Someone across the counter speaks Japanese, and a second or so later you hear English in your ear. It feels like the earbuds are doing the translating. In the products covered below, from Apple, Google and Timekettle, the earbuds are a microphone and a speaker, and the work happens on the phone they're paired with or on a server the phone talks to.
This explainer walks through what happens in that second or so: the four stages every system runs, where the processing happens on AirPods, Pixel Buds and dedicated translator earbuds, why you always wait for the end of a sentence, what breaks in noisy rooms, and how you'd build the same pipeline into your own earbuds or app.
How do translation earbuds work?
Translation earbuds pick up speech with their microphones, send the audio to a paired phone or a cloud service that recognizes the words and translates them, and play the translation back into your ear as synthesized speech. The earbuds themselves are mostly a mic and a speaker. The speech recognition, translation and voice all run on the phone or in the cloud.
That split explains most of what you'll notice in practice: why some earbuds work in airplane mode and others don't, why there's a pause before you hear anything, and why the phone screen is part of the conversation.
The four stages of live translation
Every live translation system, from AirPods to a developer's prototype, runs the same chain. It's the same cascade we cover in depth in our guide to real-time speech translation.
- Capture. The earbud microphones, and sometimes the phone's microphones, record the other person's voice and stream the audio over Bluetooth to the phone.
- Speech recognition. A streaming speech-to-text model turns the audio into text in the source language as the person talks, and decides when they've finished a sentence.
- Translation. A machine translation model or an LLM translates the finished sentence into your language.
- Speech output. A text-to-speech model reads the translation aloud, and the phone streams that audio back to your earbuds. Many apps also show the text on screen.
| Stage | What happens | Documented figure (AssemblyAI pipeline) | Source |
|---|---|---|---|
| 1. Capture | Mics record speech and stream it to the phone or app | Not published: depends on the earbuds and Bluetooth link | Measure on your hardware |
| 2. Speech recognition and end of sentence | Speech becomes text, and the model waits for a pause to decide the sentence is done | AssemblyAI-STT-3.7-Realtime checks for the end of a turn after at least 320 ms of silence, and forces it after 640 to 2560 ms depending on the mode | AssemblyAI turn detection docs |
| 3. Translation | The finished sentence is translated | "Typically within about 100 ms of the final turn" with a small model | AssemblyAI real-time translation docs |
| 4. Speech output | The translation is voiced and played in your ear | About 90 ms time to first audio for AssemblyAI-TTS-1 | AssemblyAI pricing page |
These figures come from different measurements, so they don't add up to an end-to-end number. The Apple, Google and Timekettle pages cited below don't publish per-stage figures for their earbuds.
On-device vs cloud translation
Stages 2 to 4 have to run somewhere. Earbuds that pair with a phone have two options: run the models on the phone, or send the audio to a server.
On the phone. Apple's Live Translation with AirPods is the clearest on-device example. Apple's support page says you download the language models first, and that "all processing takes place on your iPhone where all of your conversation data remains private." It needs AirPods 4 (ANC), AirPods 5, AirPods Pro 2, AirPods Pro 3 or AirPods Max 2, paired with an iPhone 15 Pro or later on iOS 26 or later with Apple Intelligence turned on. Apple's feature-availability page lists 11 languages and variants: Chinese (Mandarin, Simplified and Traditional), English (UK and US), French, German, Italian, Japanese, Korean, Portuguese (Brazil) and Spanish (Spain).
In the cloud. Google's help page for translating with Pixel Buds lists an internet connection as a requirement, along with the Google Translate app, and lists 40 languages for conversation translation.
Both. Timekettle's FAQ for its W4 Plus earbuds describes a cloud translation engine with 52 languages online, plus offline packs for 13 language pairs that you download while connected. Offline covers its two-person conversation and listening modes. Call and media translation need the internet.
| Trade-off | On-device (models on the phone) | Cloud (models on a server) |
|---|---|---|
| Works without a connection | Yes, once the language models are downloaded | No: needs Wi-Fi or mobile data |
| Where the audio goes | Stays on the phone | Sent to the provider, under its data retention terms |
| Model size and languages | Limited by phone memory and battery, so models and language lists tend to be smaller | Larger models and longer language lists are practical |
| Updates | Arrive with OS or app updates and new model downloads | The provider can update models on the server at any time |
| Phone battery | The phone runs every model | The phone mostly streams audio and plays the result |
For anyone building a product, this is a design choice. A cloud pipeline gives you larger models and more languages to work with, and the trade is a dependency on the network.
