Captions are won at the mixing desk and in the speakers' slides. Take one clean feed of the speech microphones per stage and load every name and acronym before anyone walks on; the rest of the day is fixing what those two missed.
Most of the work of captioning a conference happens before anyone walks on stage, and most of it is not software.
A phone on a lectern hears the room; the mixing desk hears the speaker. Take the feed from the desk, and be specific about which feed. Ask the sound engineer for a dedicated aux or matrix send carrying the speech microphones only: lectern, lapel mics, panel mics. Walk-in music, video playback and the applause stem stay out of it. Recognition that has to listen through a jingle makes a confident guess at the lyrics.
Test that send at the tech rehearsal, with the real microphones and the real speakers if you can get them. Otherwise you find out at 8:55 on the morning of the event that the matrix output is post-fader and that the engineer mutes the lectern between talks.
Give VocalStack that feed in one of two ways. The first is an RTMP encoder at the desk. When you add an RTMP stream as the source of a live transcription, the Dashboard shows the server URL and stream key the encoder needs, tells you whether audio is arriving yet and in what format, and lets you listen to it. The second is a stream address. If the stage is already going out as a YouTube live or an audio stream, paste that link and VocalStack pulls the audio itself; nothing runs on your side. Paste the stream's own address, not the page it plays on. If you paste a web page by mistake, the Dashboard tells you.
Either way, use one microphone feed per stage. Audience microphones and crosstalk cost accuracy that no setting wins back. The best microphone in the building belongs to the panellist who keeps leaning away from it.
Once the audio is coming in, attendees need somewhere to read the captions.
A broadcast is the link attendees open. Create it under Broadcasts, give it a name, and choose its link. The Dashboard proposes a link based on the name; you can edit it, as long as it keeps at least six characters. The same screen shows the QR code attendees scan to open the live translation and pick their language.
Decide now whether the broadcast is public or private. Public means anyone with the link can open it. Private means attendees enter a short password before the transcript appears. If the link has to change later, change it; the old one keeps working.
Then think about how the link reaches a thousand pockets. Put the QR code on the holding slide before each session, on the signage at the doors and on the lanyard card if there is one. Hundreds of phones joining at once put a real load on the venue Wi-Fi, so ask the venue what the access points were sized for and whether the guest network throttles per device. If the answer is a shrug, tell attendees to use mobile data. The pages are small; the Wi-Fi is the risk.
A conference with two stages does not need two links. One broadcast can be fed by several live transcriptions at once. Stage A and Stage B become two transcriptions attached to the same broadcast, and the attendee's phone shows both on one page. Each stage is its own live transcription, and each plan limits how many can run at the same time (1 on Pay as you go, 5 on Premium, unlimited on Enterprise). That limit comes back in the budget section below.
A link is only as good as the words behind it.
In any system, live recognition stumbles on names, product terms and acronyms. Add each speaker's script or slides before you start, and the names and terms in them take precedence over what the engine thinks it heard. During the talk the Script tab follows the words like a teleprompter.
This is also the moment to be honest about what machine captioning is. Speech recognition plus machine translation is not CART (Communication Access Realtime Translation, captions typed live by a trained human captioner), and it is not interpretation. It is fast, cheap and tireless, and it will occasionally invert a negation, round a number or translate a joke into a statement of fact. For a keynote, a product session or a panel, that trade is usually right. For a session where a wrong word has consequences (legal, medical, a vote), book a human captioner or interpreter and let the machine run alongside as the written record. The accessibility baseline is worth knowing too: WCAG 2.x asks for captions on live audio at level AA (criterion 1.2.4), and a growing number of public-sector and venue contracts simply cite it.
With the words settled, the last preparation is about who is saying them.
Speaker labels tell the audience who is talking. Set the switch to When Needed and labels appear only once a second speaker has been heard, so a keynote reads as plain text and a panel reads as a conversation.
