Batch transcription & embeddings · OpenAI-compatible API

Batch inference at a fraction of the price, on a network you can check.

Vouch Compute runs Whisper transcription and open embedding models on a network of vetted consumer GPUs and Macs. Every unit of work is counted on our servers and verified before anyone is paid. It's built for public and non-sensitive data.

For buyers

Transcribe and embed for less

Whisper large-v3-turbo, against $0.16–0.20 at Salad and $0.36 at OpenAI. Change your base URL, keep your OpenAI SDK.

For contributors

Put your idle GPU or Mac to work

Paid in USDC per verified job, never for uptime. Realistic ranges: a few dollars a month on a CPU desktop, tens of dollars on an Apple Silicon Mac, more on an RTX card when there is demand.

Distributed processing. Jobs run on a network of vetted consumer GPUs and Macs owned by independent contributors. Anyone running a job can see its input. Do not send personal or sensitive data. Read the data policy.

The network last week

Network statistics is temporarily unavailable. We don't show guessed numbers while a source is down.

Counted on the server

Audio seconds and tokens are measured by our metering service from the file we send, never from what a device reports. Buyers are billed and contributors are paid from the same ledger.

Verified before anyone is paid

Signed device benchmarks, hidden known-answer jobs, re-runs of a small sample on trusted hardware, and holds before release. Failed checks go to review, not straight to forfeit.

Honest about the trade-off

Consumer hardware can't keep inputs private from its owner. So we only take public and non-sensitive data, and we say so on every page. How it works.

Only the base URL changes

from openai import OpenAI

client = OpenAI(base_url="https://api.vouchcompute.com/v1", api_key="sk-pool-…")

job = client.audio.transcriptions.create(model="whisper-large-v3-turbo", file=open("space.mp3", "rb"))
emb = client.embeddings.create(model="bge-m3", input=["gm", "wagmi"])

Long files are cheaper through the async batch endpoint: short clips under 30 seconds yield about 4x less audio per dollar on the same GPUs. Batch API.