Get a key
00 / status 912 servers up 275 papers shipped today v4.19.2

Speech & language APIs

Illustration: audio, documents and time inputs flow through a hexagonal processing core and emerge as rows of structured data blocks.
audio → transcript

A plug-and-play toolbox of speech recognition and natural-language models, kept at the state of the art and served as REST endpoints. No GPUs to run, no models to train, no ML team to hire.

01 / one request

Call an endpoint.
Get structured text back.

Every service is a plain REST endpoint behind one account. Point your existing HTTP client at it and the model is somebody else's problem — ours.

Latency p50
41 ms
Uptime
99.98%
Endpoints
140+
prosoftip — example request 
$ curl -X POST <base-url>/nlp/sentiment \
  -H "Authorization: Bearer $PROSOFT_KEY" \
  -d '{"text":"Resolved in two minutes, thank you"}'
{
  "label": "positive",
  "score": 0.97,
  "latency_ms": 41,
  "status": "ok"
}

Illustrative request and response shape. Exact paths and parameters come with your key.

02 / in and outthree illustrative shapes

What goes in. What comes out.

Audio, text or a plain phrase on one side; a transcript, a sentiment, a time range on the other. These are illustrations of the shape, not live calls.

illustrative · not a live call
input · audio
call-0413.wav · 00:02.4 · grammar: billing
output · transcript

I was charged twice on the thirteenth

confidence 0.96
Best-in-class recognition using custom grammars so the model already knows the topic at hand.
03 / catalog6 core services · 1 account

Six services. One account.

Speech in, structured text out. Each service is its own REST endpoint; register once, call the ones you need, and we keep the models behind them at the state of the art.

prosoft.hub
v4.19.2
0 ms
hover a service 
# Move over any service to see the shape of
# its request and response. Each one is a
# plain REST call with your key.
Full service catalog
04 / sota speedupdated daily
14h

median time from
arXiv preprint to
live endpoint

preprint
A paper lands on arXiv with code.
We watch the feeds. If the repository builds, it goes into the pipeline the same day.
image
The code becomes a container image.
Pinned dependencies, reproducible build. No image, no deployment.
eval
It runs against our evaluation sets.
Side by side with the model currently serving the endpoint.
promote
The winner takes the endpoint.
Your path doesn't change. Your results do.
912+
Virtual servers
275
Papers today
1,140
Models since 2024
0
Lines of ML code you write

Why daily.

Benchmarks move every week. If your speech recognizer is six months old, it's already behind. When a paper hits preprint and we can build an image from the code, it runs against our evaluation set the same day and the winner is promoted automatically. Your endpoint path doesn't change. Your results do.

05 / start

One account.
Every service we ship.

Get an API key No minimums · volume discounts available