Atlas TTS · Production ready
Atlas TTS is natural, expressive text to speech behind an OpenAI-compatible API. Run it as a managed endpoint, or inside your own cloud.
Why Atlas TTS
Premium neural voices that hold up in live dialogue and long-form narration. Score them blind against incumbents, on your scripts.
Ten languages plus dialects, one model: English, German, French, Russian, Portuguese, Spanish, Italian, Chinese, Japanese, Korean.
Instant zero-shot cloning and named preset voices, ready to carry your brand.
Sub-120ms time to first audio at production load, fast enough that your voice agents never leave a silence.
Start free on your own scripts
Your scripts, your judges, brands stripped.
The exact content you ship today. Any of our ten languages.
Atlas TTS and your current vendor, side by side, unlabeled.
Blind, within 48 hours. If Atlas doesn't win, the conversation ends there.
Integration · OpenAI-compatible
Atlas TTS speaks the OpenAI API. Keep your SDK, keep your code, swap the endpoint.
from openai import OpenAI
client = OpenAI()
audio = client.audio.speech.create(
model="tts-1-hd",
voice="alloy",
input="Your appointment is at 9am.",
)
from openai import OpenAI
client = OpenAI(
base_url="https://api.runatlas.com/v1")
audio = client.audio.speech.create(
model="atlas-tts",
voice="capella", # or your cloned voice
input="Your appointment is at 9am.",
)
Official OpenAI SDKs (Python, Node) · WebSocket streaming · Same API on cloud and self-hosted
Capabilities
Where it fits
Deployment
Access Atlas TTS as a fully managed, OpenAI-compatible endpoint. We handle hosting, scaling, and billing.
Start in minutes, scale to production.
Text, voices, and audio never leave your infrastructure.
Atlas engineers deploy and support it, with a production SLA.
Specifications & performance
| Model | Atlas TTS, 1.7B-parameter neural TTS |
| Languages | English, German, French, Russian, Portuguese, Spanish, Italian, Chinese, Japanese, Korean + dialects |
| Voices | Named preset voices · zero-shot cloning (~3s reference) |
| Audio out | WAV / PCM · 24 kHz · 16-bit · mono |
| Latency | <120ms TTFB p50 at rated load · 38ms single stream |
| Streaming | Yes, sentence-by-sentence, for real-time agents · WebSocket: audio framed per sentence |
| Response | Audio bytes · content-type: audio/wav |
| SDK | Official OpenAI SDKs (Python, Node) |
| Access | Managed cloud or your VPC |
Performance · Throughput per GPU
| GPU | Rated load | Speed / stream | TTFB p50 | Aggregate |
|---|---|---|---|---|
| H100 | 8 streams | 3.5× RT | 115 ms | ~30× real-time |
| B300 | 10-12 streams | 4-5× RT | 79 ms | ~46× real-time |
Single stream: 38ms to first audio. Full concurrency sweep available on request.
Getting started & support
Provision a managed endpoint and call the OpenAI-compatible API in minutes.
The Atlas team builds and tunes brand voices for your use cases.
Dedicated support, pronunciation tuning, and an SLA for production.
One engine, two products
The model on its own: a managed endpoint or weights in your VPC, for teams that bring their own call logic.
Start free on your scripts →Live agents built on that voice, resolving the call inside your systems: CRM, billing, ticketing.
Explore Atlas Voice →Atlas Voice includes the model. The model also stands alone.
Next step