How do you measure latency, and what happens at the tail?

We measure the moment a caller hears sound. Bytes can arrive in a few milliseconds and open with a second of silence, which scores well on paper and feels slow on a call. Our median is 50 ms, with no silence at the front. The gap between P25 and P75 is 5 ms, the tightest of the five models tested, so a slow turn stays close to the median. Read more here: https://gradium.ai/blog/coval-perceived-ttfa-benchmark.

Which languages are supported, and do they cost extra?

We currently support five languages: English, French, German, Spanish and Portuguese, each with regional variants. All of them are included on every plan at no extra charge.

How do you handle numbers, emails and customer IDs?

Phone numbers, postcodes and email addresses are handled before synthesis, with no preparation on your side. For brand and product names, attach a pronunciation dictionary, with a separate one per client if you need it. A few cases are still hard, German licence plates among them, so test with your own data. Read more here: https://gradium.ai/blog/gradium-tts-latency-and-accuracy.

Is there a limit on how long a session can run?

Five minutes. The model holds context for the whole session, so tone and pacing stay consistent from the first sentence to the last, and you can flush between sentences without losing it. For longer calls, open a second session partway through the first one so the handover is invisible.

Can we run it inside our own infrastructure?

Of course. Three ways: our cloud by default, AWS Marketplace where usage counts toward your committed AWS spend, or your own VPC as a SageMaker image so audio never leaves your account. Teams under HIPAA or a strict internal policy usually take the VPC route. Latency is the same on all three.

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

Text-to-Speech under 50ms
that nails the hard parts

Your agent starts talking before you finish blinking. Emails, phone numbers and acronyms come out right, in all languages, over a single streaming API.

Enter your text ....or

Try the model Read the docs

The numbers that make an agent sound human

Try now See the live benchmark

No credit card required

50 ms to first audio, and none of it is silence

Coval measures when a caller actually hears sound: the first chunk, plus any silence at the start of the stream.

Perceived TTFA · Coval benchmark

TTFA P50

50 ms

Leading silence

0 ms

P25–P75 spread

5 ms

No marketing spin,just code and hard numbers

Paste one of these prompts into coding agents, it'll install the SDK and open the stream.

1Add Gradium streaming text-to-speech to this project.
2
3Read https://docs.gradium.ai/guides/text-to-speech 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 a tts_realtime session with a voice ID and an output format, as an
10   async context manager. Do not open a new socket per utterance.
112. Run the sender and the receiver as concurrent tasks. The sender pushes
12   text as it is produced and calls send_eos when the source is done.
133. In the receiver, play or buffer audio messages as they arrive and stop on
14   end_of_stream. Do not wait for the full text before starting playback.
154. Split outgoing text on whitespace only. Never split a word across two
16   send_text calls and never send punctuation on its own, because the server
17   inserts a space between consecutive messages.
18
19Keep it small: one module, typed, with clear errors when the socket drops
20mid-stream, when the voice ID is unknown, or when the key is missing.
21Do not add a UI unless I ask for one.

Concurrent send/receive for voice agents and LLM streaming.

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.