SuperWhisper Review
Local-first dictation with cloud AI modes
macOS Windows Local
+ Cloud
SuperWhisper is a voice dictation app for macOS and Windows offering both local (offline) and cloud AI transcription. Priced at $9/month or $299 lifetime. This independent review covers WER/CER accuracy across 6 test recordings, a privacy analysis, and a UX verdict.
SuperWhisper Verdict
Powerful local engine buried under a broken cloud flagship
SuperWhisper version 1.4.0 scores 6.6/10 overall in VoiceTools independent testing (tested 2026-05-30). Best local model (Whisper Standard) achieves 2.4% aggregate WER across 6 recordings. Best cloud model (Ultra) achieves 2.6% aggregate WER.
Works well for
- Whisper Standard (hidden): best-in-class local accuracy at 2.4% WER
- Genuine offline mode — no audio leaves device in local configuration
- Lifetime license option at $299 — no subscription required
Watch out for
- S1-Voice (cloud flagship default) shows 15-37% WER across test recordings
- Trial is 15 minutes total — blocks even offline models after limit
- Cloud mode sends app name, clipboard, and focused text to Modal.com beyond audio
Best for
- Power users willing to dig past default settings and switch to Whisper Standard
Not for
- Anyone who installs and expects the default model to work well
SuperWhisper Accuracy & Speed
| Model | Accuracy | Speed | ||
|---|---|---|---|---|
| English | Local | Parakeet Default 1.1 GB CPU Tested on CPU Ryzen AI 9 HX · 32 GB RAM
NVIDIA Parakeet TDT 0.6B — default local model in SuperWhisper 1.4. Fast, good on clean speech, but has no ITN: numbers and dates come out as spoken words. Default local model — users see this first | 12.4%
WER Word Error Rate
What % of words the model got wrong. 0% = every word correct.
8.1% CER Character Error Rate
Same as WER but measured letter-by-letter. Usually lower than WER.
22% PER Punctuation Error Rate
How accurately the model placed commas, periods, and other punctuation.
4 / 10 |
~3s
2–5s range
Post-stop latency
Seconds from pressing Stop to the final text appearing in your active app. Average across all test recordings.
8 / 10 |
| Whisper Standard Best accuracy hidden in UI
500 MB CPU Tested on CPU Ryzen AI 9 HX · 32 GB RAM
OpenAI Whisper large-v2 running locally. Best accuracy of all tested models but hidden from the main model picker — requires Library search to find. Hidden — Settings > Library > search "Whisper Standard" | 2.4%
WER Word Error Rate
What % of words the model got wrong. 0% = every word correct.
1.6% CER Character Error Rate
Same as WER but measured letter-by-letter. Usually lower than WER.
18% PER Punctuation Error Rate
How accurately the model placed commas, periods, and other punctuation.
8 / 10 |
~8s
6–12s range
Post-stop latency
Seconds from pressing Stop to the final text appearing in your active app. Average across all test recordings.
5 / 10 | ||
| Cloud | S1-Voice Default SuperWhisper's proprietary cloud model, presented as the headline AI feature. Applies aggressive rewriting that causes large content losses on some recordings. Default cloud model — users land here without changing settings | 22.4%
WER Word Error Rate
What % of words the model got wrong. 0% = every word correct.
14.2% CER Character Error Rate
Same as WER but measured letter-by-letter. Usually lower than WER.
38% PER Punctuation Error Rate
How accurately the model placed commas, periods, and other punctuation.
3 / 10 |
~2s
1–4s range
Post-stop latency
Seconds from pressing Stop to the final text appearing in your active app. Average across all test recordings.
9 / 10 | |
| Ultra Best Cloud Cloud Whisper-class model offered as "Ultra" tier. More conservative post-processing than S1-Voice — accurate and stable across all recording types. Better than S1-Voice on every recording — less prominently surfaced in UI | 2.6%
WER Word Error Rate
What % of words the model got wrong. 0% = every word correct.
1.8% CER Character Error Rate
Same as WER but measured letter-by-letter. Usually lower than WER.
