面向语音开发者的更新

AI 音频功能解锁了令人兴奋的全新用户体验领域。今年早些时候，我们发布了多个新的音频模型，包括 gpt-realtime，以及新的 API 功能，以帮助开发者构建这些体验。

上周，我们发布了新的音频模型快照，旨在通过提高生产语音工作流的可靠性和质量——从转录和文本转语音到实时、原生的语音对语音代理——来解决构建可靠音频代理时的一些常见挑战。

这些更新包括：

gpt-4o-mini-transcribe-2025-12-15，用于通过 Transcription 或 Realtime API 进行语音转文本
gpt-4o-mini-tts-2025-12-15，用于通过 Speech API 进行文本转语音
gpt-realtime-mini-2025-12-15，用于通过 Realtime API 进行原生实时语音对语音
gpt-audio-mini-2025-12-15，用于通过 Chat Completions API 进行原生语音对语音

新的快照包含几项共同的改进：

音频输入方面：

针对真实场景和嘈杂音频的更低词错误率
静音或有背景噪音时更少的幻觉

音频输出方面：

更自然、更稳定的语音输出，包括使用 Custom Voices 时

定价与之前的模型快照保持不变，因此我们建议切换到这些新快照，以相同的价格获得更好的性能。

如果你正在构建语音代理、客户支持系统或品牌语音体验，这些更新将帮助你让生产部署更加可靠。下面，我们将详细介绍新增内容以及这些改进如何体现在真实的语音工作流中。

语音对语音

我们正在部署新的 Realtime mini 和 Audio mini 模型，这些模型经过优化，以实现更好的工具调用和指令遵循能力。这些模型缩小了 mini 模型与全尺寸模型之间的智能差距，使一些应用能够通过迁移到 mini 模型来优化成本。

`gpt-realtime-mini-2025-12-15`

gpt-realtime-mini 模型旨在与 Realtime API 搭配使用，这是我们用于低延迟、原生多模态交互的 API。它支持流式音频输入和输出、处理打断（可选语音活动检测）以及在模型继续说话的同时在后台进行函数调用等功能。

新的 Realtime mini 快照更适合实时代理，在指令遵循和工具调用方面有明显提升。在我们的内部语音对语音评估中，与之前的快照相比，指令遵循准确率提高了 18.6 个百分点，工具调用准确率提高了 12.9 个百分点，并且在 Big Bench Audio 基准测试中也有所提升。

Speech-to-speech eval chart 1 Speech-to-speech eval chart 2 Speech-to-speech eval chart 3

这些提升共同带来了更可靠的多步交互，以及在实时、低延迟环境中更一致的函数执行。

对于代理准确率值得付出更高成本的场景，gpt-realtime 仍然是我们表现最好的模型。但当成本和延迟最重要时，gpt-realtime-mini 是一个很好的选择，在真实场景中表现出色。

例如，Genspark 对其进行了双语翻译和智能意图路由的压力测试，除了语音质量有所提升外，他们还发现延迟接近即时，同时在意图识别方面在快速交流中始终保持精准。

`gpt-audio-mini-2025-12-15`

The gpt-audio-mini model can be used with the Chat Completions API for speech-to-speech use cases where real-time interaction isn’t a requirement.

Both new snapshots also feature an upgraded decoder for more natural sounding voices, and better maintain voice consistency when used with Custom Voices.

Text-to-speech

Our latest text-to-speech model, gpt-4o-mini-tts-2025-12-15, delivers a significant jump in accuracy, with substantially lower word error rates across standard speech benchmarks compared to the previous generation. On Common Voice and FLEURS, we see roughly 35% lower WER, with consistent gains on Multilingual LibriSpeech as well.

Text-to-speech eval chart 1 Text-to-speech eval chart 2 Text-to-speech eval chart 3

Together, these results reflect improved pronunciation accuracy and robustness across a wide range of languages.

Similar to the new gpt-realtime-mini snapshot, this model sounds much more natural and performs better with Custom Voices.

Speech-to-text

The latest transcription model, gpt-4o-mini-transcribe-2025-12-15, shows strong gains in both accuracy and reliability. On standard ASR benchmarks like Common Voice and FLEURS (without language hints), it delivers lower word error rates than prior models. We’ve optimized this model for behavior on real-world conversational settings, such as short user utterances and noisy backgrounds. In an internal hallucination-with-noise evaluation, where we played clips of real-world background noise and audio with varying speaking intervals (including silence), the model produced ~90% fewer hallucinations compared to Whisper v2 and ~70% fewer compared to previous GPT-4o-transcribe models.

Transcription eval chart 1 Transcription eval chart 2 Transcription eval chart 3

This model snapshot is particularly strong in Chinese (Mandarin), Hindi, Bengali, Japanese, Indonesian, and Italian.

Custom Voices

Custom Voices enable organizations to connect with customers in their unique brand voice. Whether you’re building a customer support agent or a brand avatar, OpenAI’s custom voice technology makes it easy to create distinct, realistic voices.

Theese new speech-to-speech and text-to-speech models unlock improvements for custom voices such as more natural tones, increased faithfulness to the original sample, and improved accuracy across dialects.

To ensure safe use of this technology, Custom Voices are limited to eligible customers. Contact your account director or reach out to our sales team to learn more.

From prototype to production

Voice apps tend to fail in the same places, mainly on long conversations or with edge cases like silence, and tool-driven flows where the voice agent needs to be precise. These updates are focused on those failure modes—lower error rates, fewer hallucinations, more consistent tool use, better instruction following. And as a bonus, we’ve improved the stability of the output audio so your voice experiences can sound more natural.

If you’re shipping voice experiences today, we recommend moving to the new 2025-12-15 snapshots and re-running your key production test cases. Early testers have confirmed noticeable improvements without changing their instructions and simply switching to the new snapshots, but we recommend experimenting with your own use cases and adjusting your prompts as needed.