User guide

Content Source

Select one of the following content sources:

• Recorded file recognition: Supports audio files in MP3, WAV, M4A, WMA, AAC, OGG, AMR, FLAC, and AIFF formats, and video files in MP4, WMV, M4V, FLV, RMVB, DAT, MOV, MKV, WEBM, AVI, MPEG, 3GP, and OGG formats. The file size cannot exceed 6 GB.

• Real-time transcription: Captures and transcribes audio in real time.

• Tingwu task: Enter the task ID of a completed transcription task for the current application. This can reduce speech-to-text costs.

  NoteAfter a task is created, the system returns the TaskID field in the response.

• Meeting text content: Enter the dialogue text in the specified format. This can also reduce speech-to-text costs.

Example 1:
Speaker 1: Dialogue content.
Speaker 2: Dialogue content.
Example 2:
Sales: Dialogue content.
Customer: Dialogue content.

Audio type

This setting applies only when the content source is Recorded file recognition:

• Phone recording or 16 kHz and higher single-channel audio
• 16 kHz and higher multi-channel audio (recorded by in-vehicle devices)

Transcription model

This setting applies only when the content source is Recorded file recognition or Real-time transcription:

• paraformer-v2 (Chinese and English)
• paraformer-v2 (Chinese, English, Japanese, Korean, and Cantonese)
• paraformer-v2 (English)
• paraformer-v2 (Cantonese)
• paraformer-v2 (Japanese)
• paraformer-v2 (Korean)
• Education domain model (Chinese and English)
• Automotive domain model (Chinese, English, and Cantonese)

Transcription Configuration

These settings apply only when the content source is Recorded file recognition or Real-time transcription:

Speaker separation: Supports separation for two or more speakers.

Recognition language: The available language options depend on the selected transcription model.

Translation: Supports translating the original recognized text. Supported languages include Chinese, English, Japanese, Korean, German, French, and Russian.

Hotwords: Supports adding and managing hotword groups. Using hotwords can effectively improve speech recognition accuracy.

Analytical Model

The following models are currently supported:

• Default (official model combination for optimal results)
• tingwu-plus
• qwen-plus
• qwq

Analysis options

Configure the large model analysis options, which include the following capabilities:

Key point extraction: Supports extracting keywords and to-do items.

Summarization: Supports extracting a full-text summary, speaker summaries, key point reviews, and a mind map.

• Full-text summary: Choose between a paragraph summary or a Markdown-formatted summary.
• Mind map: Choose whether to include timestamps.

PPT analysis: When the content source is Recorded file recognition and the file is a video, this feature supports extracting and summarizing the presentation from the PowerPoint (PPT) slides in the video.

Others: Includes chapter overview, spoken-to-written text conversion, and custom prompt features.

• Chapter overview: Set the overview granularity (fine, medium, or coarse) and title length (long, medium, or short).
• Custom Prompt: Define your own large model prompts to guide the model in completing various custom tasks. The custom prompt model can be selected separately and does not conflict with the chosen analysis model. When using the custom prompt feature, you can set the large model's output format. The available formats are: speaker information and transcription results only, sentence number with speaker information and transcription results, or transcription results only.

Callback Method

• HTTP POST protocol: Enter the corresponding URL. Tingwu sends a POST request to the configured HTTP URL. The check is successful if the server returns an HTTP status code of 200.
• No callback (active polling): If you choose not to set a callback, you can actively poll for results using the getTask API for the Agent task. The default number of queries per second (QPS) for polling for each account is 100.
• EventBridge: The processing results are sent to your service based on the default configuration of EventBridge in the China (Beijing) region. You must activate Alibaba Cloud EventBridge first.

A project is created for the application only after the server-side receives the callback message and returns an HTTP status code of 200. The message format is as follows:

{
  "output": {
      "transcriptionPath": "https://***.oss-cn-hangzhou.aliyuncs.com/***",
      "status": 0
  },
  "requestId": "***",
  "taskInfo": {
      "dataId": "***",
      "userSpaceId": "llm-***",
      "appId": "***",
      "model": "tingwu-***",
      "userId": "***"
  }
}

Application name

You can copy or modify the application name here.

Application ID

You can view or copy the application ID here.

Application Description

You can add a description for the application here.

Conversation content

• If the content source is Recorded file recognition, Real-time transcription, or Tingwu task, the conversation content box displays the corresponding audio or video transcription results. The box also shows the speaker ID and dialogue timestamp. You can play the audio file for comparison.
• If the content source is Meeting text content, the dialogue timestamp information may not be displayed. However, this does not affect the analysis results.

Analysis results

In the analysis results box, you can view the analysis results for intelligent overview, mind map, and custom prompts. You can configure the analysis items on the left. For more information, see Configure and test.

The intelligent overview page includes results such as keywords, full-text summary, chapter overview, speaker summaries, key point review, and extracted PPT content.

Chapter overview example:

Speech summary:

Key takeaways:

Retrieve PPT effects:

Extracted mind map:

Test Records

All test results for the current application are saved. Click the Test Records button in the upper-right corner of the console to view them.

The test records list displays various details. These details include the test time, task ID, task status, analysis model, dialogue content, and a Details button.

Click the Details button in a test record's entry to view the complete meeting content and analysis results.