Curvature in the cloud

What is cloud curvature?

Cloud Curvature LiveData is a set of real-time interaction and AI cloud services designed for games, social applications, live streaming services, and applications intended for use in international markets.

It integrates real-time messaging, audio and video communication, multilingual translation, speech recognition, and content moderation into the developer platform.

Main functions

• Provides real-time global messaging services.
• It supports one-on-one as well as multi-person audio and video interactions.
• Provides real-time multilingual text translation.
• Convert multilingual speech into text.
• Translate the audio content into text in the target language.
• Review multilingual chat texts.
• Identify illegal content in images.
• Detect risk information in audio.
• Take screenshots of the video and conduct a review.
• REST APIs and client SDKs are provided.
• Provides console testing and data statistics.
• The open-source ViiTorVoice speech generation project.

Which users are it suitable for

Real-time Messaging RTM

RTM is used for transmitting chat, status, and business messages; official documentation indicates that the average global message latency is around 100 milliseconds.

The actual delay is influenced by the user’s location, network quality, message size, and server processing time; stress testing should be used as a reference.

• Supports high-concurrency real-time messaging.
• It can be used for games and social interactions.
• It supports horizontal scaling and dynamic capacity expansion.
• The console can simulate the server interface.
• Provides messaging and usage statistics.
• It can combine translation and review capabilities.

Real-time audio and video RTC

RTC is suitable for low-latency applications such as voice chat rooms, gaming partnerships, live streaming with audio connection, video conferences, and online interactions.

1. Choose between one-on-one or group call mode.
2. Select the terminal platforms that need support.
3. Create a project and obtain access credentials.
4. Integrate the corresponding client SDK.
5. Implement room, user, and permission logic.
6. Test weak networks, disconnections, and device switching.
7. Add recording, review, and complaint mechanisms.
8. Concurrency and cost tests are conducted before going live.

Real-time text translation

Real-time text translation is optimized for conversational speech, abbreviations, and internet slang; the official website indicates support for translation between more than 50 languages.

Speech recognition and speech translation

Speech recognition converts audio into text, while speech translation produces the text in the target language.

Different products support a varying number of languages; before integration, it is necessary to check the source language, target language, and audio requirements according to the interface documentation.

• Suitable for real-time subtitles in live broadcasts.
• Suitable for cross-lingual voice social interactions.
• Suitable for conference and teaching transcription.
• Customization services for industry-specific terminology are available.
• It can be combined with sensitive word filtering.
• As a result, sampling is required for manual verification.

Multimodal content moderation

Automatic review may lead to errors or omissions; for high-risk content, as well as for appeals and suspension decisions, a manual review process should be available.

Comprehensive tutorial on integrating cloud curvature

1. Register an account and create company information.
2. Create subsidiaries, project teams, and projects.
3. Choose from real-time interaction, translation, or review services.
4. Confirm the target platform and SDK version.
5. Obtain the project credentials and store them securely.
6. First, complete the interface simulation in the console.
7. Integrate the minimum functionality in the testing environment.
8. Handle timeouts, retries, and duplicate callbacks.
9. Use real language corpora to test translation quality.
10. Establish a manual review process for the audit results.
11. Count the number of characters, minutes, and image requests.
12. Set budgets, alerts, and rate limits.
13. Complete privacy and regional compliance checks.
14. Launch in grayscale and continuously monitor the data.

Billing method and prices

The cloud curvature service is billed based on the product, the amount of usage, and the detection coefficient; the specific pricing details can be found in the control panel or in the commercial quotes.

The coefficient is used to convert the volume of service covered by a package, and it is not equivalent to the price in RMB; the final cost should be calculated based on the capacity of the package.

How to control costs

• Select the necessary testing scenarios based on the actual business needs.
• Avoid submitting the same content repeatedly.
• For long texts, determine the segmentation strategy first.
• The video adjusts the frame rate based on risk.
• Set a maximum number of retries for failures.
• Distinguish between testing, production, and different projects.
• Regularly check the usage amount against the bill.
• An alert is issued promptly when approaching the package limit.

What is ViiTorVoice-NAR?

