How to identify voice compliance risks with Voice Moderation 2.0-AI Guardrails(AI Guardrails)-阿里云帮助中心

If you operate a live streaming, voice chat, or media sharing platform, manually monitoring audio for compliance violations is impractical. This topic shows how to use Voice Moderation 2.0 to automatically detect audio content risks such as profanity, pornography, political content, and AIGC-generated voices. You will learn how to activate the service, configure moderation rules, integrate the API, view results, and monitor usage.

Limitations

You can query moderation results from the last 30 days only.
The Result Query page stores up to 50,000 records. For longer-term storage, save API results to your own system.
Each lexicon supports up to 100,000 keywords.
You can create up to 20 lexicons per account.
A single keyword cannot exceed 20 characters. Special characters are not supported.

Compared to Voice Moderation 1.0, the Voice Moderation 2.0 service offers more capabilities for improved content security.

Feature	Voice Moderation 2.0	Voice Moderation 1.0
Business scenario support	Pre-configured for multiple business scenarios, including social live streaming, comprehensive audio and video, and voice chat rooms. This simplifies integration and expands compliance risk coverage.	The default configuration is for general scenarios and lacks business-specific features. You must define your standards and adapt the policy before use.
Voice slicing	Uses an adjustable slicing method to create fixed-duration voice slices, improving manual review efficiency.	Uses semantic segmentation, resulting in voice slices ranging from a few to dozens of seconds.
Moderation capabilities	Supports over 13 moderation labels, including new types like moaning and silence. Supports multiple moderation labels and no longer returns moderation suggestions, giving you the flexibility to make decisions based on the returned risk levels.	Supports 10 moderation labels, including terrorism, pornography, and profanity. Supports only a single risk label and returns a moderation suggestion. Supports detection of suspected AIGC audio.
API features	Returns all voice slice content and transcribed text. Provides temporary URLs for voice slices to simplify manual review. Uses the input parameter `service` to distinguish between scenarios.	Returns only voice slices that contain compliance risks. Does not provide temporary URLs for voice slices by default. Uses the input parameter `liveId` to distinguish between live streams and files.

Features

Use cases

Voice Moderation 2.0 provides two service tiers. Choose based on your analysis needs:

LLM Edition: For deep semantic analysis, including sentiment detection, metaphor identification, and ideological content recognition. Powered by a text moderation large language model.
Rule-based services: For keyword matching, cost-efficient moderation of known risk types such as profanity, pornography, and terrorism.

Note

LLM Edition services are newly available. See the first four rows below for details.

Use LLM Edition when you need deep semantic analysis of nuanced content. Use rule-based services for cost-efficient moderation of well-defined risk types.

Service name	service	Use cases
Audio and video media detection (LLM Edition)	audio_detection_byllm	Suitable for diverse audio and video content, including general videos, podcasts, and audiobooks. Uses a large language model for text moderation that performs ASR followed by deep semantic analysis. Detects positive and negative sentiment, identifies metaphors, and recognizes ideological content.
Audio and video media detection (LLM Edition, Global)	audio_detection_byllm_cb	Designed for international platforms serving global audiences. Leverages a multilingual large language model for deep semantic analysis across languages. Uses an international labeling system with region-specific risk labels for global compliance.
Social live stream detection (LLM Edition)	stream_detection_byllm	Designed for social live streaming, typically with a single speaker and potential background music. Performs ASR followed by deep semantic analysis with context-aware risk identification. Detects sentiment, identifies metaphors, and understands nuanced content beyond keyword matching.
Social live stream detection (LLM Edition, Global)	stream_detection_byllm_cb	Designed for international social live streaming platforms serving global audiences. Leverages a multilingual large language model for deep semantic analysis across languages. Uses an international labeling system with region-specific risk labels for global compliance.
Rule-based services (existing)
Social live stream detection	live_stream_detection	For social live streaming scenarios, typically involving a single speaker with potential background music. The Voice Moderation 2.0 service improves model accuracy in complex environments and reduces interference from dialects, accents, and background noise. Focuses on identifying risks such as pornography, profanity, and undesirable content, while also detecting periods of silence in live streams.
Social live stream detection (Professional Edition)	live_stream_detection_pro	Includes all features of the social live stream detection service and adds detection of political content.
Social live stream multilingual detection	stream_multilingual_cb	For international social live streaming, typically involving a single speaker with potential background music. Supports audio content in 35 languages, including Chinese, English, Arabic, German, Russian, French, Korean, Japanese, Spanish, and Italian. It uses an international labeling system, adding risk labels such as profanity and location-specific terms. For more information, see Voice Moderation 2.0 multilingual service.
Audio and video media detection	audio_media_detection	Primarily for various types of audio and video media, such as general videos and audio novels. Voice Moderation 2.0 improves support for various encoding formats and boosts detection performance. It also optimizes model effectiveness for non-semantic features. Due to the wide range of content types, it adds ideology-related detection in addition to risks like pornography, profanity, and terrorism.
Audio and video media detection (Professional Edition)	audio_media_detection_pro	Includes all features of the audio and video media detection service and adds detection of political content.
Audio and video media multilingual detection	audio_multilingual_cb	For international audio and video media, such as general videos and educational audio. Supports audio content in 35 languages, including Chinese, English, Arabic, German, Russian, French, Korean, Japanese, Spanish, and Italian. It uses an international labeling system, adding risk labels such as profanity and location-specific terms. For more information, see Voice Moderation 2.0 multilingual service.
AI-generated voice detection	voice_aigc_detector	Detects on demand if audio is likely AIGC. This service is suitable for any scenario where you need to detect and label AI-generated or synthesized audio.