20% PER Punctuation Error Rate
How accurately the model placed commas, periods, and other punctuation.
8 / 10 |
~3s
2–5s range
Post-stop latency
Seconds from pressing Stop to the final text appearing in your active app. Average across all test recordings.
8 / 10 | ||
| No models match — turn a filter back on. | ||||
SuperWhisper for Coding & IT Recommended: Whisper Standard Ultra
Ultra Cloud
Coding 7.4% 17 err / 230w
Conference 0.5% 1 err / 220w
Coding
- snake_case identifiers mostly preserved
- CLI flags correct
- No hallucinations
- "Tauri" → "Tory"
- "tokio::runtime" → "Toko runtime"
alternative wrong extra missing
okay ,so .the deploy pipeline is broken again .um ,let me walk through what we have .the repo is a rust workspace with a atari front-end and an axm back-end .cargo.tomel at the root .two members ,desktop and server .the postgres migration that landed last tuesday added a new column to the users table called um ,last underscore scene underscore at type timestamp with timezone default now diesel picked it up fine but sqlx in the server crate started failing the macro expansion at compile time because uh ,the offline query metadata json was stale run cargo sqlx prepare workspace before pushing ok ,now the kubernetes side the helm chart for the staging cluster bumped the image tag to ,um ,v1.4.0 alpha 3 ,but the deployment yaml still references v1.3.11 .that mismatch is what is throwing the image pullback off .update values.yaml under environments slash staging .the github actions workflow is using ubuntu 2204 runners ,but rust-toolchain.toml is pinned to 1.83 ,and the runner stable is 1.79 ,so either bump the runner image or pin rustup install to 1.83 explicitly .for the cloudflare worker ,wrangler.toml has the wrong account underscore id after we migrated the team ,and miniflare is failing the d1 binding lookup .run wrangler login again .finally ,explain analyze on the sessions token lookup .p95 should stay under 15 milliseconds .if it regresses ,the missing index on the ,um ,jsonb metadata column is the prime suspect .
Conference
- Best on accented English speaker — 0.45% WER
- "Kubernetes", "PostgreSQL", "microservices" exact
- Zero dropped sentences
- "load balancer" → "load balance" (once)
alternative wrong extra missing
so ,the question i want to address today is whether on-device speech recognition is finally good enough to replace the cloud .the short answer is yes ,for english at least ,and the longer answer is the interesting part .when openai released whisper in ,um ,2022 ,the smallest models were already competitive with cloud apis for clean audio ,but they were slow .three years later ,two things have changed .first ,the c++ port from georgi gurganov ,called whisper.cpp ,runs the same model significantly faster than the python reference .and ,ah ,it runs on a raspberry pi .second ,nvidia released parakeet tdt in 2025 ,which is between 5 and 10 times faster than whisper on cpu and has similar accuracy for english .so now ,the question is not ,can we run speech recognition locally ?but ,um ,should we and what do we give up ?the honest answer is that you give up the ability to update the model continuously ,which cloud providers can do every week ,and you give up some accuracy on the long tail of less common languages .what you gain is ,you know ,privacy by architecture rather than by policy ,lower latency for short phrases on modern hardware and zero per minute cost at scale .for most desktop dictation use cases in 2026 ,that trade-off is the right one .
