OpenAI launches gpt-transcribe and gpt-live-transcribe

Anyone who has built a voice workflow knows the pain: a caller says "account number 8-4-7-2," and the transcription engine decides to write "eighty-four seventy-two." That tiny formatting quirk is enough to break an automated backend.

OpenAI is trying to clean up that mess. The company just released two dedicated speech-to-text models: gpt-transcribe for batch processing and gpt-live-transcribe for real-time audio. Both are now available in Microsoft Foundry.

Built for noisy, real-world speech

Instead of just chasing clean benchmark audio, both models focus on real-world edge cases. They are tuned to isolate voices against background noise, catch clipped interruptions, and properly parse alphanumeric sequences like ID numbers, codes, dates, and addresses.

They also handle domain-specific terminology, code-mixing across multiple languages, heavy accents, and low-volume whispered speech that usually gets skipped.

gpt-transcribe takes full audio files and outputs text, positioned by OpenAI as its highest-accuracy ASR model to date. It is aimed at meeting archives, media files, and compliance reviews.

gpt-live-transcribe runs on the Realtime API for streaming speech-to-text. Notably, it introduces "tunable latency," giving builders the slider to balance response speed against accuracy depending on the use case.

Microsoft also made gpt-realtime-2.1 and gpt-realtime-mini-2.1 available for developers who need speech-in, speech-out conversational agents.

Why it matters

For builders, this solves a very pragmatic headache. While fully multimodal speech-to-speech models get the headlines, most enterprise pipelines still rely on accurate text to trigger downstream APIs, run searches, or store database records. Tunable latency means developers building contact center bots do not have to compromise on accuracy when speed matters.

We think this release signals a healthy pivot toward enterprise reality. End-to-end voice agents sound impressive in demos, but businesses will not deploy automations that butcher credit card numbers or drop short replies. Going back to polish core speech recognition shows OpenAI understands where enterprise budgets are actually getting spent.