Speech to text, any LLM and realtime text to speech behind a single API.
Every conversation card is a real session. Speech to speech replies start as fast as 204 ms after a turn ends.
Ask your agent to add Inworld.
People leave 0 to 300 ms between turns. Answer that fast, and your agent feels human.
“What are three easy dinners I could make tonight?” and then, while the agent was still answering, “Actually, just the quickest one.”
0.34 s speech_started 5.10 s speech_stopped 5.10 s transcription.completed What are three easy dinners I could make tonight? 5.10 s response.created 5.77 s output_audio (first chunk) 8.03 s speech_started 8.03 s response.done cancelled because the person cut in 11.51 s speech_stopped 11.51 s transcription.completed Actually, just the quickest one. 11.51 s response.created 11.83 s output_audio (first chunk) 12.54 s response.done completed
“Finally home. I think the interview went well.”
That is wonderful news, I am so proud of you.
“What should I do tonight to switch off?”
How about a long bath and your favorite movie?
Use each part on its own, or all together.
Speech to speech listens, reasons and talks on one connection.
Hears when the person has finished and returns the words with how they were said.
Serves the open model you pick for each session.
Speaks the reply in a voice you direct in plain words.
Listens, reasons and talks on one WebSocket or WebRTC connection, and follows the OpenAI Realtime protocol.
Real recorded sessions.
Press play to hear the person, the agent, and how fast each reply started.
“Hola, ¿practicamos cómo pedir la comida?”
Sí, ¿qué desea comer hoy?
“Quiero una ensalada y agua con gas, por favor.”
Muy bien, ¿desea algo de postre también?
“Finally home. I think the interview went well.”
That is wonderful news, I am so proud of you.
“What should I do tonight to switch off?”
How about a long bath and your favorite movie?
“Kilometre three. My legs feel really heavy.”
Keep those knees up and focus on your breathing right now.
“Okay. Can I push on the last one?”
Give it everything you have left on this final stretch.
“I can't switch off. My head's still at work.”
Let your racing thoughts drift away on your next slow exhale.
“Okay. I'm lying down now.”
Feel the weight of your body sinking deeply into the bed.
“Help me plan Saturday with Maya and Leo. They love climbing.”
Should I book a climbing gym session for you three?
“Leo can only do the afternoon though.”
I will look for climbing slots available this Saturday afternoon.
Conversational AI built for people runs on Inworld. Who are you building for?
Yes, one Inworld AI speech to speech connection listens, reasons and talks, so you don't need separate speech to text, LLM and text to speech vendors. You choose the LLM, and the connection runs over WebSocket or WebRTC. You don't manage the glue code between the parts, and you can swap any layer with a one-field change in the session configuration.
Any model in Inworld Inference works with Inworld AI's speech to speech, including Gemma 4, DeepSeek V4 Flash and GLM 5.3 Flash. You switch models by changing one field in the session configuration, so you can trade reasoning depth for speed from one session to the next. Open models start from $0.07 per million input tokens, for Gemma 4 26B.
Inworld AI's Realtime STT-1 speech to text model returns a voice profile with every final transcript, giving the emotion, age, accent and vocal style it hears and a confidence score for each. In speech to speech, your app reads the voice profile from each transcription event and can pass it to the LLM as context, so the reply can change when someone sounds tired or upset.
Inworld AI's speech to speech API returned first audio as fast as 204 ms after the turn closed and 462 ms at the median. Part of that time is the LLM writing its first words and part is the voice starting to speak, so choosing a faster LLM shortens the wait when latency matters more than reasoning depth. How long the API waits before it closes a turn depends on the turn-detection eagerness you set.
Yes, Inworld AI's speech to speech API stops talking when the user cuts in, because semantic VAD detects when the user starts speaking and cancels the in-flight response. The audio connection stays open, so the model can resume or pivot without closing the session. In one session, the response was cancelled 202 ms after the person started talking.
Yes, Inworld AI's speech to speech API follows the OpenAI Realtime protocol, with the same event names and message flow, so you can move an OpenAI Realtime client over with few changes. You change the endpoint and the key and keep your event handling. Inworld extensions such as voice profiles, backchannels and memory sit in an optional providerData field, so a client that ignores them still works.
There is no separate session fee for Inworld AI's speech to speech, so a conversation costs the sum of its parts at your plan's rates. Realtime TTS-2 costs $12.50 per 1M characters on the Growth plan and $25 on demand. Realtime TTS-2 Flash costs $7 on Growth, $15 on demand and $5 on Enterprise, up to 16× lower than the $80 list price of Eleven v4. Realtime STT-1 costs $0.10 an hour on every paid plan ($0.15 on demand), and LLM tokens bill at the model's Inference rate.
We built Inworld AI speech to speech for consumer apps where people talk to AI out loud, such as companions, language learning, fitness coaching, wellness and social apps. One connection listens, reasons and talks, and each part bills at its own rate with no separate session fee. Talkpal, a language learning app, cut its text to speech costs by 40% with Inworld.
