New Transcription Models Separate Batch Processing From Live Speech

gpt-transcribe and gpt-live-transcribe address a basic product constraint: whether the audio is already complete or is still arriving.

gpt-transcribe takes a completed file and returns the full transcript. It is intended for call archives, podcasts, and larger batch jobs, where an application can wait for the recording to finish and optimize around complete-file processing, accuracy, and throughput.

gpt-live-transcribe keeps an open connection and returns text as audio arrives. That makes it suited to captions, dictation, and voice interfaces, where latency meaningfully affects the user experience. A realtime interface shown with the model is labeled “Listening live” as it adds the speaker’s words to the transcript.

The distinction is a deployment decision rather than a general ranking between the models. One produces a transcript after a recording ends; the other supplies text during the interaction itself.

Both models support 57 languages and are presented as stronger on cases that commonly cause transcription errors: accents, multilingual speech, short answers, names, and numbers.

Custom vocabulary is a developer-supplied control. Developers can provide a list of words, proper nouns, or code terms that are especially important to get right. In the demonstration, the prompt identifies transcription models as the topic and includes “phishing,” “ARR,” and “A1C”—terms from cybersecurity, sales, and healthcare that can be easy to miss. The vocabulary gives an application a way to call attention to terminology it needs represented accurately in the resulting transcript.

That is relevant to the uses named for a completed transcript: search, summaries, action items, or a handoff to another application. The demonstration frames vocabulary prompting as a way to make domain-specific terms part of the transcription request rather than leaving them solely to general speech recognition.

Background-noise handling is presented as default model behavior rather than a prompting control. A side conversation or ambient music is said to be less likely to appear as speech. In the live interface, the spoken explanation continues to appear in the transcript while ukulele music plays in the background.

The practical settings named are a loud café, a busy conference hall, and a recording made near a chatty coworker. The claim is not that surrounding sound disappears, but that the transcript can remain focused on the speaker.

Language hints are another developer-supplied control: they can be provided to improve accuracy. But the models are also said to follow multilingual speech automatically by default. In the live demonstration, the transcript moves from English into Spanish—“Y también puedo cambiar al español en la misma sesión”—and then back to English without ending the session. The displayed transcript follows both switches in real time.

Custom vocabulary and language hints give developers ways to guide the transcription task. Reduced pickup of ambient sound and automatic tracking of language changes are presented as model behavior intended to preserve the relevant speech under less controlled recording conditions.

For gpt-transcribe, a meeting recording described as about half an hour long is submitted for offline transcription. The interface shows a selected 25-minute audio file, “AI Engineer World’s Fair recap.m4a,” while processing is in progress. The presenter says the model takes a little less than a minute to process it.

Once processing is complete, the transcript is positioned as material for the next step rather than as the final product. The uses named are search, summaries, action items, and handoff to another application.

That framing clarifies the batch model’s unit of work: a complete recording is processed, then its text can be used elsewhere. It fits workflows built around archives, podcasts, recorded calls, and larger batches—not an interaction in which a user needs to see the current utterance immediately.