gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-mini-tts

2025-03-20 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI launches new speech recognition models (gpt-4o-transcribe, gpt-4o-mini-transcribe) and a steerable speech synthesis model (gpt-4o-mini-tts) in the API to build voice agents.

What changes

  • The transcription models improve word error rate compared with Whisper, especially with accents, background noise and varying speech rate.
  • They are based on GPT-4o and GPT-4o mini, with audio pretraining, distillation and reinforcement learning.
  • gpt-4o-mini-tts makes it possible for the first time to tell the model how to speak (for example an empathetic customer service tone), with predefined synthetic voices.
  • Integration offered with the Agents SDK to assemble a voice agent.

OpenAI launches new speech recognition models (gpt-4o-transcribe, gpt-4o-mini-transcribe) and a steerable speech synthesis model (gpt-4o-mini-tts) in the API to build voice agents. ### What changes - The transcription models improve word error rate compared with Whisper, especially with accents, background noise and varying speech rate. - They are based on GPT-4o and GPT-4o mini, with audio pretraining, distillation and reinforcement learning. - gpt-4o-mini-tts makes it possible for the first time to tell the model how to speak (for example an empathetic customer service tone), with predefined synthetic voices. - Integration offered with the Agents SDK to assemble a voice agent. ### Specifications - gpt-4o-transcribe and gpt-4o-mini-transcribe: audio input, text output, 16,000-token context, 2,000-token output, cutoff June 2024 (according to the documentation). - gpt-4o-mini-tts: text input, audio output, 2,000 input tokens maximum. - Available to all developers in the audio API on the day of the announcement. ### Pricing Pricing per million tokens (USD), according to the current documentation (the announcement does not give figures). | Model | Input | Output | | --- | --- | --- | | gpt-4o-transcribe (audio) | 2.50 | 10 | | gpt-4o-mini-transcribe (audio) | 1.25 | 5 | | gpt-4o-mini-tts | 0.60 (text) | 12 (audio) | ### Benchmarks Scores published by the vendor: the announcement presents the word error rate on FLEURS (more than 100 languages) only as charts, with no numerical value in the text. It states that the models match or exceed the other frontier models for most languages. ### Key takeaways These building blocks make it possible to assemble a voice agent by chaining transcription, LLM and synthesis, as an alternative to the end-to-end realtime API. Updated snapshots were released on December 15, 2025.

Links

Tags : Audio, OpenAI, GPT-4o, Transcription, Synthèse vocale