Whisper Standard Local
Coding 7.3% 17 err / 233w
Conference 0.9% 2 err / 220w
Coding
- snake_case identifiers intact
- CLI flags like "--release" correct
- No hallucinations
- "Tauri" → "Tarry"
- "axum" → "axm"
alternative wrong extra missing
ok ,so the deploy pipeline is broken again .um ,let me walk through what we have .the repo is a rust workspace with a tarry front end and an axum back end .cargo.toml at the root .two members ,desktop and server .the postgres migration that landed last tuesday added a new column to the user's table called ,um ,last_scene_ at type_timestamp with timezone default now .diesel picked it up fine ,but sqlx and the server crate started failing the macro expansion at compile time ,because ,uh ,the offline query metadata ,json was stale .run cargo sqlx ,prepare workspace before pushing .ok ,now the kubernetes side .the helm chart for the staging cluster bumped the image tag to ,um ,v1.4.0 alpha 3 ,but the deployment yaml still references v1.3.11 .that mismatch is what is throwing the image pullback off .update values .yaml under environments / staging .the github actions workflow is using ubuntu 2204 runners ,but rusttoolchain.toml is pinned to 1.83 ,and ,uh ,the runner stable is 1.79 ,so either bump the runner image or pin rustup install to 1.83 explicitly .for the cloudflare worker ,wrangler.toml has the wrong account underscore id after we migrated the team ,and miniflare is failing the d1 binding lookup .run wrangler login again .finally ,explain analyze on the session's token lookup .p95 should stay under 15 milliseconds .if it regresses ,the missing index on the um ,jsonb metadata column is the prime suspect .
Conference
- Strong on accented English
- "PostgreSQL" and "Kubernetes" exact
- Only 1 minor substitution across 89s
- "schema" → "schema" (case flip once)
alternative wrong extra missing
so ,the question i want to address today is whether on-device speech recognition is finally good enough to replace the cloud .the short answer is yes ,for english at least ,and the longer answer is the interesting part .when openai released whisper in ,um ,2022 ,the smallest models were already competitive with cloud apis for clean audio ,but they were slow .three years later ,two things have changed .first ,the c++ ,port from georgi gurganov ,called whisper.cpp ,runs the same model significantly faster than the python reference ,and ,uh ,it runs on a raspberry pi .second ,nvidia released parakeet ttt in 2025 ,which is between 5 and 10 times faster than whisper on cpu and has similar accuracy for english .so now ,the question is not ,can we run speech recognition locally ,but ,um ,should we ,and what do we give up ?the honest answer is that you give up the ability to update the model continuously ,which cloud providers can do every week ,and you give up some accuracy on the long tail of less common languages .what you gain is ,you know ,privacy by architecture ,rather than by policy ,lower latency for short phrases on modern hardware ,and zero per minute cost at scale .for most desktop dictation use cases in 2026 ,that trade-off is the right one ,
Parakeet Local
Coding 16.7% 39 err / 234w
Conference 1.8% 4 err / 222w
Coding
- Handles prose segments cleanly
- No hallucinations
- "cargo.toml" → "Cargo .toml"
- "tokio::spawn" → "tokio spawn"
- "impl Trait" → "imp Trait"
alternative wrong extra missing
okay ,so the deploy pipeline is broken again .um ,let me walk through what we have .the repo is a rust workspace ,with a atari front end and uh an axum backend .cargo.toml at the cargo .two members ,desktop and server .the postgres migration that landed last tuesday added a new column to the user's table called um last underscore scene ,underscore at type timestamp with time zone ,default now .diesel picked it up fine ,but sqlx in the server crate started failing the macro expansion at compile time ,because uh the offline query metadata ,json was stale .run cargo sqlx prepare workspace before pushing .okay ,now the kubernetes side .the helm chart for the staging cluster bumped the image tag to um v1 .4 .0 alpha 3 ,but the deployment yaml still references v1 .3 .11 .that mismatch is what is throwing the image pullback off .update values .yaml under environments slash staging .the github actions workflow is using ubuntu 2204 runners ,but rust toolchain .toml is pinned to 1 .83 ,and uh the runner stable is 1 .79 ,so either bump the runner image ,or pin rustup install to 1 .83 explicitly .for the cloudflare worker ,wrangler .toml has the wrong account underscore id after we migrated the team ,and miniflare is failing the d1 binding lookup ,run wrangler login again .finally ,explain analyze on the sessions token lookup ,p95 should stay under 15 milliseconds .if it regresses ,the missing index on the um jsonb metadata column is the prime suspect .
