GPT-4o Transcribe (Azure)

Microsoft Foundry announced GPT-4o-Transcribe, GPT-4o-Mini-Transcribe, and GPT-4o-Mini-TTS are available in Microsoft Foundry Models. The developer guide positions GPT-4o-Transcribe as the best-quality, fast speech-to-text option, with text and audio inputs, text output, and streaming support. The guide attributes the models’ performance to targeted audio pretraining, advanced distillation, and reinforcement learning. It recommends GPT-4o-Transcribe for accurate transcription in difficult environments such as call centers and automated meeting notes, and provides Azure OpenAI authentication and setup steps.

Microsoft announced the public preview of Azure OpenAI's advanced audio models, including GPT-4o-Transcribe, GPT-4o-Mini-Transcribe, and GPT-4o-Mini-TTS, available for deployment in East US2 on Azure AI Foundry. This marks the first availability of the exact gpt-4o-transcribe model on Azure infrastructure following its release by OpenAI. The post positions GPT-4o-Transcribe as the best-quality, fast speech-to-text option for demanding environments such as customer call centers and automated meeting notes, with streaming support and text-plus-audio input. Under the hood, the models rely on targeted audio pretraining, distillation from larger models, and reinforcement learning that improves transcription accuracy and reduces misrecognition.

Perle AI benchmarked seven commercial ASR models on 1,200 hand-verified code-switched utterances across Egyptian Arabic–English, Saudi Arabic–English, Persian–English, and German–English, and OpenAI's gpt-4o-transcribe was among those tested on all four language pairs. ElevenLabs Scribe v2 won every pair with a 13.2% mean WER and 0.936 mean BERTScore, beating the next-best model by 3.5x on Egyptian Arabic. The results qualify gpt-4o-transcribe's reputation: it performs decently on Arabic but struggles with Persian, while Google Chirp 3 collapses on hard samples. The authors argue word error rate alone misleads on Arabic and Persian because transliteration variance penalizes correct transcriptions, so they pair WER with BERTScore as the primary semantic metric.

OpenAI's official model documentation for gpt-4o-transcribe describes it as a speech-to-text model powered by GPT-4o that improves on the original Whisper models for word error rate, language recognition, and accuracy. The page lists a 16,000-token context window, 2,000 maximum output tokens, a June 1, 2024 knowledge cutoff, and audio input with text output only. Pricing on that page is set at $2.50 per 1M input audio tokens and $10.00 per 1M output tokens, with the cheaper GPT-4o Mini Transcribe listed at $1.25 input for comparison. Notably, OpenAI's documentation now labels gpt-4o-transcribe as deprecated even though Azure's catalog lists the same model as generally available, a discrepancy worth checking before new deployments.

The Microsoft Foundry catalog lists gpt-4o-transcribe as a generally available speech-to-text model hosted on Azure, version 2025-03-20, with OpenAI named as both publisher and author. It accepts text and audio input and returns text, with a 16,000-token context window and a 2,000-token maximum output for longer transcripts. The catalog entry states the model delivers significant improvements in word error rate and language recognition, surpassing earlier systems such as Whisper, positioning it for precision transcription work. It is offered as a Direct from Azure model, meaning unified billing, governance, and PTU portability across models hosted on Azure.