Skip to main content
HyperWhisper Cloud gives you four featured transcription providers with automatic failover. HyperWhisper also supports bring-your-own-key (BYOK) for direct provider access, and offline local models. HyperWhisper Cloud is built in. You do not need an API key or a separate account. Select a provider for speed, for balance, or for accuracy. All of them are pay-as-you-go with no markup. You pay the rate of the provider. Each provider has an accuracy rating of Highest, High, or Medium. The rating and the cost are independent: two providers with the same rating can have very different costs per minute.
For speed, read the public latency page. It gives the measured p50, p95, and p99 times of each provider in each server region, from the last 90 days of real HyperWhisper Cloud requests. A cell needs a minimum of 3 attempts to show a number. The p99 column needs 500 attempts, because a 99th percentile from a small sample is only its slowest call. A cell below that limit shows a dash. These measurements come from the users who leave Share anonymous speed data on — read General Settings for that setting.
These four providers are the featured providers. If one provider fails, HyperWhisper Cloud sends the request to a backup provider. The full HyperWhisper Cloud picker gives all 12 supported engines with no key. The list includes Meta Muse Voice Transcribe, Soniox, Azure MAI-Transcribe, and Google Gemini 3.5 Transcribe. The Provider row of the picker shows 11 companies, because Gemini 3.5 Transcribe and Gemini are both Google and share one row — read Transcription Modes. Providers outside the featured four do not get automatic failover. The Provider Health page describes the failover chains.

Highest — ElevenLabs Scribe v2

The top-accuracy provider, and the recommended default. It gives the best results on accents, noisy environments, and technical vocabulary.~$0.59 / hour · 9.83 credits/min

High — Grok STT (SpaceXAI)

$0.10 / hour 1.67 credits/minGood multilingual accuracy at a low cost per minute.

Medium — Deepgram Nova-3

~$0.33 / hour 5.5 credits/minHigh English accuracy, low latency, and support for custom vocabulary.

Medium — Groq Whisper Large v3 Turbo

~$0.04 / hour 0.667 credits/minLatency of less than one second. Good for English and the major European languages.
One credit costs $0.001 USD. A $5 purchase creates your account key with 5,000 credits. You can add more credits with the $5 or $10 presets, or with a custom amount up to $500. Card payments have a processing fee of approximately 6%.

​
Read the latency page

The public latency page gives one row for each provider company. It names each row the way the Provider menu in the app names it. Gemini 3.5 Transcribe and Gemini are both Google, so they share one Google row. The buttons above the table select the Metric (p50, p95, p99, or the error rate) and the Clip length group. Hold the pointer on a cell to see the number of attempts behind it. Click Break down by model to open each company into its models. This control is off when the page opens. The models keep the order of the Model menu in the app, and a default badge marks the model that a new selection of that company uses. A model row counts only the attempts that ran on that model. Thus a model row shows a dash more frequently than the row of its company.

​
Use the model chooser

The public model chooser ranks the Cloud models and the on-device models that HyperWhisper ships. The Choose a model item in the site menu opens it. The page is in English only. It is not a general leaderboard: it contains only the models that the app gives you. You have 100 points. Give the points to the four properties that the ranking uses: The five presets — Balanced, Accuracy first, Cheapest, Fastest, and Fully private — set the points for you. You can then move each slider. HyperWhisper adjusts the other three sliders in proportion, thus the total stays at 100 points. Three more controls make the list smaller:
  • Your platform — macOS or Windows. The on-device models are different on each platform.
  • Language — English only, European, or Wide multilingual. A model with no published language count stays in the English list, but the page removes it from the two multilingual lists.
  • Must have — Live streaming, Custom vocabulary, or No preview models.
The page gives the best match first, then the full ranked list. For a Cloud model, the list gives the cost in credits for each minute, the word error rate, and an estimate for each audio minute. For an on-device model, it gives the download size and no cost.
The Closest region menu selects which measurements the ranking uses. The page finds your closest region for you. The measured column gives the median time for one dictation clip of less than 10 seconds, from the last 90 days. This is not a figure for each minute. A cell needs the same minimum of 3 attempts that the latency page needs. If a model is too new to have its own measurements, the page uses the measurements of its provider.

