Vidu S1 API

Vidu S1 Live API Integration

Vidu S1 is a streaming, real-time, two-way interactive digital-human video service. You create a "live session", the media flows over an Aliyun RTC channel (not HTTP), and a persistent WebSocket carries control signaling (start / hangup). This skill encodes the end-to-end flow and — more importantly — the field-tested pitfalls, several of which contradict the public doc. When the doc and this skill disagree, this skill wins (it reflects vendor clarifications).

Before you write any code

Establish these with the user. They are prerequisites, not code:

1. API Key from https://platform.vidu.cn/api-keys. Users may store it as either vda_xxx or Token vda_xxx, but every Vidu HTTP/WS request must send Authorization: Token vda_xxx. Placeholder values like vda_your_api_key_here are not configured keys.
2. Avatar image — a single-person image, as a public URL or base64 data URI.
3. Persona text — a free-form description of who the digital human is.
4. Aliyun RTC SDK — must be downloaded and integrated separately; media does NOT come over HTTP. SDK: https://help.aliyun.com/zh/live/introduction-to-real-time-audio-and-video-artc-sdk
5. Host domain — api.vidu.cn (China) or api.vidu.com (overseas). The doc writes {host} as a placeholder; always substitute a real domain.

The 6-step flow

1. POST /live/v1/lives          → create session, get live_id + RTC credentials
2. Aliyun RTC joinChannel        → publish your mic (+cam for video), subscribe to the bot
3. WebSocket connect + conn_init → announce "I'm ready"
4. Wait for conn_init_ack        → success:true means the bot is live (billing starts here)
5. Interaction                   → keep the heartbeat alive, handle forced hangup
6. Hangup + leaveChannel         → send hangup, close WS, leave RTC, optionally query the bill

Read references/api-reference.md for exact request/response schemas, error codes, and pricing. Read references/websocket-protocol.md for the WS message envelope, message types, heartbeat rules, and every hangup reason. Read references/voices.md for the system voice list.

Step 1 — Create the session

POST https://{host}/live/v1/lives with the avatar block. You get back a live object (save live.id — every later step needs it) and an rtc object (credentials for step 2). The session starts in status: waiting.

Step 2 — Join the Aliyun RTC channel

Use rtc.token and rtc.user_id to joinChannel. Some Aliyun RTC Web SDK versions accept the single token form, while others require an auth-info object such as { appId, channelId, userId, token, timestamp } built from the rtc response. If join fails, check the SDK version's expected joinChannel signature before changing Vidu credentials. Then:

• Publish your microphone — required, or the digital human can't hear the user.
• Publish your camera — only in video mode.
• Subscribe to the digital-human media stream — usually live-bot-{creatorID}-{liveID}, but treat live-video-push-{creatorID}-{liveID} as a valid media source too.

Role → RTC user-id format:

Step 3–4 — WebSocket + conn_init handshake

Connect to wss://{host}/live/ws/live/connect?live_id={live_id} with the Authorization: Token vda_xxx header. Immediately send a conn_init message, then wait for conn_init_ack:

• success: true → the bot is live. Billing starts now.
• success: false, error_code: NOT_READY → expected in video mode. Close the WS, wait, reconnect, retry with exponential backoff (2s → 4s → 8s). This is not an error. Ignore close/message events from the socket you intentionally closed for this retry.
• success: false, error_code: LIVE_CONN_INIT_FAILED → fatal. Go back to step 1 and create a new session.

Step 5 — Stay alive and handle hangups

The server pings every 5s; your client must send any message within 15s or it's dropped as a dead connection. Most WS libraries auto-reply to pings — confirm yours is enabled on the actual upstream Vidu connection. In a browser → backend proxy → Vidu design, this means the backend WS client must reply to Vidu pings. Listen for a type:6 hangup message, and add a fallback listener on the WS's abnormal-close event, because the server can drop you without a clean message.

Step 6 — End the call

Send a hangup message over the WS, close the WS, then call the RTC SDK's leaveChannel(). Optionally GET /live/v1/lives/{live_id} to read the bill (billed_seconds, credits_cost).

Pitfalls (read before implementing — these override the doc)

These are field-tested integration gotchas.

1. Do not assume a fixed live_duration. It is returned by the create API and may vary by the current service policy (real responses have returned 7200). Treat the returned positive integer as authoritative and design the UX around that session-specific cap.

2. Use rtc.token_expire_at; do not assume a fixed token lifetime. Treat the token as valid for its live session only. Don't cache/reuse it across sessions — get fresh credentials from step 1 each time. If it expires, create a new session.

3. WebSocket auth uses the Authorization: Token vda_xxx header — query params are NOT supported. Browsers cannot set custom Authorization headers on native WebSocket connections, and the header carries your raw Vidu token, so proxy the WebSocket through your own server. Same applies to HTTP calls: never ship vda_xxx to untrusted clients.

4. video mode WILL return NOT_READY on first connect — always. This is the SIP side warming up, not a failure. You must implement the disconnect → backoff → reconnect loop, or video mode will appear broken.

5. Aliyun RTC is a separate SDK integration, not an HTTP call. Budget for downloading and wiring up the ARTC SDK. Forgetting this is the most common "why can't I see/hear the avatar" bug.

6. Do not combine SDK auto-subscribe with repeated manual subscribe. Disable default remote auto-subscribe when manually selecting the bot/media user, and dedupe in-flight/completed subscribe and view-attachment work. Multiple RTC events can announce the same media user and otherwise look like duplicate streams or lag.

7. Guard against stale RTC and WS events. A previous RTC session or upstream WS can still emit late events after a retry, hangup, or new session. Ignore events unless they belong to the currently active RTC session/socket, or they can corrupt the new session's status, subscription state, or video view binding.

8. Billing starts at on_live (the moment conn_init_ack returns success:true), not at session creation. Duration rounds up to even seconds (11s bills as 12s). The account needs a minimum ~45 credits (enough for 30s) to start a session; mid-session it keeps deducting and auto-disconnects when credits hit zero.

9. Audio and audio+video modes are priced identically right now — mode choice is a product decision, not a cost lever.

10. The GET session response top-level shape is stable: { "live": { ... } }. Parse live.id, live.status, live.billed_seconds off that wrapper — don't expect the fields at the root.

11. GET /live/v1/voices returns ONLY custom cloned voices, by design. The system voice list is separate — see references/voices.md. Don't treat an "empty" voices response as an error; it just means no custom clones yet.

Voice selection

Default voice is Tina. Pass avatar.voice to pick another. System voices and their supported languages are in references/voices.md. To clone a voice, POST /live/v1/voices/clone (see references/api-reference.md); first 10 clones per user are free, then 899 credits each.