Moderation labels

This service supports a wide range of moderation labels and can return multiple labels for a single audio file. The following table lists the available labels.

Label type	Category
Voice Moderation labels (labels)	ad: Advertising violence: Violence and terrorism political_content: Political content specified_speaking: Specific speakers specified_lyrics: Specific songs sexual_content: Pornography sexual_sounds: Moaning sounds contraband: Contraband profanity: Profanity religion: Religious content cyberbullying: Cyberbullying negative_content: Undesirable content nontalk: Silent audio C_customized: Custom keyword match
AI-generated voice moderation labels (labels)	aigc: Audio suspected to be AIGC ugc: Audio is not AIGC
LLM Edition labels (labels)	Uses a large language model with deep semantic understanding capabilities. Detects sentiment, identifies metaphors, and recognizes ideological content beyond keyword-based matching. Label output varies by service: `audio_detection_byllm`, `audio_detection_byllm_cb`, `stream_detection_byllm`, or `stream_detection_byllm_cb`.

Moderation performance

The Voice Moderation 2.0 service uses a high-performance core engine that schedules dozens of models and policies with high concurrency, delivering low-latency protection for scenarios like live audio streaming and voice chat.

Performance metric	Description
File size	The maximum audio file size is increased from 200 MB to 500 MB.
QPS (Queries Per Second)	The task submission QPS is increased from 50 to 100.
Concurrency	The default concurrency limit is increased from 20 to 50.

Note

QPS for Voice Moderation is the number of API requests processed per second. Concurrency is the number of audio files or streams being detected simultaneously.

Billing

The Voice Moderation 2.0 service supports two billing methods: pay-as-you-go and resource plan deduction.

Pay-as-you-go

After you activate the service, the default billing method is pay-as-you-go, where you are billed daily based on your usage. You are not charged if you do not use the service. For more information, see Activate Content Moderation-Enhanced Edition.

Moderation type	Applicable services	Unit price
Standard Voice Moderation (audio_standard)	Social live stream detection: live_stream_detection Audio and video media detection: audio_media_detection	CNY 225 per 10,000 minutes, equivalent to CNY 1.35 per hour.
Advanced Voice Moderation (audio_advanced)	Audio and video media detection (Professional Edition): audio_media_detection_pro Social live stream detection (Professional Edition): live_stream_detection_pro Audio and video media multilingual detection: audio_multilingual_cb Social live stream multilingual detection: stream_multilingual_cb AI-generated voice detection: voice_aigc_detector	CNY 375 per 10,000 minutes, equivalent to CNY 2.25 per hour.
LLM Edition Standard (audio_llm_standard)	Audio and video media detection (LLM Edition): audio_detection_byllm Social live stream detection (LLM Edition): stream_detection_byllm	CNY 300 per 10,000 minutes, equivalent to CNY 1.80 per hour.
LLM Edition Advanced (audio_llm_advanced)	Audio and video media detection (LLM Edition, Global): audio_detection_byllm_cb Social live stream detection (LLM Edition, Global): stream_detection_byllm_cb	CNY 450 per 10,000 minutes, equivalent to CNY 2.70 per hour.

Resource plan deduction

If you have a large volume of content to moderate or a consistent moderation workload, purchase a resource plan in advance. Larger plans offer greater discounts. You can purchase and use multiple plans simultaneously. For more information, see Purchase a resource plan for Content Moderation-Enhanced Edition.

This resource plan deducts charges for the Voice Moderation 2.0 service and is not compatible with Content Security traffic packages. The deduction factors are as follows:

Moderation type	Deduction factor
Standard Voice Moderation (audio_standard)	The deduction factor is 30. This means that for every minute of audio successfully processed, 30 units are deducted from your resource plan. For example, if you purchase a resource plan with a capacity of 100 units and you moderate one minute of audio, 30 units are consumed, leaving 70 units.
Advanced Voice Moderation (audio_advanced)	The deduction factor is 50. This means that for every minute of audio successfully processed, 50 units are deducted from your resource plan. For example, if you purchase a resource plan with a capacity of 100 units and you moderate one minute of audio, 50 units are consumed, leaving 50 units.
LLM Edition Standard (audio_llm_standard)	The deduction factor is 40. This means that for every minute of audio successfully processed, 40 units are deducted from your resource plan. For example, if you purchase a resource plan with a capacity of 100 units and you moderate one minute of audio, 40 units are consumed, leaving 60 units.
LLM Edition Advanced (audio_llm_advanced)	The deduction factor is 60. This means that for every minute of audio successfully processed, 60 units are deducted from your resource plan. For example, if you purchase a resource plan with a capacity of 100 units and you moderate one minute of audio, 60 units are consumed, leaving 40 units.

