Project Management
Create and configure simultaneous interpretation projects — set up language pairs, prepare terminology, and tune translated speech speed.
This guide covers the complete workflow for simultaneous interpretation. If you haven't explored the feature yet, start with the Simultaneous Interpretation overview.
Create a project
Open the interpretation workbench, click "Create Project", and fill in:
|
Setting |
Description |
|
Project name |
Display name, e.g. "Global product launch" |
|
Project type |
Live Subtitling (bidirectional) or Simultaneous Interpretation (unidirectional) — cannot be changed after creation |
|
Source language |
The speaker's language |
|
Target Language |
The target language for translation |
After creating the project, configure speech recognition terminology, translation terminology, and translated speech voices on the project detail page.
Terminology learning
On the project detail page, click "Start Analysis", upload meeting materials, then click "Start Analysis". The system automatically analyzes the materials and extracts speech recognition and translation terminology. Each project supports up to 100 material files; drafts work fine and can be replaced with final versions later.
What to upload:
-
Speech drafts or talking points
-
Meeting agenda and speaker list
-
Product documentation, technical whitepapers
-
Previous meeting minutes (if available)
What not to upload:
-
Generic background information (the system already has strong baseline language capability)
-
Documents unrelated to this meeting
After analysis completes, review the auto-extracted results on the project page, remove unnecessary entries, then manually supplement terminology. Two input methods are supported:
-
Single entry: Add terms one at a time.
-
Batch import: Supports reading from Excel, web tables, or Markdown tables.
Auto-save mechanism: Changes to terms are automatically saved after 30 seconds of inactivity; you can also click the "Save" button to manually save your changes at any time.
Term limits: ASR terms and translation terms each support up to 500 entries. Excessive terminology consumes context window capacity and can degrade interpretation quality. Select only core terms relevant to your meeting needs.
ASR terms
Tell the system which words are likely to be misheard. Focus on person names (keep in English or Pinyin), product names and technical abbreviations, and easily confused homophones.
|
Term |
Common misrecognitions |
|
RaaS |
razz, rouse, race |
|
Qwen |
queen, quinn, kwen |
Translation terms
Lock the standard translation for proper nouns across target languages to avoid ambiguity. Only add proper nouns, not common words.
|
Source term |
Target term |
Notes |
|
大模型 |
Large Language Model |
Avoid literal "big model" |
|
睿译宝 |
Real Translation Buddy |
Official product name |
|
通义千问 |
Qwen |
Alibaba Cloud model brand |
Translated speech speed tuning
Each target language can have its own translated speech voice and speed setting (0.75x–1.5x). The default speed is a preset value for the voice and may not suit the current scenario.
Test translated speech speed before using it in a live session. Use a real speech recording to create a test project and invite target-language speakers to evaluate different speeds. Excessively fast speed makes it difficult for listeners to follow; excessively slow speed causes translated speech to lag behind the speaker with cumulative delay.
When testing by playing back a video, ensure proper sound isolation — for example, use a microphone for audio capture, or route system audio directly to the interpretation input via a virtual speaker. Playing audio through speakers without isolation causes output to feed back into the input, significantly degrading speech recognition accuracy.
Audio device setup
Microphone and speaker assignment
Many-to-many interpretation requires dedicated input and output devices for each language:
-
Select a microphone per source language: Each source language is bound to a dedicated microphone. Audio from that microphone is sent only to language pairs where it serves as the source.
-
Select a speaker per target language: Each target language is bound to a dedicated speaker for playing back translated audio in that language.
Pre-session audio check
Before starting a session, perform an audio device check:
-
Real-time monitoring: Use the built-in monitoring feature to verify that microphones capture audio clearly and speakers play back correctly.
-
Adjust device placement: Keep a reasonable distance between microphones and speakers to avoid echo. In small rooms or when using external speakers, consider using headphones for output.
Starting a session without an audio check may result in recognition failures or echo interference. Run a full device check before every formal session.
If the microphone detects audio input but no captions appear on the translation interface, check whether the browser has permission to use the microphone. Click the lock or permission icon on the left side of the browser address bar and make sure microphone access is enabled.
Handling translation latency
If translation latency gradually increases during a session, it may be caused by excessive context accumulation. In this case, you can refresh the page at an appropriate time (such as during a pause between speakers or an intermission). Refreshing clears the screen and resets the context, restoring normal translation latency.