Follow the full build: stream audio from a device, translate each finished sentence, and send captions to a display or speech to paired earbuds.
Why there's a delay before you hear the translation
Live translation always runs a beat behind the speaker, and most of that delay is deliberate. The system waits until the speaker finishes a sentence, or at least a clause, before it translates.
It has to. Languages put words in different orders, so the first half of a sentence often can't be translated until the second half arrives. A German sentence can put its main verb at the very end. In Japanese, the negation that turns "I'm going" into "I'm not going" comes last. Translate the first half early and you might say the opposite of what the speaker meant.
So the speech recognition stage watches for the end of a turn: a pause long enough, and a transcript that reads as complete. Google's Pixel Buds page describes the same rhythm from the listener's side: "The text is read aloud after each sentence is translated." AssemblyAI's real-time translation docs are just as explicit about the developer side: "Only final turns are translated."
After that, the remaining stages are short by comparison. In an AssemblyAI pipeline, the translation typically lands within about 100 ms of the final turn with a small model, and AssemblyAI-TTS-1 has about 90 ms time to first audio. Turn detection is the part you can tune. The turn detection settings trade quicker turns against more fragments: end turns sooner and you translate more half-thoughts, wait longer and the listener waits too.
What breaks in practice, and how to fix it
Live translation works best with one person speaking clearly, close to the microphone, in a quiet room. Real conversations aren't like that. Here's what goes wrong and what helps.
- Noise and distance. Earbud microphones sit in the wearer's ears, and the other person may be a meter or more away. Apple's support page notes that in noisy places the iPhone's microphones can help, and suggests moving the iPhone closer to the other speaker. In your own build, put a microphone near the speaker, and if your app is a conversational agent, the Voice Focus option on AssemblyAI-STT-3.7-Realtime-Agent suppresses background audio before it reaches the model.
- Two people talking at once. Overlapping speech confuses recognition and leaves the translator with a mixed sentence. The two-way modes below separate the sides by device: one person on the earbuds, the other on the phone. For captions of a group, speaker labels in the transcript tell you who said each line.
- Switching languages mid-sentence. Bilingual speakers often mix languages. A model that only expects one source language garbles the switch. AssemblyAI streaming is multilingual by default and follows code switching with no configuration, and the translation prompt can say "translate from whatever language is spoken."
- Names and specialist terms. A person's name, a street or a drug name can come out as a common word and then get translated as one. Keyterm prompting lets you pass a list of up to 100 terms per session for the model to recognize, and the translation prompt can carry a short glossary.
One-way listening vs two-way conversation
Most translation earbuds offer two modes, and they work differently.
One-way listening is for a tour guide, a lecture or announcements. You wear the earbuds, the system listens to the room, and you hear the translation. Nobody needs to hear you. Google calls this listening mode on Pixel Buds, which streams translated audio into your ear while your phone stays in your pocket.
Two-way conversation needs the other person to understand you, too, and the products solve it in different ways:
- The phone speaks for you. With AirPods, Apple's page says you can show a transcript on the iPhone's Translate app or play the translation through the iPhone speaker. Pixel Buds conversation mode does the same: your translated speech plays from the phone, and theirs plays in your earbud.
- Both people wear earbuds. If the other person has supported AirPods with Live Translation running, Apple says they can hear your translated response in their own AirPods.
- Share the pair. Timekettle's two-person conversation mode has you hand one earbud to the other person, so each of you hears the other's translation.
Under the hood, two-way means two pipelines running at once, one per direction, each with its own source and target language.
Building live translation into your own earbuds or app
If you're building a translation feature for earbuds, a hearing device or an app, you can run the cloud version of the four stages on AssemblyAI with one streaming connection and one TTS session:
- Speech recognition: AssemblyAI-STT-3.7-Realtime (
assemblyai-stt-3-7-realtime, $0.30/hr) transcribes 32 languages and fits one-way and conversation translation, where software doesn't reply to the speaker. It's the captioning model in the STT 3.7 family. - Translation: the
llm_gatewayconnection parameter sends each final turn to LLM Gateway with your prompt and returns the translation on the same WebSocket. It translates final turns only, so partials can drive a live caption in the source language. - Speech output: AssemblyAI-TTS-1 voices the translation in English, Spanish, French, German, Italian or Portuguese. For any other target language, show captions on the phone.
