Skip to content

Feature · Real time transcription

Real time transcription for live conversations and apps.

PyAI Hear streams words back over a WebSocket mid sentence, then corrects them once the sentence closes.

Get a sandbox key$50 credit. No credit card required.
Earliest partial in ~200 ms8 kHz audio WER 4.76%$0.001/min of speech onlyAccepts 8 kHz or 16 kHz PCM16

What PyAI gives you for real time transcription

  • PyAI real time transcription streams the first partial word in about 200 ms in-region and bills $0.001 per minute of detected speech, so your app shows text while the speaker is still talking.
  • Hear measured a 4.76 percent word error rate on live 8 kHz phone audio and 1.59 percent on clean speech, which puts it at parity with major models rather than ahead of them.
  • Billing counts only detected speech, so an open but silent socket adds nothing and the closing usage message lists the exact billed seconds.
  • Browser microphones send 16 kHz audio and phone bridges send 8 kHz, both as raw PCM16 in about 20 ms slices, and endpointing values arrive in a settings reply and change mid session without reconnecting.

What does a live transcript look like as words arrive?

Choose a scenario and press play to watch the socket events Hear emits. The lines are an example script, not a recording.

Pick a moment, then press play

SpeakerYour appPyAI Hear
Ready00:00

How does real time transcription work with PyAI Hear?

PyAI Hear keeps recognition in the same region as the socket, so each draft word is pushed to your app the instant the recognizer produces it.

Socket handshake

Your code opens a secure WebSocket to the PyAI Hear streaming API, with the key inside the subprotocol. The PyAI Edge verifies the key, the streaming permission and a free concurrency slot.

Session settings returned

Hear answers with a settings reply listing the endpointing values, after pulling in any vocabulary saved for the project in Console settings.

Audio upload begins

Binary PCM16 slices of about 20 ms flow up at 8 or 16 kHz, and recognition starts on the very first slice.

Draft text comes down

Each partial splits into settled text and draft text, which Hear may still revise when more audio lands.

Committed text follows

An end-of-sentence event signals the sentence is over. A final event then delivers punctuated, corrected wording for that utterance.

Your code closes the socket

Hear replies with a usage message holding the billed speech seconds, then settles the minutes on your balance.

Capacity is full

Admission checks capacity and tenant limits before audio is accepted, so a refused connection never reaches billing.

The live Hear protocol does not provide speaker labels. For recorded calls, use a Hear async job with diarization or channel separation to identify speakers.

Who uses real time transcription software, and for what?

Real time transcription software helps when a person or a program needs the words before the speaker stops. Developers and creator teams on PyAI use it most often for four jobs.

Support desks

Agent assist

A helpdesk tool watches each final event and surfaces the matching refund, ticket or policy for a human rep while the customer is still describing the issue.

Streaming and meeting apps

Captions

A webinar player paints settled text in full contrast and draft text faded, so viewers read along within a fraction of a second and the line rarely jumps.

Voice bot builders

Speech input

A custom-built bot treats the end-of-sentence event as its cue to hand text to a language model. Teams preferring one managed loop run PyAI Omni on the same engine.

Writers and editors

Hands-free drafting

A notes app streams a laptop mic at 16 kHz and drops each final sentence into the draft, so nobody has to clean up half-recognized words.

Real time transcription questions people ask

How accurate is live speech to text on phone calls?

PyAI Hear measured a 4.76% word error rate on 8 kHz phone recordings with accents and background noise, and 1.59% on clean speech. Major speech models land between 2 and 5% on clean English, so Hear performs at parity with them on live audio, not ahead of them.

Can ChatGPT do live transcription?

ChatGPT is a chat assistant rather than a captioning tool, though OpenAI documents a realtime transcription WebSocket in its developer API, checked October 2026. PyAI Hear provides a comparable live socket for your own product, returning partials in about 200 ms for $0.001 per minute of speech.

Is chunking audio every few seconds the same as streaming?

Chunking short files every few seconds is slower than true streaming, because each word waits for its chunk to finish. PyAI Hear accepts audio slices of about 20 ms over a WebSocket and returns draft text continuously, with the earliest partial about 200 ms after speech begins.

How much does real time transcription cost?

PyAI Hear streaming costs $0.001 per minute of active speech, and silent stretches on a connected socket are free. PyAI charges no platform fee. New accounts get $50 of credit with no credit card, and sandbox keys let developers test the live socket without spending anything at all.

PyAI is voice infrastructure. It hands you the audio and transcript, and does not replace your own models or review tools.

Open a socket, speak one line, and watch drafts settle into final text.

Start building with $50 creditSandbox keys never bill, and you can start without a credit card.

One stack: Omni, Hear, Speak, Recap and Trace.