Realtime STT-1 Speech to text
Returns what the user said and how they sounded, in the same event.
up to 13×sooner final transcript than Google
4voice traits with each transcript
up to 10×lower price per hour than OpenAI
Waiting for the bus
DeepSeek V4 Flash
Picking a film tonight
A birthday voice note
Venting about leftovers
A poem about first snow
Dinner ideas in Spanish
Qwen3.8 Flash
A song for the beach
Would you rather
Three good things
Company on the way home
A gift idea in Portuguese
A joke in French
Companions
and social apps
Build voice companions and social apps that hear how someone feels and answer in real time.
Every conversation card is a real session, with replies from 247 ms after the turn closed (median 329 ms).
Dinner ideas in Spanish
A joke in French
Waiting for the bus
Picking a film tonight
A birthday voice note
Venting about leftovers
A song for the beach
Would you rather
Three good things
Company on the way home
A gift idea in Portuguese
DeepSeek V4 Flash
Qwen3.8 Flash
Try it now
Open the playground ↗Ask your agent to add Inworld.
Paste it into your agent
What a companion needs from voice and LLM models
Speech to speech docs ↗Status users spend 1 h 36 min a day in the app. At that scale, every reply has to be fast, natural and cheap.
247 msfrom the end of a turn to the reply
6¢for a 10-minute conversation
3 sof audio to give a companion its own voice
500+languages and dialects
Each dot is one turn, timed from the moment the API closed the turn to the first audio of the reply. The fastest turn took 247 ms and the median 329 ms. In the median turn, the first token of the reply came 122 ms after the transcript and the first audio 204 ms after that. The time the API waits before it closes a turn depends on the turn-detection eagerness, which was medium here, and is not counted.
One session, three turns
Turn 1
“Hey, I'm back. I finally asked my manager for help with the launch.”
I'm so glad you finally did that. How did the conversation actually go?
Turn 2
“She said yes, which I really didn't expect.”
That is such a huge relief. Does this mean you'll finally be able to breathe a little easier tonight?
Turn 3
“Okay. What's one thing I can do tonight, just for me?”
Maybe you could just pick a favorite movie or a book and truly unplug from everything. You've definitely earned a little peace and quiet.
The second and third replies were written with the earlier turns in the prompt. Memory is one of speech to speech's extensions to the OpenAI Realtime protocol, alongside back-channel responses and fillers.
“Pick a song. All four of us can sing on the way to the beach.”
Spoken cheerful and loud, with friends laughing
emotion happy 81%style singing 88%accent en-GB 34%age adult 79%
“The bus is running late again. Tell me something good.”
Spoken tired, late at night
emotion sad 76%style normal 51%accent en-US 68%age young 75%
“Help me wish my best friend happy birthday in a voice note.”
Spoken excited and warm
emotion happy 59%style normal 82%accent en-GB 84%age young 63%
Realtime TTS-2 spoke each line with the direction shown. Each label is STT-1's top reading for that trait, with its confidence.
Spanish · voice Sarah
“Oye, ¿me ayudas a decidir qué cocinar esta noche?”
¡Claro que sí, me encantaría! ¿Tienes ganas de algo ligero o prefieres algo más reconfortante?
French · voice Olivia
“Salut ! Raconte-moi une petite blague pour bien commencer la journée.”
Coucou ! Pourquoi les plongeurs plongent-ils toujours en arrière et jamais en avant ? Parce que sinon, ils tombent dans le bateau ! Passe une super journée.
Portuguese · voice Olivia
“Oi, me ajuda a escolher um presente para minha mãe.”
Claro, eu adoraria te ajudar com isso! Me conta um pouquinho sobre o que ela gosta de fazer para a gente encontrar algo especial.
