Live Translated Audio for Wireless Conference Headsets: A Guide for Event Organizers and AV Teams

To add live translated audio to wireless conference headsets: start ConferenceCaptioning on a Mac, open the interpreter page, choose a language and a voice, and send the computer's audio output to a spare channel on your headset transmitter. Attendees switch to that channel on their headset and hear the translation about one second after each sentence ends (Rapid mode). Everything runs on the device, so you need no internet in the hall and no interpreter booth.

A two-minute demo with synthetic voices. It opens with footage of a real stage in Bali, Indonesia (10,000+ attendees); the headset demo itself uses synthetic voices. English comes out of the left ear (channel 1, from the mixer) and the translated French out of the right ear (channel 2, from the app). The narrator is in both ears. Words appear on screen shortly after they are spoken, and each French word shortly after its English word. The French is the app's real translation; the screenshot is the real interpreter page.

Video transcript

Narrator: You have built a big stage. And invited a global audience. But you have just one speaker. Introducing! Live translated audio! For wireless conference headsets! Every language! Right on the headsets your attendees already wear! Think about it, organizers: every single attendee hears the room in their own language! And AV teams? It runs right through your mixer, and uses the headsets you already own! The signal path? It is incredibly simple! Your mixer feeds a Mac running the app. Then the app sends those perfectly translated streams into separate channels, using your wireless headset transmitters! Just watch! Channel one carries the speaker in English. And now, thanks to ConferenceCaptioning, channel two carries the live translation in French! Channel 1, a woman speaking English: Good morning, everyone, and welcome to the conference. Channel 2, the app, in a female voice, in French: Bonjour à tous et bienvenue à la conférence. Narrator: Now, a man is speaking, so the dubbed French voice instantly switches to a male voice. Channel 1, a man speaking English: Thank you, Claire. Channel 2, in a male voice, in French: Merci, Claire. Narrator: Amazing! You can even choose the language, the voice, and the accent! Or simply turn on automatic detection for two people on stage! Look at the numbers! The translated voice starts just one second after each phrase or sentence ends, thanks to on-device AI! With more than 50 languages and accents to choose from, zero internet needed in the hall, and unmatched reliability thanks to automatic failover! So no hefty conference internet line fees, and no unpopular data centers used for AI! Setting up your headset system was the hardest part. Now, you can make it even more incredibly useful! Go ahead! Visit conferencecaptioning.com for a free demo today!

Key facts at a glance

What it does	Adds machine-translated audio channels to a wireless (silent-seminar-style) conference headset system. A Mac listens to the room, translates, and speaks the translation as a channel of its own.
Delay	About 1 second after a sentence ends in Rapid mode (1.1 seconds on average on our demo script).
Internet in the hall	None. Download the languages and voices once beforehand; everything else runs on the device.
Languages	More than 50 languages and accents to choose from. One language per interpreter page; add a computer for each extra language.
Voices	Choose the gender, the voice and the accent for each language. Optional automatic male or female voice for two people on stage (right about 96% of the time for one speaker at a time, tested on 66 real speakers).
Works with	Any headset system whose transmitter has a spare audio input. We have not tested every brand, so run the tone test first.
You need	A Mac running ConferenceCaptioning, a feed from the mixer or a microphone, a spare transmitter channel, and the headsets you already use. Attendees install nothing.
Also included	Live captions and translations on your screens, from the same app.
Not a replacement for	Human interpreters on legal, medical or other high-stakes sessions.

Who it is for

For event organizers

• Offer more languages without booking an interpreter for each one. Human interpreters for every language and every session are expensive, and often not possible. Translated channels let you add languages to sessions you could not staff.
• Attendees use what they already know. They pick up the headset they would wear anyway and switch to the translation channel. There is no app to install and no QR code needed for this channel.
• No dedicated internet line for translation. The AI runs on the device in the room, so you do not need to buy a conference-hall internet line for it (such lines commonly cost about $2,500 to $5,000), and the audio never leaves the room.
• Captions for everyone, too. Put live captions on the room's screens from the same app, so deaf and hard-of-hearing attendees are included as well.