Here's the connection, using the Python SDK (1.10.1 or later). It translates whatever language is spoken into English and sets up an English voice for playback:
PROMPT = (
"Translate this turn from whatever language is spoken into natural English. "
"Reply with only the translation: {{turn}}"
)
client.connect(
RealTimeParameters(
sample_rate=16000,
speech_model="assemblyai-stt-3-7-realtime",
llm_gateway=LLMGatewayConfig(
model="qwen3.5-4b-32k-fast",
messages=[LLMGatewayMessage(role="user", content=PROMPT)],
max_tokens=400,
),
)
)
tts_config = aai.TTSConfig(voice_id="aai-kyla", language="english")Each translation arrives as an LLMGatewayResponse message tied to its source by turn_order. Queue it into one long-lived TTS session and stream the audio to the earbuds. For the full program, including microphone capture, caption rendering and the speech step, follow our tutorial on live translation for smart glasses and wearables. To compare other speech translation APIs, see our roundup of the best real-time speech translation APIs, and for batch translation of recorded audio, the translation feature page.
Stream audio to AssemblyAI-STT-3.7-Realtime, translate every finished sentence through LLM Gateway, and voice it with AssemblyAI-TTS-1. Sign up free to get an API key.
What to expect from translation earbuds
Translation earbuds are a mic, a speaker and a pipeline. Where that pipeline runs decides whether they work offline, which languages they speak, and where your conversation goes. The pause before each translation is the system waiting for a sentence it can translate correctly, and the clearest way to shorten it is better turn detection, since the translation and voice stages already take a fraction of a second. Expect the next improvements to come from that end of the pipeline: earlier and more accurate decisions about when someone has finished speaking, and better handling of noisy rooms and mixed languages. For a deeper look at the streaming side, see our guide to real-time speech-to-text.
Get an API key, stream speech in any of 32 languages, and get a translation of every final turn back on the same connection.
Frequently Asked Questions (FAQs)
How does live translation work with AirPods?
AirPods Live Translation uses the AirPods microphones to pick up the other person, translates their speech on your iPhone with downloaded language models, and plays the translation in your ear. Apple says all processing takes place on the iPhone. It needs supported AirPods (AirPods 4 with ANC, AirPods 5, AirPods Pro 2 or later, or AirPods Max 2) and an iPhone 15 Pro or later on iOS 26 or later with Apple Intelligence.
Do translation earbuds work offline?
Some do. Earbuds that translate on the phone with downloaded language models, such as AirPods Live Translation, work without a connection. Others need the internet: Google lists an internet connection as a requirement for translating with Pixel Buds. Some offer both, such as Timekettle's W4 Plus, which supports offline packs for 13 language pairs.
How do language translator devices work?
Language translator devices run a four-stage chain: capture the speech with a microphone, convert it to text with speech recognition, translate the text with a machine translation model or an LLM, and speak the result with text-to-speech. In earbuds, the earbuds handle capture and playback, and a paired phone or cloud service runs the models.
Why is there a delay before I hear the translation?
The system waits until the speaker finishes a sentence before translating, because word order differs between languages and a half-finished sentence can translate wrongly. Most of the delay is that wait for the end of a turn. The translation and voice steps are short by comparison: about 100 ms for translation with a small model and about 90 ms to first audio for AssemblyAI-TTS-1, in an AssemblyAI pipeline.
Can translation earbuds handle two people talking at once?
Not well. Overlapping speech gives the recognizer a mixed signal, so the products covered here separate the two sides: one person wears the earbuds and the other speaks into, or reads from, the phone. Conversation modes on AirPods, Pixel Buds and Timekettle all work this way.
Can I build live translation into my own earbuds or app?
Yes. Stream audio from the device to a real-time speech-to-text API such as AssemblyAI-STT-3.7-Realtime, translate each final turn with the llm_gateway parameter, and voice the result with a TTS model such as AssemblyAI-TTS-1 in six languages, or show captions for any other language. The smart glasses and wearables tutorial on the AssemblyAI blog walks through the full build.