← Audio

DeepgramDeepgram

Deepgram

API平台API platform 实时流式Real-time streaming 转写Transcription TTSTTS 开发者向Developer-focused partial lifecycle: active 核验Checked 2026-10-22

Deepgram 是开发者向的语音 API 平台：Nova 转写 + Aura TTS 双线，实时流式转写是主打；它是「造语音产品」的底座，不是终端转写产品。

Deepgram is the developer-facing speech API platform: Nova transcription + Aura TTS, real-time streaming is the lead; it's the base for building speech products, not an end-user transcription product.

适合与不适合When to use it, when not to

适合：实时语音应用（客服/会议/字幕流）、低延迟 API 集成、英文向高准确率。不适合：非开发者直接使用、隐私敏感音频、中文向主打。

Fits: real-time speech apps (CS/meetings/caption streams), low-latency API integration, English-first accuracy. Not a fit: non-developers using it directly (it's an API platform), privacy-sensitive audio, Chinese-first work.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-22

出品属地Origin: 国外出品foreign-made　Deepgram（美国旧金山）。Deepgram (San Francisco, US).
国内使用Direct access: 大陆可直连direct OK　deepgram.com 大陆直连实测 200（2026-10-22）。deepgram.com measured 200 direct from mainland (2026-10-22).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-22 检索）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-22); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 自研语音模型族（Nova 系转写模型、Aura 系 TTS 模型——官方文档 models 页列线， developers.deepgram.com/docs/models 直抓核验 2026-10-22）；开发者选模型档位不训练自有模型（托管 API 形态）。
A self-developed speech model family (Nova-line transcription models, Aura-line TTS models — listed on the official docs models page, developers.deepgram.com/docs/models direct-fetched and verified 2026-10-22); developers pick model tiers and do not train their own models (hosted-API form).
运行位置Where it runs: API/SDK 为主入口（REST + WebSocket + 各语言 SDK，官方文档面）；实时流式转写是主打 （WebSocket 流式端点官方文档化）；另有浏览器/预录文件端点。
API/SDK as the main entry (REST + WebSocket + per-language SDKs, official docs surface); real-time streaming transcription is the flagship (WebSocket streaming endpoint documented officially); browser/pre-recorded-file endpoints also exist.
本地文件Local files: 云端 API——音频经 API 上传处理；数据留存策略按官方条款（未逐条核验）。
Cloud API — audio uploaded and processed via the API; data-retention policy per the official terms (not verified item by item).
关机后的任务Tasks after shutdown: 实时流式是核心形态（边说边转）；批处理端点并存；无「关机续跑」概念（API 服务）。
Real-time streaming is the core form (transcribe while speaking); batch endpoints alongside; no "keep-running-after-shutdown" concept (API service).
工具与扩展Tools & extensions: 转写（多语言/关键词 boosts/标点/说话人分离参数面）、TTS（Aura 线）、语音理解参数面（官方文档所见）；开发者工具链（SDK/示例/延迟指标文档化）。
Transcription (multilingual/keyword boosts/punctuation/speaker-diarization parameter surfaces), TTS (Aura line), speech-understanding parameter surfaces (seen in official docs); developer toolchain (SDKs/examples/latency metrics documented).
上下文与记忆Context & memory: API 无记忆概念——会话上下文完全由调用方（集成方）自己管理，多轮修正靠应用层拼。
APIs have no memory concept — conversation context is managed entirely by the calling (integrating) side; multi-turn correction is stitched together at the application layer.
权限与限制Permissions & limits: API key 体系+用量计费（D9：价格不采）；企业面有合规叙事（SOC2 类认证线官网所见，细节未逐条核验）；数据留存/训练开关按条款（未逐条核验）。
API-key system + usage billing (D9: pricing not collected); an enterprise compliance narrative (SOC2-type certification lines seen on the site, details not verified item by item); data-retention/training switches per the terms (not verified item by item).
适合什么任务Best for: 适合：开发者做实时语音应用（客服/会议/字幕流）、低延迟转写 API 集成、英文向高准确率。不适合：非开发者直接使用（它是 API 平台不是终端产品）、隐私敏感音频（云端处理）、中文向主打。
Fits: developers building real-time voice apps (support/meetings/caption streams), low-latency transcription API integration, English-leaning high accuracy. Does not fit: non-developers using it directly (it is an API platform, not an end product), privacy-sensitive audio (cloud processing), Chinese-first usage.

头号误解Top misconception

Deepgram 是 API 平台不是终端产品——「用 Deepgram 转写会议」实际是「用 Deepgram 造一个转写产品」Deepgram is an API platform, not an end product — 'transcribe meetings with Deepgram' actually means 'build a transcription product on Deepgram'
实时流式与批处理是不同端点不同计费面——按单一「转写服务」理解会错配Real-time streaming and batch are different endpoints with different billing — reading it as one 'transcription service' mis-architects
转写质量按语言/口音/领域波动——官方基准是英文向叙事，中文场景要自测Quality varies by language/accent/domain — official benchmarks are English-first narratives; test Chinese yourself

未知项清单Known unknowns

数据留存与训练用途条款细节（未逐条核验）
中文模型档位的当前状态（未逐条核验）
企业合规认证清单细节（未逐条核验）

Data-retention and training-use clause details (not verified item by item)
Current status of Chinese model tiers (not verified item by item)
Enterprise-compliance certification list details (not verified item by item)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-22，内容更新日content updated on 2026-10-22。

类型Type	名称Name	链接Link
official	Deepgram · 官网（产品与能力叙事）Deepgram · official site (product and capability narrative)	https://www.deepgram.com
docs	Deepgram · 开发者文档 models 页（模型线，2026-10-22 直抓核验）Deepgram · developer docs models page (model line, direct-fetch verified 2026-10-22)	https://developers.deepgram.com/docs/models
changelog	Deepgram · 文档 Changelog（更新线）Deepgram · docs Changelog (update line)	https://developers.deepgram.com/changelog

实测记录Measured runs

本档未跑音频质量与延迟测试；结论只覆盖能力面与许可面。

实测建议（三条，都指向官方可读页没给的东西）：

WebSocket 流式的端到端延迟实测（官方延迟指标 vs 真实网络）
中文/中英混说转写准确率（官方基准是英文向叙事）
数据留存开关的实际生效面

This site has not completed measured runs. The coding-oriented test protocol is at protocols/coding/_protocol.md (the coding task suite does not apply to creation products).

Suggested measurements (three, all pointing at things the official readable pages do not give):

End-to-end latency of WebSocket streaming (official latency metrics vs real networks)
Chinese / Chinese-English mixed-speech transcription accuracy (the official benchmark is an English-leaning narrative)
Where the data-retention switch actually takes effect