audio itself; nothing runs on your side. The address does have to be reachable from the internet. A stream that only exists on your own network has to be recorded from a device in the room. Start a live transcription at {{app}}/transcriptions/new/live and attach it to the broadcast.
Once good audio arrives, the errors you will still notice are the predictable ones.
The words a speech recogniser gets wrong in a church, a lecture or a chamber are predictable: names of people and places, scripture references, technical terms, the items on an agenda. The reason is structural. A recogniser chooses the most probable words given the sound, and it learns those probabilities from what people usually say. So a rare surname loses to a common word that sounds like it, every time. The field's answer, known as biasing or custom vocabulary, is to tell the system which rare words to expect. The useful observation for a weekly room is that these are exactly the words already written down before the event.
Add a script or slides before you start. Names and terms from it then take precedence over what the recogniser misheard. Nobody has to read it word for word; a sermon outline, a handout or the published agenda already carries most of the proper nouns. During the recording, the Script tab on the host screen follows along like a teleprompter, lighting the words spoken so far.
Scripts can be written ahead and given a start time; the library lists them under Upcoming with a Start button. That suits a rota: one person adds the text on Friday, another presses Start on Sunday.
If a word still comes out wrong, the Fix tab marks uncertain words and lets the host click a line to correct it while the recording is live; listeners see the correction. AI-assisted auto-fix, on by default, corrects likely mistakes as lines arrive; each fix is marked and has an undo.
Getting the words right is one half; the other is knowing who said them.
The speaker-labels switch has three positions. When Needed labels each line with its speaker once more than one voice is heard; otherwise the transcript is plain text. On labels every line even for a single speaker. Off gives plain text. For a sermon, When Needed means one voice reads as prose; for a council of a dozen members, every line carries its speaker.
Labels start as numbers. In a recording you can name a speaker, and the name applies to every line of theirs. For people who speak every week, go a step further and save their voice. You can make a voice from any speaker who talks for at least thirty seconds in a recording, or from a short recording they provide. The speaker is then recognised and labelled by name in future recordings, and if listeners choose to hear the translation read aloud, it is read in that speaker's own voice in every language. Without any cloning, each speaker gets a distinct stock voice of matching gender, and any speaker can be muted.
A cloned voice is biometric data. VocalStack only makes one with the speaker's explicit consent, uses it only to read that speaker's lines aloud, and lets you delete it at any time. Councillors and clergy are public figures in their rooms; that is not the same as consent, so ask.
For a council this is where the transcript becomes a record. With every line labelled by name, the minutes finally show who said "I think we've covered this" for the fourth time. Which raises the question of who the record is for.
The first beneficiaries are the regulars who are deaf or hard of hearing, not the visitors reading in another language, welcome as they are. A hearing loop (an induction loop) helps people with hearing aids and does nothing for anyone else; captions on a phone help the rest, and the two complement each other rather than compete. Public bodies in many places carry accessibility and open-meeting duties, and a transcript the public can read live and afterwards goes a long way towards both. Check what applies to you locally.
Be honest about what automatic captions are. A human captioner (the service is called CART in some countries) listens, understands and types. On a heated council debate where three members talk at once, a human will still beat a machine. Overlapping speech produces errors in any automatic system, and so does the chair murmuring off-mic. Where there is a legal requirement for verbatim captions, automatic ones are a complement, not a replacement; where the alternative is nothing, they are a large improvement on nothing.
The transcript is also worth more the day after.
Every recording lands in your library, and the broadcast page lists past recordings for anyone with the link. Someone who missed the meeting can play it back exactly as it arrived, line by line, with its audio.
Once the words are final, a title and summary are written in every language. A search in your language also finds recordings made in another, so a clerk can find the month an item was discussed and a student the lecture that covered a term. Exports come as SRT subtitles, PDF, Excel or JSON, with corrections and translations included; a PDF of the translated transcript sits well beside the minutes. If any of this should happen without a person clicking, the API does everything the Dashboard does, at the same rate, in every plan: see {{api}}/docs.
All of this has to survive a parish or committee budget.
Live transcription is $0.50 per hour of audio, billed per minute, in every plan. Listeners are free, however many there are, and translation into up to 10 languages per transcription is included.
A new account comes with $5 of one-time signup credit and asks for no card. That credit covers more than a month of Sundays. After that you top up as you go, or take Premium at $50 a month ($40 a month on the annual plan). Premium includes $50 of credit every month, which is 100 live hours. It also keeps every recording for the life of the account instead of 30 days, and it does not cut a transcription off if the credit runs out mid-meeting. Two practical notes for the treasurer. On Pay as you go, a transcription can take up to 15 seconds to start from cold, so press start before the opening hymn rather than during it. And a YouTube live that has already ended is a video, so it is transcribed as a file, at the much lower file rate.
For one room and one recurring event, Pay as you go is enough. For an institution that wants every recording searchable years later, or runs several rooms at once, Premium is the one to pick. The full table, with limits and retention, is on the pricing page, and the interactive demo shows what a listener sees.
