Does Deepgram Work in China? PIPL Cross-Border, Voiceprint Biometrics & Data Residency

For a team building voice features for mainland China with Deepgram, the question was never whether its API resolves from the mainland — it generally does. What decides it is what happens to the audio you send it. Deepgram is a speech-AI platform: its Nova speech-to-text models turn recorded audio into text, while its Aura text-to-speech models and Voice Agent API generate synthesized speech. Either way, a recording of a person speaking is that person’s voiceprint — a biometric identifier PIPL Article 28 treats as sensitive, and one you cannot anonymize — plus whatever the spoken words reveal. Deepgram’s hosted API processes that audio on its own offshore servers, with Global, EU, Australia and India endpoints but no mainland-China region. So each call is a cross-border transfer of sensitive biometric personal information (PIPL Articles 38–40, 数据出境; Article 28 adds a separate consent and a prior impact assessment), with a residency duty for a CIIO or high-volume handler and — for a China-facing synthesized-voice feature — a deep-synthesis and generative-AI filing door on top. The honest exception: Deepgram can be self-hosted in your own environment.

Deepgram in China at a glance

What you actually send — your speakers’ voiceprints and words

Deepgram is a raw speech-AI engine, not a content archive about you: audio goes in through its API, and text or synthesized speech comes back. On the speech-to-text side (Nova), you send a recording — a sales or support call, a voice message, a dictated clinical or legal note — and it returns a transcript, often with diarization that separates and labels each speaker. That recording is two things at once: the spoken content, which may itself be health, financial, or identity information, and the speaker’s voiceprint, a biometric identifier unique to that person. You can redact a transcript; you cannot anonymize the audio, because the audio is the identifier. On the text-to-speech side (Aura, Aura-2) and in the Voice Agent API, you send text and a chosen voice and get synthesized audio back — a generative feature that carries its own China duties when offered to the public.

Where does the audio go? Deepgram’s hosted API runs on its own servers, and its regional-endpoints documentation names the geographies: a default Global endpoint, plus EU (api.eu.deepgram.com), Australia (api.au.deepgram.com), and India (api.in.deepgram.com), each of which Deepgram says keeps requests “never routed outside” that region. None of those regions is in mainland China, and there is no China endpoint to select. For a caller or employee in China, each request is their voice leaving the country in real time.

On retention and training, Deepgram runs a Model Improvement Partnership Program: you can keep any request out of it by adding mip_opt_out=true, and Deepgram states that “Data from opted-out requests is retained only for the duration necessary to process the request.” A single-tenant Dedicated tier adds regional data residency. These are real data-protection controls — but they govern retention and training, not the border crossing itself, and none of the hosted regions sits in the mainland.

It’s a cross-border transfer of sensitive biometric data — under PIPL

Sending a China-origin recording to Deepgram’s servers in the US, the EU, Australia, or India is a cross-border transfer of personal information under PIPL Articles 38–40 (数据出境). As the handler, you — not Deepgram, the processor — must give notice, obtain a separate consent for the export, and put one transfer mechanism in place: a CAC security assessment, the CAC standard contract, or certification.

What sharpens this case is that a voiceprint is sensitive personal information. PIPL Article 28 treats biometric data as a sensitive category, so processing or exporting it requires a specific purpose, strict necessity, a separate specific consent, and a prior personal-information protection impact assessment (PIPIA) before the audio moves. The spoken content can pile on further sensitive categories — a diagnosis in a telehealth call, a card number read aloud to support, an ID recited to verify identity. And because the audio is the identifier, you cannot strip the sensitive part out and keep the rest: the voiceprint travels whenever the recording does.

Two more duties attach at the edges. Where you are a critical information infrastructure operator, or your volumes cross the regulators’ thresholds, mainland personal information must be stored in the mainland (PIPL Article 40; Cybersecurity Law Article 39 (formerly Article 37)), and a CAC-led data-export security assessment can be required before any of it lawfully leaves. And where you build a China-facing feature that synthesizes a voice with Aura or the Voice Agent API, China’s deep-synthesis rules and the generative-AI measures add a service-filing and content-labeling duty on that feature — a second door on top of the transfer.