For AV production engineers

• Simple signal path. The mixer (or a microphone) feeds the Mac, the Mac's sound output goes to one spare channel on the transmitter, and the English stays on channel 1 straight from the mixer. The routing steps are below.
• Test before doors open. Test tones for left, right and both, and a speak-sample button, let you check every cable and channel with a headset in hand.
• Plan the extras. Give the Mac a static IP, keep it awake, open one interpreter tab per language (each on its own sound output), and add a backup device for mission-critical shows (see the Pro Tips).
• No cloud dependency. Translation and voices run on your side of the room, so a venue Wi-Fi problem does not take the translated channel down.

Why add translation channels to wireless headsets?

Silent-seminar-style headset systems already solve the hard part of listening in a busy hall: every attendee wears a headset and picks their own channel, and the systems on the market offer many of them. For example, the makers of the conference headsets at Silent Seminars describe 10 and 45 channel systems with programmable, lockable channels. What each extra channel needs is an audio source in another language. Booking human interpreters for every language and every session is expensive, and often not possible. ConferenceCaptioning turns a computer into that source: it listens to the room, translates, and speaks the translation, on a channel of its own.

What you get

• The translation read aloud, sentence by sentence, with a voice you choose, including the Premium voices that ship with macOS.
• Your choice of voice: pick the gender, the voice and the accent your listeners hear for each language, or turn on automatic gender detection so the voice follows who is speaking when two people share the stage.
• Test tones for the left, right and both sides, and a "speak sample" button, so your AV team can check every cable and channel before doors open.
• Setup instructions on the page itself for Dante Virtual Soundcard, USB interfaces and a second computer.
• No internet in the hall. The page, the translation and the voices all run on your side of the room.
• On-screen captions at the same time. The same app shows live captions and translations on your screens, which headsets cannot do for attendees who are deaf or hard of hearing.

How it works

A microphone or the mixer feeds the app. The app captions the speech, translates each finished sentence, and the interpreter page reads the translation aloud through the computer's sound output. That output goes to your transmitter, which sends it to the headsets as one channel. The page runs in a web browser on the computer next to the transmitter (or on any computer on the same network), and the app serves it at http://{device IP}:8080/interpreter.html.

Set it up in six steps

Step 1: Prepare, once, with internet

1. Install the app and download the translation languages you will use.
2. Install a good voice for each language: on a Mac, System Settings > Accessibility > Spoken Content > System Voice > Manage Voices, and download a Premium or Enhanced voice.
3. After installing a voice, fully quit your browser (Cmd+Q) and reopen it. Browsers only list new voices after a restart. Online-only voices will not work in a hall with no internet, so the page never picks them on its own.

Step 2: Start captioning

Connect your audio source (mixer, audio interface, Dante or a microphone), choose the speaking language and the language to translate to, and start captioning in Rapid or Mix.

Step 3: Open the interpreter page

1. In the app, tap the Local pages button (the network icon) and open the interpreter card, or open http://{device IP}:8080/interpreter.html in a browser.
2. Choose the language, pick a voice, and press Start speaking. Browsers require one click before they play sound.

Step 4: Send the audio to a channel of its own

By default the page speaks through the computer's current sound output. To send it to a specific device instead, open the page in Chrome as http://localhost:8080/interpreter.html on the Mac running the app, press Show device names (Chrome asks for the microphone once only to reveal the names; nothing is recorded) and pick a device under Sound output. The app then renders each sentence with the same voice and the page plays it on that device. Then route that device to the transmitter:

• Dante Virtual Soundcard: set the output to Dante Virtual Soundcard and subscribe only TX channel 2 to the transmitter's channel 2 input. English stays on channel 1 from the mixer.
• USB interface or headphone jack: use a stereo-to-two-mono splitter and connect only the right leg to the transmitter's channel 2 input.
• Another computer on the same network: open the same address on it and wire its output the same way. This is another way to add a language.

More than one language from one computer

When the app is translating into several languages, the page lists them under Language to speak and explains this: one page speaks one language. Press Open … in a new tab to open another tab for the next language (the link carries ?slot=1 for the second language, and &out=Headphones picks an output by name), choose that tab's Sound output, and repeat for each language. Every tab keeps its own language and output, so one Mac can feed several transmitter channels, not just two. In our test, one tab spoke French on the headphones and another Arabic on the Mac speakers at the same time. We have not measured CPU or memory with many tabs running, so check a full set-up before the event.

