Run text-to-speech behind a managed API

Serve open text-to-speech models like Parler or Zonos from an autoscaling endpoint that returns hosted audio URLs. Beam manages the GPUs and only bills while speech is being generated.

input

"So what we're building with Beam is a faster way to ship models."

audio

first audio · 0 ms

Scalable Voice AI Infrastructure

Open models, your endpoint, flat compute pricing — with the serving plumbing already built.

Any open TTS model

Parler and Zonos have step-by-step guides, and anything that runs in PyTorch deploys the same way.

Hosted audio outputs

Save generated audio with Output and return a public URL straight from the endpoint — no storage buckets to wire up.

Warm models per container

on_start loads the model once per container, so generation starts immediately on every request.

Keep-warm for voice apps

Hold containers hot with keep_warm_seconds when response time matters — and scale to zero the rest of the time.

Autoscale with traffic

Beam adds containers as requests grow and removes them when traffic stops — no idle GPUs on your bill.

Per-second pricing

Pay for seconds of GPU generation, not per character or per voice — cost scales with compute, not usage tiers.

Ship a voice API in three steps

01

Define the endpoint

Load your TTS model in on_start and declare the GPU and Python packages inline.

02

Deploy it

beam deploy returns an authenticated HTTPS URL backed by autoscaling GPUs.

03

POST text, get audio

The endpoint returns a hosted URL for the generated file — ready to drop into your app.

terminal

$ beam deploy app.py:speak
=> Deployed 🎉
=> https://parler-tts-abc123.app.beam.cloud

$ curl -X POST https://parler-tts-abc123.app.beam.cloud \
    -H 'Authorization: Bearer YOUR_TOKEN' \
    -d '{"prompt": "Welcome! Your order has shipped."}'

{"url": "https://app.beam.cloud/output/id/3f2a..."}

Frequently asked questions

Which TTS models can I run?

Beam has guides for Parler-TTS and Zonos, and any open text-to-speech model that runs in PyTorch deploys the same way — including your own fine-tuned voices.

How is audio returned?

Save the file with Output and return a hosted public URL, or base64-encode the audio into the JSON response.

How do I keep latency down for real-time apps?

Combine on_start model loading with keep_warm_seconds so containers stay hot between requests. Cold boots on Beam take seconds, not minutes.

What does it cost when nobody is talking?

Nothing. Endpoints scale to zero when idle and billing is per-second while audio is being generated.

Can I batch-generate narration for long documents?

Yes — run the same function as a task queue or fan it out with .map() to synthesize many segments in parallel.

Start shipping on infra you won’t outgrow.

Run sandboxes and GPU workloads on your cloud, and scale out to ours when you need to. No infra to manage.