​
Meta Muse Voice Transcribe

Meta Muse Voice Transcribe 1.0 is available through HyperWhisper Cloud or with your own Meta Model API key. The Cloud route costs 3 credits per audio minute, or 0.003perminute∗∗.Thisis∗∗0.003 per minute**. This is **0.18 per audio hour. The direct route uses your Meta account and does not use HyperWhisper credits. For direct use, add the key in Model Library → API Keys, then select Your Provider → Meta → Muse Voice Transcribe 1.0. The app stores the key in the operating system’s secure credential store. Your device sends the audio directly to api.meta.ai; the request does not reach HyperWhisper servers. Meta has extensively evaluated 25 languages. Meta does not publish the closed language-code list in the Model API reference. HyperWhisper therefore does not claim a language code until it can map that code to Meta’s API. The shared catalog decodes this capability record for Muse: These values describe the upstream model. HyperWhisper currently uses Meta’s pre-recorded push-to-talk API. HyperWhisper returns the transcript text, but it does not expose all turn metadata in the app. Meta also provides a real-time Muse API. HyperWhisper does not have a Meta WebSocket relay, so Muse is not available in HyperWhisper Live. The Live streaming filter correctly removes this model. Do not use Meta’s streaming benchmark as a batch WER or speed result. For the primary sources, read Meta’s Muse launch article and speech-to-text API reference.

​
You pay only for speech

HyperWhisper Cloud detects silence and blank audio automatically. If a recording contains no speech, you pay 0 credits. HyperWhisper does not bill for silence at the start of a clip or for pauses between thoughts. HyperWhisper also does not bill for an empty recording that you start by accident. In a typical working day of push-to-talk dictation, you pay only for the minutes that you spoke.

​
Accuracy by language

For English, any of these providers gives good results. ElevenLabs Scribe v2 (Highest) is the most accurate on accents, noisy audio, and technical vocabulary. Deepgram Nova-3 and Groq Whisper Large v3 Turbo share the same Medium accuracy rating — Groq is not less accurate than Deepgram, it is simply cheaper and lower-latency. Grok STT (High) costs a little more than Groq but less than ElevenLabs, and is a few percent less accurate than the top provider.
SpaceXAI does not publish a table of WER values per language for Grok STT. HyperWhisper does not apply older third-party benchmark numbers to this provider, because those numbers measure a different provider.

​
Cost examples

At 1 credit = $0.001 USD, this table gives the cost of each featured provider at typical usage levels. HyperWhisper bills only for speech. Therefore “30 min/day” means 30 minutes of speech, not 30 minutes with the app open. The monthly cost at 30 min/day is ~$9 ElevenLabs / ~$1.50 Grok / ~$5 Deepgram / ~$0.60 Groq. A single $5 top-up gives approximately 8 hours of speech on ElevenLabs Scribe v2, or approximately 50 hours on Grok STT.

​
Alternatives

If you have API credits, or your own free tier (Deepgram $200, AssemblyAI $50), add a key on the API Keys page. You then pay the provider directly at the published rate.HyperWhisper also supports Mistral and Google Gemini for BYOK. The API Keys page shows how to configure them.
When you use your own key, you must remove your audio from model training. Each provider has a different setting in its dashboard. The Data Privacy & Model Training page gives an LLM prompt that you can copy. This prompt finds the current opt-out method for any provider.

​
Increase accuracy with any provider

  • Custom vocabulary — add domain terms such as product names, frameworks, specialist terms, and the names of colleagues. This gives the largest single increase in accuracy for technical or professional use.
  • Low-noise environment — background noise makes every model less accurate. The Best Practices page gives more information.
  • Natural pace — speech that is too fast or too slow also makes accuracy lower.
If you use custom vocabulary with Deepgram Nova-3, set the language and do not use auto. On Nova-3, the keyterm parameter is active only in monolingual mode.

​
FAQ

Does HyperWhisper Cloud add a markup to the cost of the provider? No. You pay the same rate per minute that you pay with a direct API key from the provider. What happens if my provider is temporarily unavailable? HyperWhisper Cloud sends the request to another provider in the chain. The transcription still succeeds. You pay the rate of the provider that did the work. Which provider gives the highest quality? ElevenLabs Scribe v2 has the Highest accuracy rating, and it is the recommended default. It gives the best accuracy on accents, noisy audio, and technical vocabulary. Deepgram Nova-3 costs less and is a good choice for everyday dictation.