Inworld

A one-sentence summary

Inworld AI is a development platform designed for real-time voice applications; it integrates text-to-speech, speech-to-text, full-duplex dialogue, model routing, and hosted inference within a single accounting and billing system.

Tool Introduction

Inworld AI is operated by Theai, Inc. It initially gained attention for its capabilities related to game characters and AI NPCs; today, its focus has shifted to the infrastructure needed for real-time voice processing and low-latency inference in consumer applications. It functions more like a developer API platform rather than a voice-editing tool intended for ordinary users.

The main applications include voice companionship, language learning, games and interactive media, customer service, outbound sales calls, and intelligent phone agents. Developers can use just one voice model, or they can combine various elements through the Realtime API to create complete conversation flows.

Current product composition

Realtime TTS

Realtime TTS offers two model options: TTS-2 and TTS-2 Flash. The former focuses on natural language control and expressiveness, while the latter emphasizes lower latency and lower character costs.

The latency metric is measured on the server side and does not take into account the user’s network, application buffering, or the processing time of the playback device. The actual end-to-end response time also depends on text generation, connection quality, and the client implementation.

Voice control and output

• It is possible to control emotions, enunciation, tone, volume, pitch, speed, and speaking style using natural language.
• Instructions, pauses, and non-verbal sounds can be added to the text.
• It supports stream-based generation, making it suitable for synthesis and playback simultaneously.
• It can return word, phoneme, and lip-sync timestamps for subtitles and character lip synchronization.
• It supports control of speech speed, tone, custom pronunciation, and various audio encoding settings.
• Playground makes it easy to listen to a sound first, before using the same sound identifier in the program interface.

Voice cloning and sound design

Instant Voice Cloning allows for the creation of a voice using about 5 to 15 seconds of reference audio, enabling that same voice to be produced in over 200 different languages. Voice Design, on the other hand, generates new voices directly based on textual descriptions such as age, accent, timbre, and energy level.

1. Confirm that the recorder has given clear consent to cloning, storing, and the intended use.
2. Record clear samples with no background music, no echo, and only one person present.
3. After creating a sound, save the identifier of the resulting sound.
4. Use short texts to test the original language, cross-lingual capabilities, and various expression instructions.
5. Check for similarity, risk of misguidance, sensitive words, and commercial licensing.
6. Add identity disclosure, access control, and abuse monitoring to public deployments.

Professional Voice Cloning is available as an additional feature for Developers; Growth includes 1 such option, while Enterprise versions are configured according to the contract terms. Voice cloning involves data that can be considered biometric information, and it is not permissible to copy someone else’s voice without authorization.

Realtime STT

Realtime STT is responsible for converting speech into text, and it can identify characteristics such as age range, pitch, emotion, speaking style, and accent through voice profiling. This product can be integrated with Inworld’s own transcription services as well as other speech recognition providers.

Realtime API

The Realtime API combines speech recognition, large models, and speech synthesis through a continuous WebSocket session, enabling users to interrupt, resume, and speak alternately in a natural manner. It is compatible with the OpenAI Realtime protocol, facilitating the migration of existing clients.

• It features built-in voice activity detection to determine when the user starts and stops speaking.
• Smart Turn helps to distinguish between brief pauses and the actual end of a speech.
• You can use the models available in the Router, or you can choose to integrate your own models as per your design.
• It supports multi-modal sessions that consist of audio, text, and certain image content.
• The underlying TTS, STT, and LLM services are charged based on actual usage.
• A persistent connection requires handling heartbeats, reconnections, cancellations, and session status restoration.

Build a voice agent

1. Create accounts, workspaces, and API credentials intended solely for use on the server side.
2. First, determine the TTS voice, STT settings, and target model in Playground.
3. Design system instructions, tools, user interruption rules, and conversation termination rules.
4. Establish real-time sessions and handle audio formats, events, and streaming playback correctly.
5. Set fallback paths for network disruptions, model timeouts, and tool failures.
6. Record latency, transcription errors, interruption success rate, task completion rate, and cost per minute.
7. Launch on a small scale first, then increase concurrency and verify that the billing matches the actual usage.

Realtime Router

Realtime Router provides access to over 220 models through an interface that is compatible with the calling methods of OpenAI and Anthropic. Third-party models are resold at the costs specified by their providers, without any additional charges; however, a fee is applied when proprietary keys are used.

