Live, on-screen text of what the other person is saying on a phone or video call, transcribed by streaming AI speech recognition as they speak — with sub-second latency, not after the call.
Last verified · 2026-07-17 · by Moe Ameen
Real-time phone call captions are the words of a live call rendered as text on screen while the conversation is still happening. A streaming automatic speech recognition (ASR) engine listens to the audio, emits a running guess word by word, and finalizes each phrase as the speaker moves on — the whole loop targeting well under a second of latency so the reader stays in step with the talk. This is the same core technology behind live meeting captions and burned-in video subtitles, but pointed at a two-party voice call, where the audio is narrowband, unscripted, and often has both people talking at once.
The term spans two worlds that use nearly identical tech for different ends. The first is accessibility: captioned telephone service for people who are deaf or hard of hearing, plus on-device features like Google's Live Caption and Apple's Live Captions that transcribe any call audio locally on the phone. The second is business voice: contact-center and unified-communications platforms — Dialpad, RingCentral, and the like — that transcribe sales and support calls in real time to feed live agent assist, sentiment scoring, and a searchable transcript afterward.
Under the hood, quality turns on a handful of things: streaming (not batch) inference so text appears as speech arrives; speaker diarization to label who said what on a two-way call; custom vocabulary so names, brand terms, and jargon get transcribed correctly; and robustness to the 8 kHz narrowband audio, accents, and crosstalk that make phone calls harder to transcribe than a clean studio recording. Top consumer live-transcription engines land around 85–95% accuracy in quiet conditions, versus 98–99% for a professional human transcriptionist — a gap of a few word-level errors per minute.
For a creator or operator, the caption itself is the visible feature, but the durable asset is the transcript it leaves behind. A live-captioned discovery call, podcast interview conducted over the phone, or customer support conversation ends with a clean, timestamped, speaker-labeled text record — which is exactly the kind of raw source a content engine can turn into posts, clips, and articles.
Live phone captioning began as an accessibility service, not a consumer feature. Captioned Telephone Service (CTS), and later its internet version IP CTS, gave people with hearing loss a phone with a screen that displayed real-time captions of the other party's speech. In the US these services are regulated by the FCC and funded through a charge on consumers' phone bills, provided at no cost to eligible users. For years the captions were produced by a human communication assistant who re-voiced the call into speech-recognition software; over the 2020s the FCC certified providers such as CaptionCall to caption using automatic speech recognition alone, moving the human out of the loop as ASR accuracy caught up.
The consumer, non-accessibility version rode the same ASR curve. Google shipped Live Caption in 2019 on the Pixel 4, transcribing any audio playing on the device on-device and privately, and later extended it toward call audio. Apple previewed Live Captions in May 2022 — in an accessibility-features announcement ahead of Global Accessibility Awareness Day, not at WWDC — and shipped it in iOS 16 later that year, generating on-screen text for FaceTime and phone-call audio locally on the iPhone so the audio never leaves the device. Both leaned on the same insight that made short-form video captions ubiquitous — that people increasingly consume audio as text — applied to the phone.
In parallel, business telephony absorbed real-time transcription as a core feature rather than an add-on. Cloud phone systems and contact-center platforms built streaming ASR directly into the call path, so a transcript scrolls live during the call and powers agent-assist prompts, compliance monitoring, and post-call summaries. By 2026 the expectation for a serious live-transcription API had hardened: finalize text in well under a second, label speakers inline, accept turn-by-turn custom keyterms to steer recognition mid-conversation, and hold up when a speaker switches languages — the difference between a passive after-the-fact transcript and the live backbone of a responsive voice agent.
