Build speech generation, transcription and live voice conversations with shared API infrastructure. A browser playground adds voice design, cloning and multi-speaker audio creation.
Used this tool? Rate it
Used this tool? Rate it
Inworld AI is a platform for realtime voice models and LLM inference in consumer applications. It provides text to speech through Realtime TTS-2 and TTS-2 Flash, transcription through Realtime STT-1, and speech-to-speech conversations that combine listening, reasoning and spoken replies. Its target uses include companions, learning, social apps, personal agents, health, fitness and entertainment.
Speech, incoming audio and LLM tokens have separate meters within a shared platform.
The browser TTS Playground also creates downloadable clips without application code.
Realtime TTS-2 supports natural-language steering and 200+ languages and locales; TTS-2 Flash is positioned for latency-critical, high-volume workloads. The model IDs are inworld-tts-2 and inworld-tts-2-flash.
The model documentation reports 100 ms for TTS-2 and 20 ms for Flash at server-side P90 time to first audio byte, excluding network latency. These vendor measurements cover synthesis alone.
Earlier TTS-1 and TTS-1.5 models are deprecated and are not recommended for new projects.
Instant voice cloning uses as little as 3 seconds of audio, up to 30 seconds; Inworld says longer samples improve similarity. The 30-second limit applies to clones created on or after August 15, 2026; earlier clones retain a 15-second prompt.
Voice Design creates a voice from a written description rather than a reference recording. It offers a Freeform mode using an English description of up to 1,000 characters and a Structured mode for editing the profile directly. Editable attributes include dialect, age, emotion, pitch and texture; language and accent also apply to preview scripts.
A design generation produces up to 3 voice previews, and selected voices can be named and saved to the voice library.
Professional Voice Cloning is a beta workflow that uses more audio and an asynchronous training step. It requires at least 10 minutes of combined audio, accepts up to 50 samples totaling 1 GB, and requires confirmation of the rights to clone the voice. Training typically takes 30–60 minutes and can take longer during busy periods.
Realtime STT-1 transcribes 30 languages, detects the spoken language automatically and supports language changes within a conversation. It runs through synchronous and WebSocket endpoints.
Voice profiling returns signals such as emotion, vocal style, estimated age and accent alongside the transcript. These system interpretations can become LLM context and TTS-2 delivery directions.
The speech-to-speech API takes speech in and returns speech over one WebSocket or WebRTC connection, using an LLM chosen by the developer. It can call tools and use regular, cloned or designed TTS-2 voices, with turn detection and interruptions handled within the session. Phone integration uses Twilio Media Streams with G.711 μ-law or A-law audio at 8 kHz.
Realtime controls include semantic turn detection, interruption handling, function calls and changes to the voice, LLM or tools during a session. The service follows the OpenAI Realtime protocol. When a user speaks over a reply, Inworld says playback stops and conversation history retains only what the user heard. Tool definitions can be supplied initially or added later.
Inference accepts OpenAI-compatible Chat Completions requests with an Inworld endpoint, API key and model ID.
Inworld also offers dedicated deployments for a customer's own model, with reserved capacity and a model ID. The company lists the US, Europe, India and Asia Pacific for these dedicated deployments, with zero data retention available on the models it hosts.
Multi-Voice Generation creates a dialogue, scene or podcast as one stitched track, with a separate voice, model and settings for each line. Lines support steering tags, individual regeneration, and comparison or restoration of earlier takes and settings.
The documented browser workflow is:
WebSocket streaming targets low latency; async generation targets long-form audiobooks, podcasts and voiceovers.
For realtime agents, an interruption stops unsent segments; synthesis already requested remains billable.
Inworld positions Creator for content creation and small projects, Builder for growing projects and small teams, Developer for production applications, and Growth for large deployments and compliance needs.
Professional-clone language must match the recordings. Multilingual support is experimental; English en-US is fully supported.
Growth advertises ZDR, HIPAA and BAA add-ons; Enterprise advertises SLA and DPA arrangements plus EU and India data residency.
Inworld provides Node.js and Python SDKs whose generate() methods handle chunking, retries and encoding.
| Synthesis method | Documented boundary | Result delivery |
|---|---|---|
| Non-streaming HTTP | Up to 2,000 characters per request | Complete audio in one response |
| HTTP streaming | Up to 4,000 characters per request | Audio chunks over HTTP |
| WebSocket | Persistent connection | Audio chunks as generated |
| Asynchronous | Up to 100,000 characters per job; 10,000 on On-Demand | Result after background generation |
| Batch | 20% less per character than realtime synthesis | Independent requests processed in the background |
Playground limits are separately 2,000 characters on On-Demand and 40,000 on paid self-serve tiers.
Portal instant cloning accepts one WAV, MP3 or WebM file up to 12 MB and automatically trims audio longer than 30 seconds. Direct recording is also supported, and the flow requires rights confirmation.
Playground generation history is stored locally per device and workspace; it is not synced across browsers or shared between teammates. Clearing browser storage resets it. History supports replay, restoration, filtering and MP3 downloads.
As captured on October 27, 2026, paid tiers include credits equal to their monthly price. Rates use the site's $ symbol, with TTS per million characters and STT per audio hour; the specific dollar currency was not verified.
| Plan | Monthly charge | TTS-2 / 1M characters | TTS-2 Flash / 1M characters | STT / hour |
|---|---|---|---|---|
| On-Demand | $0, pay as you go | $25 | $15 | $0.15 |
| Creator | $25 | $20 | $10 | $0.10 |
| Builder | $100 | $17.50 | $9 | $0.10 |
| Developer | $300 | $15 | $8 | $0.10 |
| Growth | $1,500 | $12.50 | $7 | $0.10 |
| Enterprise | Custom | As low as $5 | Below $5 | Custom |
Annual subscriptions are billed once per year. The website estimates about 1,000 characters per audio minute; TTS is actually metered in characters.
