Will the transcript change after it appears?

No. Once a word appears, it stays. Many speech-to-text systems go back and rewrite earlier words as they hear more, which makes live captions jump around. We chose stability over squeezing out a little extra accuracy, because that is what live transcripts and dictation need.

Do we need a separate model to know when someone stops talking?

No, it is built in. The same model that transcribes also tells you when the speaker has finished, so there is only one thing to run. It listens to how the sentence is built, not just to silence, so a quick "uh-huh" or "mmm" is not mistaken for an interruption. Switch it on in your session settings: it is off by default.

Can it tell speakers apart, and transcribe recordings?

Telling speakers apart is in beta and works for two people today, which covers the usual agent-and-caller call. Support for more is on the way. You can also send pre-recorded audio through the same API, though the model is designed first for live conversations. Confidence scores are in development. If you need them sooner, we can prioritize them as part of an enterprise plan.

Spotlight post

Gradium Extends Funding to $100 Million and Expands to Silicon Valley

Gradium extends its seed funding to $100 million, welcoming new investors including NVIDIA, and opens a San Francisco Bay Area office to scale its real-time voice AI.

July 8, 20263 min read

Gradium Extends Funding to $100 Million and Expands to Silicon Valley

Models

Text-to-SpeechTurn text into natural sounding voice Speech-to-TextTranscribe audio into text instantly Live TranslationConvert voice to voice in real time Voice CloningInstance or pro voice clone Voice DesignA brand-new voice from a text prompt On-Device TTSOffline real-time natural voice

Showcase

Agent DemoTest our voice agent live GradbotBuild voice agents in a single prompt CommunityBuilders, researchers and creators App GalleryApps built on Gradium

Speech-to-textbuilt for live conversation.

Streaming transcription that knows when your user is done talking, and learns your vocabulary. Semantic turn detection, keyword boosting and tunable latency over one WebSocket.

Your transcription will appear here...

Tap to record

Try the model Read the docs

The numbers that make an agent listen like a human

every 80ms
Knows when you've finished
Listens to your pauses and your words, so the agent replies right when you're done
0.56s to 3.84s
Speed you can tune
Shorter for a fast live agent, longer for the most accurate transcript
up to 500 words
Custom vocabulary
Add product names, people and jargon. No retraining, and each word's odds can rise ~20x

Try now Talk to us

No credit card required

A pause is not the end of a turn

A pause to think sounds exactly like the end of a sentence, so systems that only listen for silence cut people off. Ours listens to what is being said, so your agent waits for people to finish, then answers without an awkward gap.

Semantic VAD

Do you like any other… uh… other animals or anything?

I gotta figure out this… uh… packing tape.

No marketing spin,just code and hard numbers

Paste one of these prompts into your coding agent and it will read the guide, install the SDK and open the stream. Nothing else to do, you're ready to go.

1Add Gradium real-time speech-to-text to this project.
2
3Read https://docs.gradium.ai/guides/speech-to-text first and follow it over
4anything you assume. Use the official Gradium SDK for this codebase's
5language. Read the API key from an environment variable, never hardcode it.
6
7Then implement the flow, in this order:
8
91. Open an stt_realtime context manager with the input format and a
10   json_config holding the language and delay_in_frames. Use 16 frames unless
11   I ask for something more reactive.
122. Push 80 ms PCM frames from the audio source in a producer task and call
13   send_eos when it ends.
143. Iterate the session in a consumer task. Accumulate text messages into the
15   current turn, and close the turn when a step message shows the inactivity
16   probability crossing 0.5 on the 1 s, 2 s or 3 s horizon. Make the horizon
17   and the threshold configurable.
184. When the app decides a turn is over before VAD does, call send_flush with
19   an ID and wait for the matching flushed message before treating the
20   transcript as final.
21
22Keep it small: one module, typed, with clear errors when frames are the wrong
23size, when the socket drops, or when the key is missing. Do not add a UI
24unless I ask for one.

Push-based, for microphones and telephony

No distortion under load

Latency stays flat as concurrency climbs, flexible deployment that fits your setup.

Real-time API

Bidirectional WebSocket streaming. Start sending before ready returns; first audio in ~200 ms. Telephony formats built in.

in · pcm 24 kHz

CLIENT/AGENT

WEBSOCKET

Gradium API

wss://api.gradium.ai/
api/speech/s2s

BidirectionalTelephony ulaw / alaw~200 ms

out · pcm 48 kHz

VOICE ENGINE

Copied!

Predictable & secure

Flat latency under load, enterprise SLAs, zero data retention, ISO 27001 certified.

1 conversationQuiet hour

I’d like to move my appointment.

50ms

Of course. Which day works?

10 000 conversationsPeak

I’d like to move my appointment.

50ms

Of course. Which day works?

Runs where you need it

Cloud API, dedicated instances, self-hosted, on-prem for data sovereignty, and cloud-provider marketplaces.