Muse Voice Transcribe API 指南
面向正在评估 Meta Model API 的开发者：这里整理已核验的模型名、价格、实时能力和隐私边界，不编造 endpoint、鉴权字段或响应结构。
Last verified: September 5, 2026
Model
API price
Equivalent
主要来源
Meta 的 speech-to-text developer resource 当前位于 Meta Model API 文档中，并从 Muse Voice Transcribe 发布页链接进入。该开发者资源可能需要登录；本页只保留可核验的模型名、价格和能力边界，不补写未公开的 endpoint、headers 或 JSON schema。
先区分 API、Mac 和 Muse Code
三个入口都出现在 Meta 公共资料中，但只有 Meta Model API 是开发者应按 API 文档评估的集成路径。
开发者评估和生产集成应从 Meta Model API 的 speech-to-text 文档开始，模型名为 muse-voice-transcribe-1.0。
打开文档Meta 表示 Mac 听写由 Muse Voice Transcribe 提供支持；这是产品使用路径，不是 API endpoint 或计费说明。
查看 Meta AIMuse Code 中的语音输入同样被 Meta 列为访问路径。开发工作流细节应以当前 Muse Code 页面为准。
查看 Muse CodeAPI 评估时应核对的能力
Meta describes Muse Voice Transcribe as a real-time audio perception model for streaming ASR, so text can be produced while speech is still arriving.
The launch documentation describes native diarization for conversations with 20+ speakers, including long audio examples beyond one hour.
Endpointing is documented as part of the same model family, marking when speech begins and when a turn has finished.
The model supports multilingual speech, code-switching, and language, keyword, and context biasing for terms such as names, products, and places.
不猜 endpoint 的集成清单
获取音频权限并确认你有权录制正在捕获的声音。
选择官方支持的实时流或文件转录模式，不假设二者的重试、静音和计费细节完全相同。
把转录、说话人标签和 endpoint 信号作为模型输出处理，并保留人工校验路径。
上线前在 Meta 控制台确认账户权限、地区、速率限制、数据保留和生产条款。
哪些内容仍需看官方文档
Authentication
Assume authentication is account-scoped through Meta Model API until the official docs for your account state the exact token, header, and project setup. This guide intentionally avoids placeholder keys.
Endpoint and schema
This site does not publish guessed request paths, authentication headers, WebSocket formats, or response JSON. Use the current Meta Model API docs and cookbook for implementation details.
ZDR and data handling
Zero data retention, storage, regional processing, and enterprise controls should be confirmed in Meta's current API terms before sending production audio.
Cookbook examples
Treat official cookbook examples as the starting point for streaming behavior, file transcription behavior, error handling, and production retries.