Step 5: Test the whole chain

Press Tone: left, Tone: right and Tone: both, and Speak sample with this voice, then listen on a headset tuned to the channel. Do this before the event, with the transmitter powered and the headsets in range.

Step 6: Turn on "Match the speaker" (optional)

Tick Match the speaker, then choose a male voice and a female voice for the language. Each sentence is now read in the voice for whoever was speaking when it was said. The page shows what it hears, for example Speaker sounds female (about 243 Hz). Leave "keep the voice used last" ticked: it stops the voice changing when the app is not sure.

---

Choose the voice, or let the app match the speaker

For each language you choose the gender, the voice and the accent your listeners hear, for example a British or a Canadian French voice. If one person speaks, that is all you need. If two people share the stage, turn on automatic gender detection and give the page a male and a female voice.

Hearing a translation in a voice that matches the person on stage makes a conversation easier to follow, especially in a panel or a Q&A. The app measures the pitch of the microphone audio, which separates most male voices (typically about 85 to 155 Hz) from most female voices (about 165 to 255 Hz), and the interpreter page picks the voice you chose for that sound. Nothing is shown to the audience, no audio is stored, and you can switch it off.

It is a hint, not a fact. Pitch alone cannot separate every voice: a low-pitched woman or a high-pitched man sits in a zone where the app says "not sure" and keeps the voice it used last. Please read the numbers below before you rely on it for an event with many speakers.

How fast is it, step by step?

After a speaker finishes a sentence, the translated voice starts about one second later in Rapid mode. This is where that second goes, measured on a Mac with the demo script:

Step	Time
The speech recognizer decides the sentence is finished	about 0.6 s (0.3 to 1.0 s)
The app notices, translates (0.03 s) and hands the text to the voice page	about 0.2 s
The voice starts speaking	about 0.1 s
Total when the voice is free	about 0.9 s

The biggest piece is the recognizer, which waits for a short pause before it will call a sentence final. We used to wait for the next sentence to start before speaking, which added about two seconds; the app now speaks a sentence as soon as the recognizer finalizes it. A sentence only waits longer when the voice is still reading the one before it, which happened to about a third of the sentences in our demo script and added about a second each.

How good is it?

We measured the whole chain on a short demo script (a host and a speaker taking turns), and measured the voice matching separately on real recorded speech. These are our numbers, with what they do and do not show:

What we measured	Result	What to know
Delay: end of a sentence to the translated voice being called (Rapid)	1.1 seconds on average. Typically 0.9 to 1.3 s, 95% under 2.1 s. Plus about 0.1 s for the voice itself to start.	French, one Mac, the 11-line demo script, synthetic voices. Before our November 2026 speed-up the same test gave 3.3 s. Mix, measured earlier, gave 2.8 s.
Delay: end of a sentence to the first sound, voice free (Rapid)	About 0.9 seconds.	0.6 s recognizer + 0.2 s app + 0.1 s voice start-up. Measured in pieces on the same Mac; the pieces add up to about the total.
Translation of the demo script	11 of 11 lines with no meaning errors. Similarity to a human-written translation: 79 and 83 out of 100 for the two translation modes.	Checked by hand against a reference. We reworded one line that machine translation gets wrong (see the note below).
Male or female voice chosen correctly, demo script	12 of 12 sentences with Mix, 12 of 14 with Rapid.	Two clearly different voices. Real speakers vary more.
Voice chosen correctly, 66 real speakers	96% of the voice decisions were right, one speaker at a time. The voice changed in the middle of one person's speech in 1% of recordings.	Read audiobook speech in quiet rooms, tuned on other speakers. Not yet tested in real conference rooms.
Voice chosen correctly, people taking turns every few seconds	81% to 84% of sentences got the right voice.	The first sentence after a change is the one most likely to keep the previous voice.

Speech recognition accuracy depends on the speaker, the microphone and the room, so we do not quote one number for this setup. Our open measurements of the captions themselves are in the live captioning benchmark.