Conference
- Clean transcription of accented English speaker
- Technical terms like "API" and "SDK" correct
- "Kubernetes" → "Cubernetes"
alternative wrong extra missing
so ,the question i want to address today is whether on device speech recognition is finally good enough to replace the cloud .the short answer is yes ,for english at least ,and the longer answer is the interesting part .when openai released whisper in um 2022 ,the smallest models were already competitive with cloud apis for clean audio ,but they were slow .three years later ,two things have changed .first ,the c port from georgi gurganov called whisper .cpp runs the same model significantly faster than the python reference .and uh it runs on a raspberry pi .second ,nvidia released parakeet tdt in 2025 ,which is between five and ten times faster than whisper on cpu and has similar accuracy for english .so now ,the question is not can we run speech recognition locally ,but um should we ,and what do we give up ?the honest answer is that you give up the ability to update the model continuously ,which cloud providers can do every week ,and you give up some accuracy on the long tail of less common languages .what you gain is ,you know ,privacy by architecture rather than by policy ,lower latency for short phrases on modern hardware ,and zero per minute cost at scale .for most desktop dictation use cases in 2026 that trade off is the right one .
S1-Voice Cloud
Coding 35.9% 85 err / 232w
Conference 24.6% 56 err / 220w
Coding
- Faster response than local models (~2s)
- "impl AsyncRead" → "RUMP_INSTAL"
- Dropped entire code block (lines 14–17)
- "CloudFace" hallucinated (not in source)
alternative wrong extra missing
okay so .the deploy pipeline is broken again .um let me walk through what we have :the repo is a rust workspace with a atari front end and an axm backend .cargo dot tomel at the root ;two members desktop and server .the postgres migration that landed last tuesday added a new column to the users table called um last_seen underscore at type timestamp with time zone .default now .diesel picked it up fine but sqlx in the server crate started failing the macro expansion at compile time because ,uh .the offline query metadata json was stale .run cargo sqlx prepare workspace before pushing .the github actions workflow is using ubuntu 220.4 runners but rusttoolchain.toml is pinned to 1.83 and the runner stable is 1.79 so either bump the runner image or pin rustup install to 1 point 83 explicitly .for the cloudface ,the rump_instal is a bit of a runaway .the github actions rupp_instr is using ubuntu 22.04 a little bit more than one day .and the runner stable is has been pinned to the runner image .the rustup is pinned down to 1 0.83 explicitly .as a flare worker ,wrangler.toml has the wrong account_id after we migrated the team ,and miniflare is failing the d1 binding lookup .run wranglar login again .finally ,explain analyze on the sessions token lookup :p95 should stay under 15 milliseconds .if it regresses ,the missing index on the ,um ,jsonb metadata column is the prime suspect .
Conference
- Fast turnaround even on longer clip
- "distributed systems" → "destructive systems"
- Dropped 3 full sentences mid-recording
- Non-deterministic: 37% WER run 1 vs 16% run 2 — same audio
alternative wrong extra missing
so .the question i want to address today is whether on-device speech recognition is finally good enough to replace the cloud ?the short answer is yes - for english at least .and the longer answer is the interesting part .when openai released whisper in .um 2022 ,the smallest models were already competitive with cloud apis for clean audio but they were slow .three years later two things have changed :first the c++ port from georgi gerganov called whisper.cpp runs the same model significantly faster than the python reference and it runs on a raspberry pi second nvidia released parakeet tdt so now ,which is between five and ten times faster than whisper on cpu and has similar accuracy for english so now the question is not can we run speech recognition locally? but um .should we ?and what do we give up? the honest answer is that you give up the ability to update the model continuously which cloud providers can do every week .and you give give-up some accuracy on the long tail of less common languages .what you gain is ,you know ,privacy by architecture rather than by policy ,lower latency for short phrases on modern hardware and zero per minute cost at scale .for most desktop dictation use cases in 2026 that trade off is the right one
SuperWhisper for Everyday & Long-form Recommended: Whisper Standard Ultra
Whisper Standard Local
Casual 0.0% 2 err / 190w
Long-form 2.3% 15 err / 525w
Casual
- Perfect: zero word errors
- Disfluencies handled cleanly
alternative wrong extra missing
hey ,so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are three separate problems and we have been treating them as one .first ,the email verification step has a 48% drop-off rate ,which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .second ,the time zone selection screen is buried inside settings and people who skip it end up with notifications at 3 in the morning ,which is ,you know ,the worst possible experience .we need to bring that into the main flow .third ,the welcome email goes out two hours after sign up ,which feels kind of random .it should go out either immediately while the user is still excited or ,or i mean ,strategically the next morning ,but not at this awkward in-between moment .so my proposal is ,in order of impact ,fix the email verification first ,then move the time zone picker ,then retime the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .
