Benchmarks

Benchmarks and methodology

Our own measurements aren't published yet. They come after the pilot, measured on real jobs with the method below. Until then, this page shows public third-party figures that our design is based on, each with its source. None of them are our results.

Methodology (what we will publish)

  1. Workload. A fixed, public test set: long-form English and multilingual podcasts and Spaces (10 to 90 minutes), plus a short-clip set under 30 seconds. Embeddings: a public forum corpus.
  2. Accuracy. Word error rate against human reference transcripts, after standard text normalisation. Embeddings: retrieval recall@10 against the reference model run on our own hardware.
  3. Throughput. Audio hours per wall-clock hour for a batch of 100 files, measured end to end from submission to the last result.
  4. Cost. What the buyer pays per audio hour at list price, next to sourced competitor list prices on the same date.
  5. Distribution, not best case. We report p50 and p95 and include failed and re-run jobs.

Reference figures from public sources

Third-party figures cited in our feasibility research. Not our measurements.
FindingSource
Whisper large-v3 on 100 consumer-GPU replicas: nearly 8,000 audio hours in about 10 hours for roughly $110 all-inSalad
Clips under 30 seconds yielded about 4x less audio per dollar than long files on the same GPUsSalad
RTX 4090, faster-whisper large-v3: about 14x real time single-stream, 175–250x batchedRunPod
CPU-only: a Ryzen 9 5950X needed 8.1 s to encode a 30 s window; an i5-8250U laptop needed 61.5 s (slower than real time)whisper.cpp #89
Phones lose 15–45% of peak speed within minutes of sustained loadarXiv 2603.23640

Market prices on the comparison date

Planned First write-up: after the pilot. Current prices.

Benchmarks · Vouch Compute