What to know before the event

• It is machine translation. It is very good on everyday conference language, but idioms can come out wrong: in our own tests "Exactly right" was translated as the French for "exactly to the right". Run your key phrases through it first, and keep human interpreters for legal, medical or other high-stakes sessions.
• Fast talkers. A voice cannot read faster than people comfortably listen. If the speaker talks continuously and quickly, the translation falls behind. In one 16-minute test with continuous speech, half of the sentences waited more than 4.6 seconds because earlier sentences were still being read. The page then speaks up to 30% faster to catch up. For a fast speaker, ask them to leave short pauses.
• One language per page. Open the page once per language; each tab can send its sound to its own output device (Chrome, opened as localhost on the Mac running the app). Several languages from one Mac are tested with two tabs; larger set-ups and Dante or USB hardware are not yet tested.
• Overlapping speakers are not handled: when two people talk at once the voice choice is a guess.
• Your headset system is your own. We have not tested every brand of transmitter. Anything that takes an audio input and sends it to a channel should work, and the tone buttons tell you quickly.

---

Pro Tips

1. Keep the computer awake (System Settings > Lock Screen, or run caffeinate in Terminal) and keep the interpreter tab open and in front of nothing that mutes it.
2. Keep the address from changing: give the computer a static IP so the page address works from one event to the next. Here's how on a Mac.
3. Speed: if sentences pile up, the page speaks a little faster, so the headsets stay close to live.
4. Mission-critical show? Add a backup device with hot-failover redundancy.
5. Put captions on the screens too. Headsets help people who hear; screens help everyone, including deaf and hard-of-hearing attendees.

Frequently asked questions

Does this work with silent seminar and wireless conference headsets?

Yes, in principle. The app produces audio from a computer, and any wireless-headset system with a spare audio input on its transmitter can carry it as a channel of its own. Attendees switch channels on the headset as they always do. We have not tested every brand of transmitter, so test your own chain with the tone buttons before the event.

Do I need internet in the hall?

No. Download the translation languages and install the voices once at the hotel, and everything else runs on the device. The interpreter page works over your local network. The computer needs internet only for that one-time setup.

How long after a sentence ends does the translated voice start?

About one second after a sentence ends in Rapid mode, when the voice is free: roughly 0.6 seconds for the speech recognizer to decide the sentence is finished, 0.2 seconds for the app to translate it and hand it over, and 0.1 seconds for the voice to start. On our demo script the voice was called 1.1 seconds after the end of a sentence on average. Mix mode, which we measured earlier, took about 2.8 seconds. If the speaker talks faster than the voice can read, the voice falls behind: it then speaks a little faster and skips ahead rather than drifting further.

How many languages can I send at once?

One language per interpreter page, but you can open the page several times on the same Mac: each browser tab takes one language and sends it to its own sound output (Sound output on the page), so one computer can feed more than two channels. Use Chrome, open the page as http://localhost:8080/interpreter.html on the Mac running the app, and press Show device names once per tab. We tried two tabs at once with real voices (French to the headphones, Arabic to the Mac speakers); we have not yet tested it with Dante or a USB interface on a transmitter, or measured how many languages a Mac can run at once. Another computer per language still works too.

How does the automatic male and female voice work?

The app measures the pitch of the microphone audio, which tells it whether the speaker sounds male or female, and the interpreter page reads each sentence in the male or female voice you chose. It is a hint, not a fact: in our tests on 66 real speakers it was right about 96% of the time for one speaker at a time, and 81% to 84% of sentences got the right voice when people took turns every few seconds. When it is unsure it keeps the voice used last, so the voice does not flip back and forth while one person talks. No audio is stored.

What equipment do I need to add translated audio to a headset system?

A Mac running ConferenceCaptioning, an audio feed from your mixer or a microphone, a spare audio input (channel) on your headset transmitter, and the headsets you already use. Attendees need nothing new: they switch to the translation channel on their headset. A sound-routing tool such as Dante Virtual Soundcard, a USB audio interface or a headphone-jack splitter sends the computer's sound to that one channel.

Is this a replacement for human interpreters?

It is machine translation. It is very good on everyday conference language, but idioms can come out wrong, so keep human interpreters for legal, medical or other high-stakes sessions. For most sessions it lets you offer more languages than you could staff with interpreters.

Which languages are supported?

More than 50 languages and accents to choose from. For each language you choose the voice and the accent your listeners hear. Each interpreter page speaks one language, so open it once per language (one tab each, or one computer each) when you need several at once.

What if the computer fails during the event?

ConferenceCaptioning supports hot failover to a backup device for mission-critical shows (see our zero-downtime guide). Whatever your setup, keep a spare computer on the same network and test the switch with your own transmitter before doors open.