Models and billing methods

The price information was verified on August 23, 2026; the actual amounts, taxes, exchange rates, and discounts may vary, and the final figures will be those displayed on the settlement page.

The Router covers providers such as OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, xAI, as well as open-weight models optimized for Inworld. The list of models and their respective prices change frequently, so it is necessary to retrieve this information dynamically from the current model catalog.

Playground and API development

The platform offers playgrounds such as TTS and Realtime, which facilitate testing of sounds, models, and parameters before writing code. For actual applications, integration is achieved through server-side APIs or WebSocket; temporary tokens should be used for real-time browser sessions in order to prevent the exposure of the main key.

• REST is suitable for one-time TTS, STT, and resource management requests.
• Stream-based TTS returns audio content in multiple lines of JSON, which requires decoding line by line.
• WebSocket is suitable for full-duplex real-time voice communication and maintaining a persistent state.
• Routers are compatible with common SDKs; however, it is still necessary to check those extension fields that are not identical during migration.
• Credentials are subject to concurrency limits based on the account, and should be separated by environment as well as rotated regularly.
• Production systems must implement timeout, exponential backoff, quota protection, and log anonymization.

Subscription plans

The payment options include monthly or annual billing; the annual fee is charged in a single installment. The price of the package is converted into equivalent usage points in dollars, and any additional usage is covered by further purchases at the applicable rate or through automatic top-ups.

Unit prices for TTS and STT

Any excess usage under the same plan is charged at the price applicable to that tier. The Realtime API does not have a fixed rate per minute; instead, it charges separately based on the actual number of TTS characters, the duration of the STT audio, and the number of LLM tokens used.

Concurrency and capacity

The concurrency limits are shared across accounts and API keys; requests that exceed these limits will be processed only as capacity permits. Capacity planning should be based on peak values and stress tests, and the estimated number of sessions cannot be used as a guarantee for service performance.

Validity period of points and account changes

• Unused credits from a paid plan can be carried over for up to 3 months, as long as the subscription remains active and the value of the plan does not decrease.
• The points issued each month have a separate 3-month validity period.
• The minimum amount to purchase additional points is 10 dollars, and they are valid for 1 year after purchase.
• Automatic top-up can be set up to replenish points when the balance falls below a certain threshold.
• The upgrade takes effect immediately, and the full cost of the new plan plus any unused usage up to that point will be charged.
• The downgrade or cancellation takes effect at the end of the current cycle, and any remaining points will be deleted.
• After consecutive payment failures and 7 days of retry attempts, the account will be downgraded to On-Demand status and its API will be disabled.

Responsibilities for sound and privacy

The privacy policy includes chats, recordings, and voice characteristics derived from sound among the data that may be processed; some of these characteristics may constitute biometric information in certain regions. Voice cloning and analysis require appropriate consent, and mechanisms for deletion and withdrawal must be provided.

Privacy Policy Note: Inworld may use certain information for research and development as well as for model training; it also offers an option that allows companies to avoid any data retention. For sensitive or regulated workloads, the terms related to termination of training, retention periods, and the scope of applicable products must be specified in the contract.

Safety and compliance

GitHub, SDKs, and open-source status

Inworld has an active official GitHub organization that makes available TTS code, API examples, runtime templates, MCP tools, voice migration tools, as well as various sample applications. The TTS code, API examples, and those templates are licensed under the MIT license.

The public examples and templates do not imply that the hosting platforms, models, or all SDKs are open source. Certain Runtime binaries, as well as older versions of the Unity and Unreal SDKs, are subject to specific licensing terms; it is necessary to check the license for each repository and each component before using them.

Which users are it suitable for

• Form a team to develop consumer applications for voice-based companionship, social interaction, language learning, and interactive features.
• Developers who need low-latency multilingual TTS and voice cloning.
• The goal is to integrate STT, LLM, and TTS into a single voice intelligence agent for real-time conversations.
• A platform engineering team that is required to handle routing, testing, and failover among more than 220 different models.
• Companies that require large-scale concurrency, dedicated capacity, or on-premises deployment.
• Developers who use Unity, Unreal, Node.js, and common voice frameworks to create interactive experiences.