| Platform | Behavior |
|---|---|
| iPhone (Apple Live Captions) | On-device live captions for phone and FaceTime audio, introduced with iOS 16 (announced May 2022). Runs locally so audio never leaves the device; a system-wide accessibility feature rather than a per-app one. Accuracy is solid for clear speech, weaker on heavy accents and crosstalk. |
| Android / Pixel (Google Live Caption) | On-device captioning first shipped on the Pixel 4 in 2019 for any playing media, later extended toward calls. Private by design (no cloud round-trip) and works offline, at the cost of a smaller model than a cloud service can run. |
| Captioned phones (IP CTS / CapTel) | FCC-regulated accessibility service for people with hearing loss, funded by a fee on US phone bills and free to eligible users. Captions come from ASR alone or ASR plus a human communication assistant. FCC rules impose strict confidentiality and generally bar providers from retaining call content after the call. |
| Business phone systems (Dialpad, RingCentral) | Streaming transcription built into the call path, live during the call and searchable after. Powers agent assist, sentiment, and auto-summaries. Custom vocabulary for account and product names materially lifts accuracy on domain terms. |
| Video-call platforms (Zoom, Meet, Teams) | Not phone calls, but the same real-time captioning applied to VoIP meetings — live captions plus a saved transcript. Cleaner audio than a cellular call, so accuracy tends to be higher, and speaker labels come from the platform rather than pure diarization. |
| Voice agents / IVR | Real-time captions here are the input to an AI, not a human reader — the transcript stream feeds a language model that decides the agent's next turn. Latency matters most in this mode, since every extra fraction of a second of transcription delay is added to the agent's response time. |
The feature everyone looks at is the scrolling text; the thing worth keeping is the transcript. Real-time phone captions are, functionally, a free speech-to-text pipeline running on conversations you're already having — discovery calls, customer interviews, expert calls, phone-recorded podcast episodes — and each one ends with a clean, timestamped, speaker-labeled document. That document is content raw material, and most people delete it.
To be clear about the boundary: Kompozy does not caption your live calls. The real-time layer belongs to your phone's OS, your accessibility service, or your contact-center platform, and it should stay there. Where Kompozy picks up is the moment the call ends and the transcript exists. That transcript is exactly the kind of source it ingests — the same [content-repurposing](/glossary/content-repurposing) flow that turns a podcast or a newsletter into a week of posts works on a call transcript. Paste or pipe it in, and the engine can spin out text posts, an article, [quote graphics](/glossary/subtitle), and even a [Persona Short](/glossary/persona-shorts) where your avatar delivers the best line from the call, all governed by your [Persona Brief](/glossary/persona-brief) so it sounds like you and not like a transcript.
The practical habit: stop treating call captions as an in-the-moment convenience and start treating the transcript as an input. The calls you're already on are a content source hiding in plain sight — the captioning tech just made them machine-readable for free.
They are live on-screen text of what the other person is saying during a phone or video call, produced by streaming AI speech recognition as they speak rather than after the call. The system emits a running word-by-word guess and finalizes each phrase, targeting under a second of latency so the reader keeps pace with the conversation.
Top consumer live-transcription engines reach roughly 85–95% accuracy in quiet conditions, versus 98–99% for a professional human transcriptionist. Phone audio is narrowband and often has crosstalk and accents, so expect a few word-level errors per minute. Loading custom vocabulary for names and jargon can add another 5–10% on domain-specific terms.
It depends on where the transcription runs. On-device features like Apple Live Captions and Google Live Caption process the audio locally so it never leaves your phone. Cloud-based business transcription services send call audio to a server. Accessibility services regulated by the FCC (IP CTS) are bound by strict confidentiality rules and generally cannot retain call content after the call.
The technology is nearly identical, but the purpose differs. Accessibility captioning — captioned telephone service and on-device Live Captions — helps people who are deaf or hard of hearing follow a call. Business transcription on platforms like Dialpad and RingCentral captions sales and support calls to power live agent assist, sentiment scoring, and a searchable post-call transcript.
Yes, and that is often the most valuable output. A live-captioned call ends with a clean, timestamped, speaker-labeled transcript, which is exactly the kind of raw source a content engine can repurpose. Kompozy does not caption the live call itself, but once the transcript exists it can turn it into text posts, an article, quote graphics, or a persona video, governed by your Persona Brief.
Usually, yes. A caption is a recording in text form, and many jurisdictions require one-party or all-party consent to record or transcribe a call. Check the rules where you and the other party are located before enabling live transcription on a call, especially for business or legal use.