Long-form
- Best long-form accuracy tested — no drift over 3:42
- Zero hallucinations across full recording
- "whisper.cpp" partially garbled near end
- Occasional sentence-boundary missed
alternative wrong extra missing
okay ,so i want to talk for a few minutes about why we ended up building our own voice dictation app instead of just using one of the existing ones .this is not going to be a sales pitch .i want to talk about the specific moments where the existing tools broke for me because i think they will sound familiar if you have tried to dictate seriously for any length of time .the first thing that happened was that apple dictation timed out .i was writing a long email ,maybe 200 words ,and somewhere around 60 seconds in ,the microphone icon just disappeared .no warning ,no progress indicator ,just gone .i had to manually press the keyboard shortcut again ,which broke my train of thought and also did not paste the previous content together cleanly .i tried to find a setting to extend the time out .there is no such setting .multiple apple support threads confirm this is by design .for short voice messages ,fine , for drafting any real piece of writing ,useless .the second thing was whisper flow ,which is ,excellent at the actual transcription , but uploads every recording to their servers .i noticed this because my company has a policy against sending internal product discussions to third party cloud services ,and i had to stop using whisper flow inside one specific app because of that .i am not blaming whisper flow for the policy . i am saying that , you know , for the specific job of brainstorming client work into a voice memo ,an app that requires network upload was the wrong tool .and in 2025 , they had a documented incident where they were also capturing screenshots periodically .they have since changed that , but it set a precedent .the third thing was that whatever i tried on mac never had a windows equivalent .i work across both .mac whisper is excellent .there is no mac whisper for windows .super whisper had mac for years before it shipped a windows version .whatever workflow i built up on the mac side , um , disappeared every time i switched to my windows desktop ,which was at least once a week .i wanted feature parity across both platforms , and there was nothing offering that as a primary product goal .so we started snail text with ,three commitments .one ,the audio never leaves the device .we use whisper dot cpp on the cpu and the gpu .we used parakeet tdt as an optional second engine .neither one calls home .two ,no time out .hold the hot key . talk for 10 minutes if you want .and three ,mac and windows are first class on day one ,not mac first with windows as an afterthought .what i did not expect was , um , how much subtle engineering work was actually involved in making the gpu acceleration work reliably across the heterogeneity of consumer windows laptops .i will save that story for another time .the short version is that what looks like a single line of configuration in the whisper dot cpp documentation turns into ,you know ,a couple of weeks of edge cases when you actually ship to thousands of users .but that is a different conversation .
Ultra Cloud
Casual 0.0% 0 err / 190w
Long-form 2.5% 14 err / 521w
Casual
- Perfect: zero errors on casual speech
- Natural disfluency handling
alternative wrong extra missing
hey ,so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are three separate problems and we have been treating them as one .first ,the email verification step has a 48% drop-off rate ,which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .second the time zone selection screen is buried inside settings and people who skip it end up with notifications at three in the morning which is you know the worst possible experience we need to bring that into the main flow third the welcome email goes out two hours after sign up which feels kind of random it should go out either immediately while the user is still excited ,or i mean strategically the next morning ,but not at this awkward in-between moment .so my proposal is ,in order of impact ,fix the email verification first ,then move the time zone picker ,then retime the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .
Long-form
- No drift — consistent quality start to finish
- Zero hallucinations across 3:42
- App name spacing inconsistent once
- "whisper.cpp" spacing once
alternative wrong extra missing
okay ,so i want to talk for a few minutes about why we ended up building our own voice dictation app instead of just using one of the existing ones .this is not going to be a sales pitch .i want to talk about the specific moments where the existing tools broke for me because i think they will sound familiar if you have tried to dictate seriously for any length of time .they were lo- the first thing that happened was that apple dictation timed out i was writing a long email ,maybe 200 words and somewhere around 60 seconds in the microphone icon just disappeared no warning ,no progress indicator just gone i had to manually press the keyboard shortcut again which broke my train of thought and also did not paste the previous content together cleanly .i tried to find a setting to extend the time out .there is no such setting .multiple apple support threads confirm this is by design .for short voice messages ,fine .for drafting any real piece of writing ,useless .the second thing was wispr whisperflow ,which is excellent at the actual transcription but uploads every recording to their servers .i noticed this because my company has a policy against sending internal product discussions to third-party cloud services ,and i had to stop using wispr whisperflow inside one specific app because of that .i am not blaming wispr whisperflow for the policy . i am saying that , you know , for the specific job of brainstorming client work into a voice memo ,an app that requires network upload was the wrong tool .and in 2025 , they had a documented incident where they were also capturing screenshots periodically .they have since changed that .but it set a precedent .the third thing was that whatever i tried on mac never had a windows equivalent .i work across both .mac whisper is excellent .there is no mac whisper for windows .super whisper had mac for years before it shipped to windows version .whatever workflow i built up on the mac side disappeared every time i switched to my windows desktop which was at least once a week .i wanted feature parity across both platforms and there was nothing offering that as a primary product goal .so we started snail text with three commitments .one ,the audio never leaves the device .we use whisper.cpp on the cpu and the gpu we use parakeet tdt as an optional second engine .neither one calls home .two ,no timeout .hold the hotkey . talk for 10 minutes if you want .and three ,mac and windows are first class on day one . not mac first with windows as an afterthought .what i did not expect was , um , how much subtle engineering work was actually involved in making the gpu acceleration work reliably across the heterogeneity of consumer windows laptops .i will save that story for another time .the short version is that what looks like a single line of configuration in the whisper.cpp documentation turns into , you know , a couple of weeks of edge cases when you actually ship to thousands of users .but that is a different conversation .
Parakeet Local
Casual 0.0% 0 err / 197w
Long-form 3.6% 19 err / 528w
Casual
- Perfect: zero word errors on casual speech
- Disfluencies (um, uh) preserved naturally
alternative wrong extra missing
hey ,so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are um three separate problems ,and we have been treating them as one .first ,the email verification step has a 48 % drop off rate ,which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .second ,the time zone selection screen is uh buried inside settings ,and people who skip it end up with notifications at three in the morning ,which is ,you know ,the worst possible experience .we need to bring that into the main flow .third ,the welcome email goes out two hours after sign up ,which feels um kind of random .it should go out either immediately while the user is still excited ,or i mean strategically the next morning ,but not at this awkward in between moment .so my proposal is in order of impact ,fix the email verification first ,then move the time zone picker ,then re time the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .
