← Audio

AssemblyAIAssemblyAI

AssemblyAI

API平台API platform 转写+理解Transcription + understanding LeMURLeMUR 开发者向Developer-focused 英文向English-focused partial lifecycle: active 核验Checked 2026-10-22

AssemblyAI 是「转写+理解」一体化的语音 API：转写之上做摘要/实体/问答（LeMUR 线），文档化体验是它的开发者口碑；与 Deepgram 的实时性能叙事错位竞争。

AssemblyAI is the 'transcription + understanding' speech API: summarization/entities/Q&A on top of transcripts (the LeMUR line); documentation experience is its developer reputation — offset competition with Deepgram's real-time performance narrative.

适合与不适合When to use it, when not to

适合：转写后理解管线（摘要/抽取/问答）、英文向、文档体验优先团队。不适合：非开发者直接使用、隐私敏感音频、中文向主打。

Fits: post-transcript understanding pipelines (summarize/extract/Q&A), English-first, docs-experience-first teams. Not a fit: non-developers, privacy-sensitive audio (cloud), 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　AssemblyAI（美国旧金山/丹佛双线叙事）。AssemblyAI (San Francisco / Denver dual-line narrative, US).
国内使用Direct access: 大陆可直连direct OK　assemblyai.com 大陆直连实测 200（2026-10-22）。assemblyai.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: 自研统一语音理解模型线（best-seller 叙事：一个模型转写+理解，官方站点）；托管 API 形态，开发者不选裸模型权重。
A self-developed unified speech-understanding model line (best-seller narrative: one model transcribes + understands, official site); hosted-API form, developers do not pick bare model weights.
运行位置Where it runs: API/SDK 为主入口（REST + WebSocket + SDK，官方文档面）；预录文件转写是传统强项，实时流端点并存（官方文档）。
API/SDK as the main entry (REST + WebSocket + SDKs, official docs surface); pre-recorded file transcription is the traditional strength, with real-time streaming endpoints alongside (official docs).
本地文件Local files: 云端 API——音频经 API 上传；数据留存策略按条款（未逐条核验）；官方有「不用于训练」类承诺叙事（站点所见），条款级细节未逐条核验。
Cloud API — audio uploaded via the API; data-retention policy per the terms (not verified item by item); the official site carries a "not used for training"-type promise narrative (seen on the site), terms-level details not verified item by item.
关机后的任务Tasks after shutdown: 批处理（文件→结果轮询/webhook）+ 实时流双形态；无「关机续跑」概念（API 服务）。
Batch (file → result polling/webhook) + real-time streaming, two forms; no "keep-running-after-shutdown" concept (API service).
工具与扩展Tools & extensions: 转写+说话人分离+摘要+主题/实体/情感等理解层参数（官方文档列面全）； LeMUR 线（转写后 LLM 问答/抽取，官方文档叙事）——「转写+理解」一体化是它的差异叙事。
Transcription + speaker diarization + summarization + topics/entities/sentiment and other understanding-layer parameters (fully listed in official docs); the LeMUR line (post-transcription LLM Q&A/extraction, official docs narrative) — "transcription + understanding" in one is its differentiating narrative.
上下文与记忆Context & memory: API 无记忆概念；LeMUR 对既有转写做问答是「转写产物上的 LLM」，不是产品记忆。
APIs have no memory concept; LeMUR Q&A over existing transcripts is "an LLM over transcript artifacts", not product memory.
权限与限制Permissions & limits: API key+用量计费（D9：价格不采）；企业合规叙事（SOC2/HIPAA 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 + usage billing (D9: pricing not collected); enterprise-compliance narrative (SOC2/HIPAA-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「转写+理解」管线（摘要/抽取/问答一步到位）、英文向高准确率、文档化体验优先的团队。不适合：非开发者直接使用、隐私敏感音频（云端）、中文向主打。
Fits: developers building "transcription + understanding" pipelines (summarization/extraction/Q&A in one step), English-leaning high accuracy, teams that prioritize documented developer experience. Does not fit: non-developers using it directly, privacy-sensitive audio (cloud), Chinese-first usage.

头号误解Top misconception

与 Deepgram 同是 API 平台不是终端产品——差异在 AssemblyAI 把「理解层」（摘要/实体/问答）做进 APILike Deepgram it is an API platform, not an end product — the difference: AssemblyAI builds the understanding layer (summary/entities/Q&A) into the API
「不用于训练」类承诺要落到条款与 DPA 才算数——站点叙事≠条款原文（未逐条核验）'Not used for training' promises only count once they land in terms and the DPA — site narrative ≠ terms text (not verified line by line)
转写质量按语言/口音波动——官方基准英文向，中文场景要自测Quality varies by language/accent — official benchmarks are English-first; test Chinese yourself

未知项清单Known unknowns

数据留存与训练条款细节（未逐条核验）
中文模型档位状态（未逐条核验）
企业合规认证清单细节（未逐条核验）

Data-retention and training-clause details (not verified item by item)
Chinese model-tier status (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	AssemblyAI · 官网（产品与能力叙事）AssemblyAI · official site (product and capability narrative)	https://www.assemblyai.com
official	AssemblyAI · Terms（官方条款，2026-10-22 直抓核验）AssemblyAI · Terms (official terms, direct-fetch verified 2026-10-22)	https://www.assemblyai.com/terms
changelog	AssemblyAI · Changelog（文档更新线）AssemblyAI · Changelog (docs update line)	https://www.assemblyai.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）：

LeMUR 问答对长会议转写的召回质量
说话人分离在中文会议的实际表现
「不用于训练」承诺的条款落点（DPA 原文）

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):

LeMUR Q&A recall quality over long-meeting transcripts
Speaker diarization's actual behavior in Chinese meetings
The terms-level landing of the "not used for training" promise (DPA original text)