Each session is one speech to speech connection with a Realtime TTS-2 library voice, and only the instructions set the language. The library has 275 voices, and you can clone a voice from 3 seconds of audio or design one from a text description.
| Part | What it bills | On-Demand | Growth |
|---|---|---|---|
| Realtime STT-1 | 10 minutes of audio in | 2.5¢ | 1.7¢ |
| Gemma 4 26B on Inference | 40 turns, 3,500 tokens in and 200 out each | 1.3¢ | 1.3¢ |
| Realtime TTS-2 | 5 minutes of reply speech, about 4,649 characters | 11.6¢ | 5.8¢ |
| Conversation, with TTS-2 | 15.4¢ | 8.7¢ | |
| Conversation, with TTS-2 Flash | 10.7¢ | 6.2¢ |
STT-1 costs $0.15 an hour on demand and $0.10 on paid plans, TTS-2 costs $25 and $12.50 per 1M characters, TTS-2 Flash costs $15 and $7, and Gemma 4 26B costs $0.07 per 1M input tokens and $0.34 per 1M output tokens. Growth list prices make the same conversation 1.8× cheaper than on demand (15.4¢ → 8.7¢ with TTS-2). The example assumes the person and the agent each talk for half of the 10 minutes, that each turn sends 3,500 prompt tokens, and that reply speech runs at 15.5 characters a second. Cached prompt tokens would lower the LLM line.
Four parts, one session
Pricing ↗Use each part on its own, or all together.
Speech to speech listens, reasons and talks on one connection.
Inference LLMs
Route free and paid users to different models by the plan in each request.
104msto first token, fastest turn
up to 90%cache hits on repeated prompts
$0.07per 1M input tokens, Gemma 4 26B
Realtime TTS-2 Text to speech
Gives each companion its own voice, cloned from 3 seconds of audio or designed from a text prompt.
up to 3×sooner first audio than ElevenLabs
500+languages and dialects
up to 16×lower price than ElevenLabs
Speech to speech One connection
One session handles turn detection and barge-in, so users can trail off and cut in.
247msto first audio, fastest turn
6.2¢per 10 minutes, TTS-2 Flash on Growth
1connection per session
Use cases
Speech to speech ↗Real recorded sessions.
Press play to hear the person, the agent, and how fast each reply started.
A voice companion people can talk to at any hour, about their day or whatever is on their mind.
0:41
Person
“Hey, I'm back. I finally asked my manager for help with the launch.”
Agent
I'm so glad you finally did that. How did the conversation actually go?
Person
“She said yes, which I really didn't expect.”
Agent
That is such a huge relief. Does this mean you'll finally be able to breathe a little easier tonight?
Person
“Okay. What's one thing I can do tonight, just for me?”
Agent
Maybe you could just pick a favorite movie or a book and truly unplug from everything. You've definitely earned a little peace and quiet.
The person's lines are generated with text to speech.
Assistants inside social apps that help friends make plans together.
0:29
Person
“Help me plan a birthday picnic for Sam this Saturday. There are six of us.”
Agent
That sounds like a blast! Should I start looking for some nice park locations for Saturday?
Person
“Two of them are vegetarian and. Joe can't get there before two.”
Agent
Got it, I'll make sure we find plenty of veggie options and aim for a start time after two. Do you want me to send out a poll to see what everyone's bringing?
The person's lines are generated with text to speech.
The person and the companion make up a story together, taking turns line by line.
0:54
Person
“Can we make up a story together? It starts with a lighthouse keeper who finds a letter.”
Agent
Elias had lived alone on the Crag Rock lighthouse for twenty years, preferring the company of the sea to the noise of the town. One stormy evening, while sweeping the lantern room, he found a waterlogged envelope tucked behind a loose floorboard. What does the letter say, or who do you think it was meant for?
Person
“The letter is from her younger self. What does? It say.”
Agent
The ink had bled into ghostly blue smudges, but the words were still legible: "If you are reading this, it means you finally decided to stay." Below the message was a date from fifty years ago, a day Elias had tried desperately to forget. What does the letter say next, or how does Elias react to seeing his own handwriting?
The person's lines are generated with text to speech.
Guided conversations that follow instructions you write, such as a reflection before a hard talk.
0:33
Person
“Can you give me a short reflection to start the day? I have a hard talk with my brother later.”
Agent
Remember that your intention is to be heard and to understand, not just to be right. Approach this conversation with a soft heart and steady breath.
Person
“I want to listen more than I talk this time.”