On-Demand includes up to 70 minutes of TTS or 400 minutes of STT for free. The reviewed pricing and billing pages did not specify an expiry date for that free allocation. API and Playground share credits across TTS, STT, Inference and Realtime.
| Plan | TTS generations at once | TTS WebSocket connections | STT streams / Realtime sessions | Custom voices / professional clone slots |
|---|---|---|---|---|
| On-Demand | 5 | 50 | 10 / 10 | 100 / 2 |
| Creator | 10 | 100 | 20 / 20 | 500 / 4 |
| Builder | 50 | 500 | 100 / 100 | 3,000 / 10 |
| Developer | 150 | 1,500 | 300 / 300 | 10,000 / 25 |
| Growth | 500 | 5,000 | 1,000 / 1,000 | 30,000 / 40 |
| Enterprise | Custom | Custom | Custom / Custom | Custom / 100+ |
Concurrency is shared across account API keys; above guaranteed limits, requests are best-effort. Connection and generation ceilings differ.
Unused subscription credits roll over for up to 3 months while an active paid subscription remains on a plan of equal or higher value. Each grant keeps its original expiry. Additional purchases and auto-reload credits instead expire one year after purchase and use the current tier's usage rates. The billing guide sets a $10 minimum for adding credits.
An upgrade takes effect immediately, starts a new billing cycle and charges the full new plan price. Existing metered usage is settled separately, and carried credits keep their original expiry. Downgrades and cancellations take effect at the end of the current billing period, when remaining credits are forfeited. Cancellation is available through Manage plan on the Portal Billing page.
The Terms allow cancellation at any time but state that payments are non-refundable except where required by law, without overriding mandatory local cancellation rights.
Some premium LLMs require a card on file despite an existing credit balance, and asynchronous transcription also requires a payment method.
ElevenLabs offers multi-speaker dialogue support and tools to clone a user's own voice or design a new one. Its overlap is dialogue and narration, with separate limits and rights.
Deepgram markets text to speech designed to retain conversational context and offers cloud, private-cloud and on-premises deployment options. These are the competitor's deployment claims.
These feature descriptions do not establish a comparative quality ranking.
Speaker diarization in STT-1 is experimental; speaker numbers persist within one stream or job and are not carried across separate streams. Drafts and failed professional-clone attempts also occupy voice slots.
Live-response latency includes turn detection as well as synthesis. Inworld reports a fastest first audio of 204 ms after turn closure among 37 turns and a median of 462 ms over 17 measured turns, while high-eagerness semantic VAD took a median 954 ms after the user's last word to close the turn.
The On-Demand plan lists a commercial license. The service-specific terms say customers retain input and output rights as between themselves and Inworld, except as expressly stated, and permit use of downloadable outputs outside the service subject to the Terms and Acceptable Use Policy. Downloads are enabled for some services, not every service.
Voice models must use the customer's voice or a voice they are authorized to share. Customers must have the necessary rights to provide inputs, create outputs and grant the stated licenses. The voice terms license input use for providing, improving and securing the services, and license user voice models for providing and improving them.
For roleplay apps, Inworld discusses NSFW/SFW model selection and routing to match developers' own content policies. This does not establish blanket platform permission for all adult content.
The Acceptable Use Policy requires clear, prominent disclosure that users are interacting with AI rather than a human. In the Voice and Speech Services section, the service-specific terms prohibit inputs containing PHI. A separate Managed GPU Services section refers to a separate BAA for PHI; that exception is stated for that service and does not itself rewrite the voice section.
The Terms say Inworld will not train generally available models on nonpublic Materials, with exceptions for feedback and materials flagged for trust and safety review. The feedback provision grants an unrestricted, perpetual, irrevocable license to use feedback or suggestions; the safety provision allows use or analysis to improve detection and enforcement. The no-training statement has these qualifications.
The Privacy Notice has its own scope: it covers website and related-service information and excludes information processed on behalf of enterprise customers. Within that scope, it describes using information for research and development, including training AI models. Their information categories and conditions differ.
The notice says voice biometric information may be used to provide services and to analyze and improve them unless the user opts out. It retains that information until the collection purpose is satisfied or for up to three years after account access, whichever occurs first. General personal information is retained as long as necessary for the stated purposes, legal requirements or other business purposes explained in the notice.
For residents covered by the notice's section for certain US states, depending on residence and subject to limitations, the notice provides rights to request access, portability, deletion and correction, with verification of requests.
ZDR is included on Enterprise and Enterprise-Trial and is a Growth add-on. It can be configured at organization or workspace level, with the workspace override taking precedence. Enabling it stops future eligible content retention; it does not delete content collected earlier. Eligible inputs and outputs are processed in memory and redacted after processing, with different coverage by product.
Billing metadata remains retained. Voice-cloning samples and Voice Design descriptions or preview scripts are not eligible, and third-party routed LLM requests are outside the hosted-LLM ZDR coverage.
Asynchronous and batch results are normally kept in Inworld-managed storage for about 7 days for download. With ZDR enabled, a job that would retain results there is refused before synthesis; delivery to customer-owned storage is the documented alternative.
Playground and API activity consume the same cross-product balance.
Instant cloning uses a short sample. Professional cloning is beta and adds a larger dataset, background training and separate voice-slot limits.
No. Server-side TTS measurements exclude network latency; recognition, reasoning and turn detection add other parts of the wait.