Step 1: Activate service

Before using the Voice Moderation 2.0 feature, you must activate the service.

Go to the Content Moderation-Enhanced Edition page, and then read and agree to the service agreement.
Click Activate Now.

Step 2: Configure moderation rules

If the built-in labels are insufficient for your business needs, create custom lexicons and configure rules to flag or ignore specific terms.

Log on to the Content Security console.
In the left-side navigation pane, choose Machine Moderation > Plus > Voice Moderation > Rules.
Configure a lexicon.
1. On the Keyword Library Management tab, click Create Library.
2. In the Create Library panel, provide a Library Name and add words by using Add words by this page or Upload File.
  Select the Create a library first and add words later option to create a library without immediately adding keywords. Add them later based on your business needs. Add up to 100,000 keywords and create up to 20 libraries in a single account. A single keyword cannot exceed 20 characters, and special characters are not supported.
3. Click Create Library.
  If the lexicon fails to be created, a message appears. Follow the instructions in the message to try again.
  In the Operation column, click Modify or Clear All to modify keywords or remove all keywords at once.
Configure rules.
1. On the Rules Management tab, select the target service, and then click Set Thesaurus in the Operation column.
2. In the Settings panel, select the ignore lexicon. Then, click Next.
  This feature is used to whitelist keywords, preventing them from being flagged.
  For example, if the keywords in the ignored lexicon are moderator and fan, and the text converted from speech for review is Welcome to the livestream, double-tap to like, fan badge plus moderator gets streamer friend spot, then the keywords moderator and fan are first ignored. A risk detection is then performed only on the remaining text: Welcome to the livestream, double-tap to like, badge plus gets streamer friend spot.
3. Select the match lexicon, and then click OK.
  If any keyword from a lexicon is successfully matched with the text transcribed from the audio, the labels parameter returns C_customized when you call the Voice Moderation 2.0 API. The value C_customized indicates a match in a custom lexicon that you created. This scenario is primarily used to detect violation risks in the text transcribed from the audio.
  For example, suppose the keywords in a matched lexicon are room manager and fans. If the text converted from audio for moderation is Welcome to the livestream. Double-tap to like. Get a fan badge and become a room manager to be added as the streamer's friend, the audio risk detection process matches the keywords room manager and fans. When you call the Voice Moderation 2.0 API, the value of the returned labels parameter includes C_customized in addition to any matched built-in labels.
The rule configuration takes about three minutes to take effect.

Step 3: Integrate the service

The Voice Moderation 2.0 service supports the following integration methods.

Submit a Voice Moderation task by calling the DetectAudit API. Specify the service parameter to select a moderation service. For example, to use the LLM Edition service for audio and video media detection:


curl -X POST 'https://green-cip.ap-southeast-1.aliyuncs.com/?Action=DetectAudit&Version=2022-03-02' \
  -H 'Content-Type: application/json' \
  -d '{
    "Service": "audio_detection_byllm",
    "AudioUrl": "https://example.com/sample.mp3"
  }'

Replace the endpoint with the one that matches your region. For the Service parameter values, see the Supported business scenarios table above.

Call the API to integrate the service. For more information, see Voice Moderation 2.0 API.
Use an SDK to integrate the service. For more information, see SDK Integration Guide.
Use HTTP to integrate the service. For more information, see HTTP Integration Guide.

Step 4: View moderation results

View moderation results to analyze common violation types in your audio content.

Log on to the Content Security console.
In the left-side navigation pane, choose Machine Moderation > Plus > Voice Moderation > 文档页结果.
On the 文档页结果 page, view the moderated audio, matched labels, and request times.
Search for specific information by setting a Time Range or by filtering by Request ID, Text (from transcribed audio), or Label. Query data from the last 30 days. The Result Query page can store up to 50,000 records. If you require longer-term storage, save the API results to your own system.
When searching by label, use the following filter options:
- Contains: Returns results where the label field includes the specified label value.
- Does not contain: Returns results where the label field does not include the specified label value.
- Empty: Returns results that did not match any label.
- Not empty: Returns results that matched any label (no label value input is needed).
Locate a specific transcribed text entry and click View in the Operation column.
In the View details panel, review the detailed moderation information for that transcribed text.

Step 5: Track usage statistics

Track call volumes to monitor service usage for your Alibaba Cloud account and its RAM users.

In the left-side navigation pane, choose Machine Moderation > Plus > Voice Moderation > Statistics.
On the Statistics page, view risk statistics for Voice Moderation, including high, medium, low, and no-risk counts, as well as risk trends and distributions.
On the Statistics page, view the Voice Moderation call volume.
Specify a custom time range to query call volume data from the last 365 days. Query the data Daily or Monthly. By default, the Daily view displays the daily call volume for the last 30 days, and the Monthly view displays the monthly call volume for the last 12 months. Also view call volume data based on Alibaba Cloud accounts and their RAM users.