To have speakers recognised by name from the first line, enrol their voices in advance under Voices. A clip of one person speaking is enough: MP3, M4A or WAV, at least thirty seconds long, though a few minutes is better. The person must have agreed, and the Dashboard asks you to confirm you have permission before it creates the voice. Treat that as more than a checkbox. Under the GDPR, a voiceprint used to identify a person is generally treated as biometric data (Article 9). That means explicit consent, and a right to have the voiceprint deleted. Comparable rules exist elsewhere, and speaker agreements are the easy place to cover it. A cloned voice can be deleted at any time. While it exists, it also lets listeners hear that speaker's translations in a matching voice.
With the plan settled, the rest is the day itself.
Every live transcription has a Speed tab with a latency mode and a chart of the delay over time. Three settings:
For a scripted keynote with a clean desk feed, Auto is right; for a technical panel with four accents and no script, Accuracy is worth the extra delay.
Start-up is the other timing to plan for. On Pay as you go, a cold start can take up to 15 seconds before transcription begins. Premium usually starts warm, and Enterprise always does. Press start during the walk-in, not on the first word.
Whatever the mode, some words will be wrong, so someone should be watching.
Someone at the desk should keep the Fix tab open. It has two switches: Mark uncertain words colours and underlines the words the engine was unsure of, and Click to correct opens a line for review when you click it. The tab walks you through the uncertain segments one by one, and a correction reaches every listener in every language.
AI-assisted auto-fix is on by default. It corrects likely mistakes as lines arrive, which helps with noisy audio and heavy accents, and every fix is marked and can be undone.
The audience microphone is the weak link in any room, and Q&A is where that shows. Ask the moderator to repeat every audience question into their own mic before answering. The captions will thank them, and so will the back row, who didn't hear it either.
Between sessions, the desk has a different job.
The host screen can pause and resume a recording, and the listener's page shows a notice: "Paused. The transcript continues when the host resumes." The pause is for a changeover, not a lunch break. Each recording has a maximum pause time, set in the Dashboard to 30 seconds, 1 minute or 3 minutes; a recording left paused for longer ends by itself.
For a longer gap, stop the recording and start a new one when the next session begins. Every recording stays listed on the broadcast's page, so in the afternoon attendees can scroll back to the morning's sessions. A single live recording has a ceiling of a few hours on Pay as you go and longer on Premium (the exact figures are on the pricing page), so a long day is several recordings in any case.
Midway through the morning, check who is actually reading.
The Broadcast tab's audience view shows who is in, how many are listening now, the languages they have open, and listeners over time. Each transcription includes up to 10 languages, and each listener picks one on their own device. The view therefore tells you which languages the room actually reads, which is rarely the list the organisers guessed. Delegates registered from Madrid turn out to read English; the Portuguese row fills up from a table nobody expected. Take the list home for next year.
Every one of those listeners is free. What costs money is the audio.
Live transcription is $0.50 per hour of audio, billed per minute, in every plan and through the API. Two stages running eight hours each is 16 hours of live audio, or $8.00. Translations, broadcasting, summaries and exports are included, however many people listen.
The plan matters for capacity rather than rate. Pay as you go gives $5 of one-time signup credit with no card required, one transcription at a time and a 5-hour maximum recording. A live transcription stops immediately when the credit reaches zero, so top up before the doors open. Premium is $50 a month ($40 a month on the annual plan) and includes $50 of credit every month, which covers 100 live hours. It allows five concurrent transcriptions and 10-hour recordings, and adds a grace period: a live transcription keeps running for a while past zero, and the extra minutes are billed at the same rate. Two parallel stages need Premium. Enterprise has custom pricing, unlimited concurrency, 24-hour recordings and a dedicated engine pool.
See the pricing page for the full table, and check your balance under Billing in the Dashboard the week before.
The bill stops when the audio does; the transcript doesn't.
Transcripts are kept for 30 days on Pay as you go and without limit on Premium and Enterprise. On Pay as you go, source audio is kept for only 7 days. Export within the week if you are on the free plan.
For how the link looks from the attendee's seat, see One link, every language, or try the live demo on the home page.