Product advantages

• TTS, STT, Realtime API, and the model Router share accounts, credits, and development documentation.
• TTS-2 takes into account expression control, while Flash focuses on low latency and low costs at large scales.
• Sound cloning samples are short in length, and they allow the preservation of the voice’s identity across more than 200 languages.
• The Router is compatible with common calling methods and supports conditional routing, testing, and automatic failover.
• The subscription price is converted into available credits; the premium plan sees a reduction in the unit price along with an increase in concurrent usage.
• Options include SOC 2 Type II, enterprise ZDR, on-premises deployment, and data residency options.
• The official GitHub examples and templates cover various programming languages as well as real-time speech frameworks.

Usage restrictions and cost risks

• The total cost of real-time voice processing is composed of TTS, STT, and LLM – it’s not possible to consider only the cost per character.
• Paid points can be carried forward for a maximum of 3 months; any unused balance is cleared in the event of cancellation or downgrade.
• High concurrency, professional voice cloning, ZDR, HIPAA, and local deployment require higher-tier solutions or additional features.
• The server latency metrics do not include network delays, model generation time, or client playback time.
• Voice profiling is a form of probabilistic inference; it cannot replace identity verification, emotion detection, or human judgment.
• Third-party models and proprietary keys may introduce additional contracts, warranties, and costs.
• The core cloud services and models are not open source; the public repositories contain mainly code, templates, and examples.

Go-live checklist

1. Determine whether only TTS or STT is needed, or if a full Realtime session and Router are required.
2. Test sound quality, transcription, and interruptions using representative languages, devices, and networks.
3. Estimate characters, audio hours, LLM tokens, and concurrent sessions based on the peak value.
4. Compare package points, unit price, rollover period, and excess fees to set up budget alerts.
5. Obtain explicit authorization for voice cloning, and assess biometric and recording regulations.
6. The master key is stored on the server, while the browser uses only short-term credentials.
7. Verify third-party model retention, ZDR, HIPAA, data residency, and deletion requirements.
8. Handle deployment timeouts, reconnections, model rollback, and manual intervention, then gradually increase the scale.

Frequently Asked Questions

What is Inworld AI doing mainly these days?

The current focus is on real-time voice and reasoning infrastructure, including TTS, STT, full-duplex voice APIs, model routers, and hosted computing, rather than just the earlier AI role tools.

Is Inworld AI free?

On-Demand can be started for free, offering up to around 70 minutes of TTS service or 400 minutes of STT service. Beyond this free quota, charging applies based on actual usage; some advanced models require a payment method to be linked.

Is Chinese supported?

The TTS-2 series supports over 200 languages, including Chinese. Before putting it into use, it is necessary to test each type of output – standard Mandarin, different accents, numbers, proper nouns, and custom pronunciations.

How many recordings are needed for voice cloning?

Instant cloning typically provides clear audio within 5 to 15 seconds. Professional cloning is part of higher-tier packages or additional services, and it requires compliance with sound licensing requirements.

How is the Realtime API priced?

There is no separate fixed price per minute; the cost for TTS characters, STT audio duration, and LLM tokens within a session is calculated based on the respective rates outlined in the current plan.

Will the Model Router have an additional cost?

Third-party models are routed through Inworld, and billing is based on the provider’s costs; no additional fee is charged for the router. When using an existing key, a 4% service fee is specified in the documentation.

Will the points be carried over?

The monthly points under a paid plan can be carried over for up to 3 months, as long as the subscription remains at the same level or at a higher one; additionally purchased points are valid for 1 year. The remaining points are deleted when the subscription is canceled or downgraded.

Is Inworld AI open-source?

The core platform and models are not fully open source. The official team has made available TTS code, API examples, and various templates under the MIT license, but some SDK binaries are subject to separate licensing terms.

Summary

Inworld AI is suitable for teams that need to integrate natural speech processing, low-latency conversations, and multi-model reasoning into large-scale consumer applications. It can provide TTS or STT as standalone services, or it can be used together with Realtime APIs and routers to create a complete voice intelligence framework.

When selecting a plan, it is necessary to take into account the length of the text, the audio duration, the number of model tokens, the level of concurrency, and the validity period of the credits. In cases involving voice cloning, medical data, or sensitive information, authorization procedures, ZDR rules, contracts, and human oversight must also be implemented in addition to technical testing.