Reaching the API isn’t the question — keeping the audio in-country is

The reflex for a China problem is to make a slow or failing third-party call respond faster. On this axis that reflex is backwards: making Deepgram’s offshore API more reachable for mainland users does not reduce the exposure — it simply exports more voiceprints, more smoothly, without supplying the lawful basis the transfer needs.

What actually works is to keep China-origin audio’s processing in-country, and here Deepgram offers a genuine lever that a pure cloud service does not: its models can be self-hosted. Deepgram states its voice AI is “available to self-host on your own infrastructure, both in the cloud or on-premises,” delivered as containers (Docker, Podman, Kubernetes) with “the same API and feature set as our hosted API,” and it describes the deployment plainly — “Your audio never leaves your environment.” Run that self-hosted stack on in-country infrastructure and the recording, the transcript, and the voiceprint are processed inside the mainland rather than shipped abroad. (Self-hosting is an enterprise-contract option; confirm the current terms and which models it covers with Deepgram.) Where self-hosting does not fit, route China audio through a licensed in-country speech service whose data stays in the mainland. Either way, obtain the Article 28 separate consent, run the PIPIA, minimize what you capture and keep, and — for any synthesized-voice feature you expose in China — meet the deep-synthesis and generative-AI filing and labeling duties. None of this is a tunnel that ships the audio offshore while hiding that it does.

This is a risk map, not a verdict. Whether you owe a separate consent, a transfer mechanism, a data-export assessment, in-country storage, a generative-AI filing, or an ICP filing — and in what combination — turns on your entity, your data volumes, whose voices you capture, and exactly what you send to transcribe or synthesize. Settle the specifics with counsel before a single mainland recording leaves the country.

The lawful path — map, localize, deliver

You do not have to drop Deepgram for your other markets to handle China voice lawfully. 21YunBox is a compliant overlay, not a migration — and, for a service you already use, a partner alongside your stack rather than a competitor to the speech vendor. There are three moves, and they fit together.

Map. Our China compliance team inventories what audio flows to Deepgram — which calls, recordings, and voice features; whose voiceprints and what spoken content they carry; where Deepgram processes and stores the audio; whether it is retained or used to train models; whether any feature synthesizes a voice; and where you lack a lawful basis for the cross-border leg (a separate Article 28 consent and a PIPIA, a transfer mechanism, a deep-synthesis/generative-AI filing). The exposure is written down before anything changes.

Localize. Because the risk is China-origin audio leaving the country, localize here means keeping that audio’s processing in-country: self-host Deepgram’s speech models inside China, or route China voice through a licensed in-country speech service whose data stays in the mainland; obtain the Article 28 separate consent; disable training retention; and, for a China-facing synthesized voice, meet the deep-synthesis and generative-AI filing and labeling duties. It never means a tunnel that makes the offshore call anyway. 21YunBox never uses or suggests circumvention of any kind.

Deliver. The China-facing site or app that captures or plays the audio is a public internet service with an ICP filing duty and needs compliant, in-country delivery — the 21YunBox Optimizer provides it, ICP-filed and in-country, in front of the stack you already run. No rebuild, no second codebase.

The goal is plain: your voice features run legally and compliantly for your users in China.

Get a compliance assessment →

Related reading:

• Cross-border data transfers under PIPL
• China’s Cybersecurity Law and data localization (Article 39)
• China’s data-export security assessment measures
• China’s generative-AI measures (生成式人工智能服务管理暂行办法)
• How to get an ICP filing for China

RELATED TECHNOLOGIES

Amazon Polly & Transcribe

Amazon Polly & Transcribe

AssemblyAI

AssemblyAI

ElevenLabs

ElevenLabs

Google Cloud Speech-to-Text

Google Cloud Speech-to-Text

Speechmatics

Speechmatics

OpenAI Whisper

OpenAI Whisper