Sauti / self-hosted
Whisper small (self-hosted)
Open weights, fast lane
WER
12.3%
95% CI 10.7%–14.0%
Rank
#7
of 12 measured
Latency
44.9s
mean, per clip
Cost
$0.00
per audio hour
Every other model on the board is a statistical tie with this one. Its confidence interval overlaps all 11 of them, so the rank above tells you where it landed, not that it is better or worse than anything else here.
What it is
The small Whisper checkpoint on the same self-hosted bridge, used as the fast lane where latency matters more than the last point of accuracy.
Pick it when
When you are running your own inference and want throughput. The gap to large-v3 was well under half a point of WER with fully overlapping intervals, for roughly a fifth of the wall-clock time on the same hardware.
Think twice when
When you want the best open-weight accuracy available and have the compute to spare, or when you need a model with published per-language figures to plan against.
Assessment written against the board of 2026-08-08. The figures above are read live from the current published snapshot (2026-08-08), so if those dates differ, trust the figures.
What the vendor publishes
OpenAI (Whisper paper) publishes 3.4% on LibriSpeech; we measured 12.3% on People’s Speech. Those are different corpora, so this is a difference in test set as much as in the model. It is comparable only when the benchmark matches. Source (read 2026-08-15).Whisper paper Table 8 (English transcription WER, greedy decoding): LibriSpeech test-clean. This is the MULTILINGUAL `small` we run, not `small.en` (3.1) — and LibriSpeech is clean read audiobooks, so it is not comparable to our conversational People's Speech number.
How this was measured
250 clips (2.9 hours) of unscripted conversational English from People’s Speech (MLCommons, CC-BY), human-transcribed. Scored on a held-out split (N=77) never used to tune anything, through one fixed text normalizer, with bootstrap 95% confidence intervals. One corpus, one language, one modality: batch pre-recorded. It does not tell you how this model handles your accents, your domain vocabulary, or streaming.
Cost is self-hosted’s own published pre-recorded pay-as-you-go list price, read from their pricing page on 2026-08-08. Streaming and committed-volume rates differ.Marginal cost on hardware we already run. Amortised capex and power are excluded.