Inworld AI

Inworld AI is a voice and inference platform for consumer applications, built around realtime text to speech, speech to text, and speech to speech models that developers add to apps people use every day. Its homepage names the audiences directly: companions, learning, health and fitness, social, personal agents, and entertainment. The voice at the center of the offer is Realtime TTS-2, a text to speech model you can direct in plain language, and Realtime TTS-2 Flash, which the vendor calls the world's fastest voice. Both models stream audio back as it is generated, rather than waiting for a finished file. The performance claims on the homepage are specific and framed as the vendor's own published benchmarks. Realtime TTS-2 reaches first audio in as little as 90 milliseconds and speaks about 500 languages and dialects, and it launched at number 1 on the Artificial Analysis Controlled Voice board. Realtime TTS-2 Flash reaches first audio in as little as 50 milliseconds on the Coval benchmark cited on the page and ties Eleven v3 in blind listening, again as the vendor reports it. Realtime STT-1 handles the other direction, producing streaming transcripts with four voice traits: emotion, style, age, and accent. For a full conversation, the speech to speech API listens, reasons, and talks back over one WebSocket or WebRTC connection, and the developer chooses the language model for each session. Inworld also runs an inference service for open models, serving models such as Gemma, GLM, and DeepSeek through an OpenAI compatible API on infrastructure in four regions: the US, Europe, India, and Asia Pacific. Voice runs in every region, and the page lists SOC 2 Type II, HIPAA, and GDPR references for buyers who need to check compliance fit. Inworld suits product teams building realtime voice into a consumer app who care most about latency, language coverage, and keeping the voice, transcription, and model layers in one provider. Test it with your hardest realtime script, measure time to first audio on your own traffic, and check how direction in plain language changes delivery before you commit a live use case to it.