Long-form
- No quality drift over 3+ minutes
- Consistent pace throughout — no mid-recording degradation
- "whisper.cpp" split across sentences once
- Occasional dropped filler mid-paragraph
alternative wrong extra missing
okay .so i want to talk for a few minutes about um why we ended up building our own voice dictation app instead of just using one of the existing ones .this is not going to be a sales pitch .i want to talk about the specific moments where the existing tools broke for me ,because i think they will sound familiar if you have tried to dictate seriously for any length of time .they were the first thing that happened was that apple dictation timed out .i was writing a long email ,maybe two hundred words ,and somewhere around sixty seconds in ,the microphone icon just uh disappeared ,no warning ,no progress indicator ,just gone .i had to manually press the keyboard shortcut again ,which broke my train of thought ,and also did not paste the previous content together cleanly .i tried to find a setting to extend the timeout .there is no such setting .multiple apple support threads confirm this is by design .for short voice messages ,fine , for drafting any real piece of writing ,useless .the second thing was wispr whisperflow ,which is um excellent at the actual transcription , but uploads every recording to their servers .i noticed this because my company has a policy against sending internal product discussions to third party cloud services ,and i had to stop using wispr whisperflow inside one specific app because of that .i am not blaming wispr whisperflow for the policy . i am saying that , you know , for the specific job of brainstorming client work into a voice memo ,an app that requires network upload was the wrong tool .and in 2025 , they had a documented incident where they were also capturing screenshots periodically .they have since changed that .but it set a precedent .the third thing was that whatever i tried on mac never had a windows equivalent .i work across both .macwisper is excellent .there is no mac whisper for windows super whisper had mac for years before it shipped to windows version .whatever workflow i built up on the mac side , um disappeared every time i switched to my windows desktop ,which was at least once a week .i wanted feature parity across both platforms , and there was nothing offering that as a primary product goal .so we started snail text with uh three commitments .one ,the audio never leaves the device .we use whisper .cpp on the cpu and the gpu .we use parakeet t d t as an optional second engine .neither one calls home .two ,no timeout .hold the hotkey ,talk for 10 minutes if you want .and three ,mac and windows are first class on day one ,not mac first with windows as an afterthought .what i did not expect was um how much subtle engineering work was actually involved in making the gpu acceleration work reliably across the heterogeneity of consumer windows laptops .i will save that story for another time .the short version is that what looks like a single line of configuration in the whisper .cp documentation turns into you know ,a couple of weeks of edge cases when you actually ship to thousands of users .but that is a different conversation .
S1-Voice Cloud
Casual 37.0% 71 err / 192w
Long-form 14.4% 80 err / 520w
Casual
- Lost 2 of 3 sections — only opening paragraph survived
- Heavy rewriting distorts meaning of what remains
alternative wrong extra missing
hey !so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are three separate problems and we have been treating them as one :first ,the email verification step has a 48% drop-off rate which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .and the time zone selection screen is buried inside settings .and people who skip it end up with notifications at 3:00 in the morning which is you know the worst possible experience .we need to bring that into the main flow third the welcome email goes out two hours after sign up which feels um kind of random .it should go out either immediately while the user is still excited or i mean strategically the next morning but not at this awkward in-between moment so my proposal is in order of impact to fix the email verification first then move the time zone picker ,then re-time the welcome email what do you think let me know if you want to grab 15 minutes tomorrow to talk through it .
Long-form
- Manages to complete the full recording without timeout
- Large sections rewritten — 14% WER from aggressive post-processing
- Brand names and app identifiers mangled or dropped in later paragraphs
alternative wrong extra missing
okay so i want to talk for a few minutes about why we ended up building our own voice dictation app instead of just using one of the existing ones .this is not going to be a sales pitch .i want to talk about the specific moments where the existing tools broke for me because i think they will sound familiar if you have tried to dictate seriously for any length of time .they were lo- the first thing that happened was that apple dictation timed out for me .i was writing a long email maybe two hundred words ,and somewhere around 60 seconds in ,the microphone icon just disappeared .no warning ,no progress indicator - just gone .i had to manually press the keyboard shortcut again ,which broke my train of thought and also did not paste the previous content together cleanly .i tried to find a setting to extend the time out .there is no such setting .multiple apple support threads confirm this is by design .for short voice messages : fine .for drafting any real piece of writing : useless .the second thing was wispr whisperflow which is excellent at the actual transcription but uploads every recording to their servers .i noticed this because my company has a policy against sending internal product discussions to third-party cloud services and i had to stop using wispr whisperflow inside one specific app because of that i am not blaming wispr whisperflow for the policy . i am saying that , you know , for the specific job of brainstorming client work into a voice memo an app that requires network upload was the wrong tool .and in 2025 they had a documented incident where they were also capturing screenshots periodically .they have since changed that but it set a precedent .the third thing was that whatever i tried on mac never had a windows equivalent .i work across both .mac whisper is excellent ! there is no mac whis per for windows .super whisper had mac for years before it shipped to windows version .whatever workflow i built up on the mac side disappeared every time i switched to my windows desktop which was at least once a week .i wanted feature parity across both platforms and there was nothing offering that as a primary product goal .so we started voicenote snailtext with three commitments : 1 the audio never leaves the device .we use whisper dot cpp on the cpu and the gpu .we used parakeet tdt as an optional second engine .neither one calls home .2 no time out .hold the hotkey . talk for 10 minutes if you want .3 mac and windows are first class on day one not mac first with windows as an fine .what i did not expect was how much subtle engineering work was actually involved in making the gpu acceleration work reliably across the heterogeneity of consumer windows laptops i will save that story for another time .the short version is that what looks like a single line of configuration in the ip documentation turns into a couple of weeks of edge cases when you actually ship to thousands of users .but that is a different conversation .
