Core Concepts

Task

A Task is the top-level concept in VAS, representing one complete speech-processing job. Each Task has a unique task_id (UUID) and, once completed, can be queried and managed through the REST API.

A Task can be created in the following ways:

• Live recording: Created automatically after voice translation over WebSocket
• Audio import: Created automatically after an uploaded audio file finishes processing

Recording

A Recording is the core data of a Task, containing the audio file and the transcript. Each Task corresponds to one Recording.

Recording Types

Type Differences

Since v1.7.0, record (plain recording) is a lightweight speech-recognition-only type: translation and TTS are not supported (sending translation_languages or tts_enabled=true returns 400), summary is off by default and only generated when summary_template or summary_mode=custom is provided, and speaker separation is optional.

Processing Status

A Recording goes through several processing statuses from creation to completion. You can query active tasks with GET /api/v1/tasks?status=active.

Live recording / broadcast: recording → uploading → processing → completed / failed
Audio import:               importing → uploading → processing → completed / failed

Session

A Session is an active working session within a single WebSocket connection, from start to stop.

Session Lifecycle

Connect WebSocket
    │
    ▼
  start  ──→  session_started (obtain session_id + task_id)
    │
    ▼
  Send audio (audio) ──→ Receive recognition results (result)
    │
    ├── pause / resume (optional)
    ├── config (optional, update terminology)
    ├── retranslate (optional, re-translate)
    ├── switch_language (optional, switch language)
    ├── set_tts (interpretation mode, toggle TTS on/off and settings)
    │
    ▼
  stop   ──→  status (stop confirmation)
    │
    ▼
  task_complete (audio and transcript upload complete, obtain task_id)

Speaker Recognition Modes

In multi-speaker mode, the system automatically recognizes different speakers and labels them as Guest-1, Guest-2, and so on. You can manage speakers with the following operations:

• Rename: Rename Guest-1 to the speaker's actual name
• Reassign: Change which speaker a particular sentence belongs to
• Merge: Merge two speakers into one

For details, see the Speaker Management Guide.

Channels and Multi-Channel Mode

A channel represents the independent audio source of one physical microphone. Multi-channel mode (recognition_mode: "multi_channel") lets a single recording take in up to 8 channels at the same time, with each channel running its own independent speech recognition. Speaker identity is determined directly by the channel: speech received on a channel is always attributed to the speaker configured for that channel (speaker_id format: channel_{N}), rather than inferred by AI.

When to choose it: Choose multi-channel mode when every speaker can have a dedicated directional or close-talking microphone (e.g., panel discussions, interpretation booths, multi-guest interviews), when you need fully deterministic speaker attribution, or when different speakers use different languages; when everyone shares a single microphone, use multi_speaker. Multi-channel mode requires activation before use and supports PCM audio only. During recording you can dynamically add or remove channels (add_channel / remove_channel) or change a single channel's language (set_channel_language).

For details, see the WebSocket voice-translation reference and the Speaker Management Guide.

Data Flow

Microphone audio
    │
    ▼
  Speech recognition (STT) ──→ Original text (origin)
    │
    ▼
  Translation    ──→ Translation result (translation)
    │
    ├── Real-time translation: translate immediately after each sentence is recognized
    └── Non-real-time translation: translate only after is_final
    │
    ▼
  TTS speech synthesis (optional) ──→ Audio (tts_ready)
    │
    ▼
  Meeting summary (optional) ──→ Summary result

Real-Time vs. Non-Real-Time Translation

Conversation mode (conversation) is not affected by this setting and always translates after the sentence is finalized; broadcast mode (broadcast) always uses real-time translation.

Interpretation Mode

Interpretation mode lets two people who speak different languages hold a real-time interpreted conversation over a single WebSocket connection. The system automatically detects the language of each sentence and maps the translation direction accordingly, and the entire process is completely transparent to users.

Person A speaks Chinese → auto-detect zh-TW → translate to en-US → TTS (en-US voice) → tts_ready
Person B speaks English → auto-detect en-US → translate to zh-TW → TTS (zh-TW voice) → tts_ready

Interpretation Mode Characteristics

Interpretation Flow Overview

1. WebSocket: Start with the conversation type and specify the two languages
2. Send audio: The system automatically detects the language, translates to the other language, and returns TTS
3. Automatic language detection: Each sentence is detected independently; origin.language reflects the detected language
4. Stop: Stop recording; a Task is created automatically once processing completes

For details, see the interpretation mode section of the Real-Time Voice Translation Guide.

Broadcast Architecture

The broadcast feature delivers one presenter's speech to many viewers in real time.

Presenter (WebSocket)
    │
    ├── Send audio ──→ STT ──→ Translation ──→ TTS (optional)
    │
    ▼
  VAS server
    │
    ▼
Viewer 1 (SSE) ──→ Live subtitles + TTS audio
Viewer 2 (SSE) ──→ Live subtitles + TTS audio
Viewer N (SSE) ──→ Live subtitles + TTS audio

Broadcast Phases

Broadcast Flow Overview

1. REST API: Create a broadcast channel and obtain a broadcast_token
2. WebSocket: The presenter starts recording with the broadcast type
3. SSE: Viewers connect through the share link to receive subtitles
4. WebSocket: The presenter stops recording, and the broadcast ends

For details, see the Broadcast Guide.

Summary Templates

A summary template defines the output format for automatic summaries. In transcribe recordings, summary_template is a required parameter; in conversation, broadcast, and record recordings, it is optional (record generates a summary only when summary_template or summary_mode=custom is provided, and does not generate one by default).

You can query the list of available templates via the Summary Templates API.

Terminology and Text Processing

VAS provides three text-processing settings that can be configured before or dynamically during recording:

For details, see the config action in the WebSocket voice-translation reference.

Webhook Callbacks

VAS supports Webhook notifications that proactively push events to a URL you specify when recording processing completes or fails, or when your credit balance runs low or is exhausted, eliminating the need for polling.

Supported Events

Configuration

• Per request: Add the callback_url parameter to the API request (audio import, broadcast)
• Per API Key: Specify webhook_url in the API Key settings (used when no callback_url is specified; real-time recordings and credit balance notifications are always sent here)

For details, see the Webhook Callback Guide.

Version: V1.24.1 Last Updated: 2026-10-07