ViiTorVoice-NAR is a non-autoregressive speech generation system made available by viitor-ai, which enables voice cloning and local speech editing.

• Use the reference audio to clone sounds.
• Only the modified partial segments are regenerated.
• Supports control of emotions and paralinguistic cues.
• Provides separate gRPC services.
• Provides an end-to-end HTTP gateway.
• Supports Docker and local deployment.
• The model weights are made public under the Apache-2.0 license.

ViiTorVoice Local Deployment Tutorial

1. Prepare a Linux environment, Python, and a GPU-compatible setup.
2. Download the official viitor-voice-nar repository.
3. Read the license and upstream model terms.
4. Initialize the runtime environment script.
5. Download the model file to the specified directory.
6. Verify that the model file is not a broken soft link.
7. Start the encoder, model, and decoding services.
8. Check the HTTP health status.
9. Use authorized reference audio for testing.
10. Compare the timbre, pronunciation, and local seams.
11. Add authentication, rate limiting, and audit logging.
12. It will be made available for use only after passing the security review.

Compliance requirements for voice cloning

Voice cloning can only use one’s own voice or audio for which explicit permission has been obtained; it is not allowed to create deceptive content that impersonates others.

• Record the scope of authorization for the sound owner.
• Define the purpose, deadline, and channels for publication.
• Impersonating public figures and legitimate organizations is prohibited.
• It must not be used for fraud, harassment, or false endorsement.
• Clearly label the synthesized audio.
• Save the generated records and operator logs.
• Mechanisms for revoking authorization and deletion are provided.
• Stricter restrictions apply to the voices of minors.

Data security and privacy

• The credentials are stored only on the secure server.
• Chat and audio data are minimized as needed.
• Verify the data transmission rules for cross-border transactions.
• Restrict employees from exporting the original audit content.
• Sensitive data is anonymized before it is recorded in the logs.
• Establish a review process for model and SDK updates.
• Delete the data that is no longer needed once it has expired.

APIs, SDKs, and open-source status

Product advantages

• Integration of real-time interaction with AI capabilities.
• It covers multi-language text and voice scenarios.
• Supports review of text, images, audio, and video.
• Suitable for gaming, social interaction, and live streaming on an international scale.
• Complete documentation, APIs, and SDKs are provided.
• There are open-source voice projects that can be deployed locally.

Usage restrictions

• The price of the complete package needs to be inquired about or viewed by logging in.
• Errors may occur in the translation and recognition results.
• Content moderation cannot completely replace human oversight.
• Real-time performance is affected by the network and location.
• Different products support a varying number of languages.
• Open-source voice projects require local computing power.
• Clear authorization is required for voice cloning.

Frequently Asked Questions

What services does cloud curvature mainly provide?

It offers real-time messaging, real-time audio and video communication, multilingual text translation, speech recognition, voice translation, and multi-modal content moderation services.

What types of content moderation is supported by cloud curvature?

It supports the review of text, images, audio, and video; videos are usually processed by taking screenshots, while high-risk results still require manual verification.

How is the pricing for cloud curvature?

The platform charges based on the product, the amount of usage, and various coefficients; for amounts that exceed the scope of the annual package, the standard charge is 1.6 yuan per thousand units of usage. The total cost of the package is determined according to the current pricing.

Does cloud curvature support real-time translation?

Real-time text and voice translation is supported; different interfaces cover a range of different languages, so it is necessary to check the specific product documentation before integration.

What can ViiTorVoice-NAR do?

It can clone voices based on authorized reference audio, and it can also identify changes in the text and regenerate only the corresponding partial audio segments.

Can ViiTorVoice-NAR be used for commercial purposes?

The main model page is licensed under Apache-2.0, but before it can be used for commercial purposes, it is necessary to examine the code, model components, training data, and related licenses, as well as obtain any required sound-related permissions.

Is the Cloud Curvature Platform open-source?

Commercial cloud platforms are not open source; although the developers have made projects such as ViiTorVoice-NAR available publicly, it cannot be assumed that an individual open-source repository represents an open-source platform as a whole.