Latency

Speech-to-text latency, measured

Every transcription we run times the provider that answered it. These are those timings — 51,521 provider attempts over the last 90 days, grouped by the region that made the call.

Metric

Clip length

Detail

Row order

Pick the region closest to you

#1 overall
Groq
295ms
#2 overall
Mistral
646ms
#3 overall
ElevenLabs
719ms


264	234	291	328	295	221	362	300	521	322	147	—	339	292	37	134	295
521	362	391	693	646	498	967	—	905	—	401	889	—	1037	617	1269	636
820	909	719	304	548	803	750	534	1861	575	603	—	364	929	547	825	446
760	789	909	456	457	571	748	—	1001	—	493	509	—	835	294	1244	393
777	694	773	585	551	728	1439	—	1844	—	808	2571	—	1834	475	1498	563
585	850	724	1068	1242	636	1509	—	1974	—	512	—	—	889	374	1577	845
992	1072	948	452	542	1016	1118	478	2498	413	908	1288	418	1895	463	1193	513
1271	845	1296	1433	968	929	1174	—	1370	—	699	1664	—	839	1262	926	877
2733	2678	2594	1506	1951	2457	2781	1336	3700	1939	2620	4259	1894	4352	2316	5406	1966
2651	2849	2772	2163	1823	2606	2374	1869	3916	2156	3053	3014	1683	2736	2071	3181	1665
3499	2802	3158	1721	2423	3440	3226	—	5746	—	2666	2860	—	4889	2729	4235	2583

Milliseconds, provider call only. A cell needs at least 3 attempts to show a number. Click a region to sort by it, or a provider to pin its row.

Method

How this is measured

What the number is: The time one provider took to answer one call, measured at the edge machine that made it — everything that attempt spent, including handing the audio to the provider and waiting on a long-running job. Your upload to us, authentication, and our own credit checks are excluded. So is the network between you and us — this is the provider's time, not your round trip.
Failed attempts count: A provider that times out still spent your time, so its attempt is in the latency numbers. When one provider fails we fall back to another, and both attempts are recorded separately. The error-rate metric shows how often that happens. What is never counted is a request we turn away ourselves before calling anyone — audio past a provider's size cap, or in a format it does not accept: that provider never received the call, so it is not charged for it here.
A row is a provider, until you open it: Most of these providers offer several models, and HyperWhisper lets you pin the one you want. A provider row covers whichever model of that provider actually ran — its default for most people, plus whatever anyone else picked — so read it as “how fast this provider answers us”. Break down by model splits each one into its models. Those rows count only the attempts that ran on that model, so they fill in more slowly and a young model may show nothing but dashes for a while.
The names match the app: Provider and model names here are the same ones the app's Provider and Model menus show, read from the same shared catalog. Google covers both Chirp and Gemini in one row, exactly as the app does — the model you pick under it is what decides which of the two runs.
Why clip length matters: Longer audio takes longer to transcribe, so comparing a provider handed 5-second clips against one handed 5-minute files would say nothing. Every cell compares clips of similar length, grouped by an estimate taken from the audio's size and format — the same estimate for every provider, so no cell is flattered by a provider that measures its own audio differently.
What it covers: The batch transcription endpoint. Live streaming transcription is not measured here.
What we store: A timing row carries the provider, the model, the region, which clip-length group the audio fell in, which attempt in the chain it was, and whether the call worked. It carries no account, no key, no request id, and no audio or text. The clip's own length is not kept, and the timestamp is stored only to the hour, so nothing ties a row to a person or reassembles one transcription.
Who is counted: Only the HyperWhisper apps, from the versions that carry the setting below — macOS 2.43.0 and Windows 1.10.0 onward. Sharing is on by default, and a default is only fair if you could have changed it, so builds that shipped before the setting existed are not measured at all. Neither is anything calling the API directly.
How to opt out: In the HyperWhisper app, turn off Share anonymous speed data under Settings → General. Your transcriptions work exactly the same — the same providers, the same speed, the same result — we just stop keeping the timing. Local models never send a timing at all.