ALIYUN · REALTIME ASR

Alibaba Cloud Paraformer realtime transcription

See words while you speak, with sentences confirmed by the service after a pause. Reuse a saved Aliyun connection without entering the key again in recording settings.

Save the connection before recording

Save the Aliyun connection and recognition model in Setup, then select the service in Recording settings. Return to Setup when changing the model; do not enter the key again on other pages. First try a short recording with your own account and input device to check access, recognition quality and charges. Speed and quality depend on the service, connection and recording conditions.

Before you start

An enabled Alibaba Cloud Model Studio account, service access in Beijing, and the corresponding API key. Service usage may be billed under your account terms.
A working recording input and network connection. Audio is streamed to the selected Alibaba Cloud service. No local ASR model download is required for this route.

First use: configure in order

Open Setup → API configuration. Find Aliyun and enable the connection. Reuse an existing saved Aliyun connection if you already have one; a separate key entry is unnecessary.
Save the Beijing endpoint and API key. You can keep the default https://dashscope.aliyuncs.com/compatible-mode/v1 or use your actual Beijing workspace root https://<WorkspaceId>.cn-beijing.maas.aliyuncs.com. Replace the workspace ID with your own value. This integration currently accepts these Beijing service hosts; Singapore and third-party hosts are not interchangeable.
Set the ASR model to paraformer-realtime-v2. Finish the wizard and save. This field is for the realtime recognition model rather than a file-transcription, TTS or conversational-audio model.
Open Recording settings. Choose Realtime API as the microphone recognition method, then Aliyun · Paraformer as the saved service. The saved endpoint and key are reused. Once its ASR model is saved in Setup, it is shown read-only here; do not enter the model or key again on the recording page.
Check the language, then try a short recording. Use zh (Chinese), en (English), ja, yue, ko, de, fr or ru, or leave the hint empty for automatic recognition. An explicitly selected source language remains bound to this recording; hints do not guarantee recognition accuracy. Return to live subtitles, click Start, speak a sentence at normal volume and pause briefly.
After Stop, wait for completion. After capture stops, the app still waits for the remaining results and task-finished acknowledgement. Temporary text is not treated as a confirmed final transcript.

Recording settings selecting Realtime API, Aliyun Paraformer and paraformer-realtime-v2 — Actual production UI with controlled IPC configuration. The screenshot is copied without image edits and is not evidence of live cloud, billed-account or physical-microphone acceptance.

Edit the connection or model through Manage service

In the Realtime API area of Recording settings, click Manage service. It opens Setup directly at API configuration.
Find Aliyun and check the Beijing endpoint, API key and ASR model. Realtime recognition requires paraformer-realtime-v2. Finish the wizard and save your changes.
Return to Recording settings, choose Aliyun · Paraformer, and check the displayed model before starting a new recording. Saved changes apply to the next recording. An active recording does not switch models or services midway.

If “The saved model has changed” appears, refresh the settings or reselect the service. If a file-transcription, TTS or other unsupported model was saved, correct it through Manage service and save. The app asks you to correct the setting rather than silently substituting another model.

What the states mean

While speakingInterim text may be revised by later results.

Sentence confirmedThe server uses an approximately 400 ms pause threshold for segmentation; this is not a 400 ms caption-latency guarantee.

Stop and finishWait for task-finished. One final sentence does not mean the complete recording task has ended.

A confirmed final sentence received after Stop still reaches captions and the meeting record. It is also submitted for translation when translation is enabled. No further text is typed into other windows after Stop. Wait for this task to finish before starting another recording; unconfirmed drafts are not added as final text.

If something goes wrong

Aliyun · Paraformer is absent: Check that your version includes realtime transcription, available in r466 or later. The centralized model-management steps on this page describe the r468 candidate. Then confirm Aliyun is enabled, the model name is correct and the wizard was saved.
The model is read-only or unsupported: This is the ASR model saved in Setup. Choose Manage service, change the model to paraformer-realtime-v2, finish saving, then return and reselect the service. A file-transcription model cannot replace this realtime model.
Endpoint, key or service access fails: Check that the Beijing region, workspace and key belong together and that the account has access. Edit the key in Setup; do not duplicate it in other pages or send it to a third-party host.
Connection fails or the final sentence is missing: Confirmed results and the original recording are retained. Drafts are not promoted to final text. The app does not automatically reconnect or resend audio. Review the error and connection; if needed, manually transcribe the saved recording through the separate file-transcription workflow and review its service and possible charges.

Scope of this feature

This is realtime speech-to-text. File transcription, voice cloning and complete speech-to-speech interpreting are separate workflows. Translation, translated audio and Linglu reports still need their own capabilities and configuration.

Protocol references: Client events · Server events(2026-10-11).