Agent
That is a beautiful intention. By holding space for his words, you are offering him the gift of being truly seen.
The person's lines are generated with text to speech.
A friendly voice to talk to every morning, for people who live alone or just like the company.
0:22
Person
“Good morning. I've just come back from my walk around the park.”
Agent
Good morning! It sounds like a lovely way to start the day, how was the weather out there?
Person
“My granddaughter is visiting on Sunday, and I want to bake something.”
Agent
Oh, how wonderful! Do you have a special treat in mind that she loves?
The person's lines are generated with text to speech.
For teams building companions and social apps
Case studies →
AI costs down 20× at more than 500,000 daily users, who spend 1 h 36 min a day in the app
“When you work with external parties you never know what you're going to get, but Inworld has been really hands-on and fast. It's honestly been the most positive and best experience I've ever had working with another company on a problem like this.”
Fai Nur, CEO, Wishroll (Status)
10× lower TTS cost, with all five core engagement metrics flat or better
“Inworld reduced our TTS costs by about 10x, while remaining neutral or better on all the metrics we care about, which is pretty incredible.”
Creston Brooks, Co-founder and CTO, Luvu
Start building
Companions built for people run on Inworld. Who are you building for?
Questions
What voice AI should I use for an AI companion app?
For an AI companion app, Inworld AI's speech to speech runs the whole conversation on one connection that listens, reasons and talks, and it hears how the user sounds as well as what they say. Each companion can have its own voice. A 10-minute companion conversation costs 6.2¢ with Realtime TTS-2 Flash at Growth list prices.
Can each companion have a distinct voice?
Yes, each companion built on Inworld AI's text to speech can have its own voice, picked from 275 library voices, cloned from 3 seconds of audio or designed from a text description. The Growth plan holds up to 30,000 voices of your own, and cross-lingual cloning keeps a companion's voice the same across supported languages.
Can my companion detect how a user is feeling?
Yes, Inworld AI's Realtime STT-1 speech to text model returns a voice profile with every final transcript, giving the emotion, vocal style, age and accent it hears and a confidence score for each. Pass those signals to your LLM as context and have it adjust the tone and pacing of the reply, for example speaking more softly when someone sounds sad.
How fast does an Inworld AI companion reply?
A companion running on Inworld AI's speech to speech API started its reply as fast as 247 ms after the user's turn closed, with a median of 329 ms across 21 turns. Each time runs from the end of the turn to the first sound of the reply. One speech to speech session handles listening, semantic turn detection, interruptions and the spoken reply.
How much does an AI companion conversation cost?
A 10-minute companion conversation on Inworld AI costs 6.2¢ with Realtime TTS-2 Flash at Growth list prices, so a dollar covers about 16 of them, or 8.7¢ with Realtime TTS-2. Each part bills at its own rate. Realtime STT-1 costs $0.10 an hour on paid plans, Realtime TTS-2 costs $12.50 per 1M characters on Growth, half the $25 on-demand rate, and the LLM bills at its Inference price. The example assumes each side talks for five minutes over 40 turns, with no cache discount.
Can free and paid users get different models?
Yes, Inworld AI's Inference API routes each request by metadata such as plan, region or mode, so free users can stay on a fast, lower-priced model while paying users get a larger one. You can also A/B test models and prompts with weighted variants, keeping each user on one variant, and measure retention. Status by Wishroll cut its AI costs by more than 95% with Inworld.
Can conversations stay private?
Yes, conversations on Inworld AI can run with zero data retention, available on Enterprise plans and as an add-on on the Growth plan, so audio, transcripts and replies are processed in memory and redacted from our logs as soon as processing completes. Enterprise plans add EU and India data residency. Inworld AI is SOC 2 Type II certified and GDPR compliant.
Which languages can an AI companion speak?
Inworld AI's Realtime TTS-2 text to speech model speaks about 500 languages and dialects, and Realtime STT-1 transcribes 30 and detects the spoken language automatically. Cross-lingual synthesis keeps the companion's voice the same when it switches language, even inside a single reply, so people can talk with a companion in their own language.