SuperWhisper for Numbers & Structured Data Recommended: Whisper Standard S1-Voice
Whisper Standard Local
Numbers/ITN 2.5% 3 err / 40w
Numbers/ITN
- Dates and currency nearly exact
- "$12,400.75" and phone number correct
- "March 15th, 2026" → "March 15, 2026" (minor format)
alternative wrong extra missing
okay ,the meeting is on march 15 ,2026 at 3.30 p.m .pacific time .the proposed budget is $12,400.75 which is about 1.6% under the original estimate .phone number for the conference bridge is 1-800-555-0123 extension 479 .order id is abc-1 23456 .
S1-Voice Cloud
Numbers/ITN 7.3% 3 err / 41w
Numbers/ITN
- "$12,400.75" exact
- Phone number format correct
- "March 15th, 2026" → "March 15, 2026"
- "ABC-123456" → "ABC 123456" (hyphen dropped)
alternative wrong extra missing
okay .the meeting is on march 15th ,2026 at 3:30pm pacific time .the proposed budget is $12,400.75 which is about 1.6% under the original estimate .phone number for the conference bridge is 1800 555 0123 extension 479 .order id is abc-123456 .
Ultra Cloud
Numbers/ITN 10.0% 2 err / 39w
Numbers/ITN
- Phone number and date format correct
- "$12,400.75" → "$12400.75" (comma dropped)
- "Order ID" label partially dropped
alternative wrong extra missing
okay ,the meeting is on march 15th ,2026 at 3:30 3.30pm pst .the proposed budget is $12,400.75 which is about 1.6% under the original estimate .phone number for the conference bridge is 1-800-555-0123 extension 479 .order id is abc-123456 .
Parakeet Local
Numbers/ITN 86.0% 37 err / 43w
Numbers/ITN
- No ITN — numbers output as spoken words throughout
- "$12,400.75" → "twelve thousand four hundred dollars and seventy five cents"
- Phone number and order ID completely garbled
alternative wrong extra missing
okay .the meeting is on march fifteenth ,twenty six ,at three thirty p .m .pacific time .the proposed budget is twelve thousand four hundred dollars and seventy five cents ,which is about one point six percent under the original estimate .phone number for the conference bridge is one eight hundred five zero one two three extension four seven nine .order id is abc one two three four five six .
SuperWhisper: Noise Resistance Recommended: Parakeet Ultra
Parakeet Local
Noisy Cafe 0.0% 0 err / 195w
Noisy Cafe
- Noise has zero effect — identical output to clean version
- Café background at SNR 5 dB not detected
alternative wrong extra missing
hey ,so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are um three separate problems and we have been treating them as one .first ,the email verification step has a 48 % drop off rate ,which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .second ,the time zone selection screen is uh buried inside settings ,and people who skip it end up with notifications at three in the morning ,which is you know the worst possible experience .we need to bring that into the main flow .third ,the welcome email goes out two hours after sign up ,which feels um kind of random .it should go out either immediately while the user is still excited ,or i mean strategically the next morning ,but not at this awkward in-between moment .so my proposal is in order of impact ,fix the email verification first ,then move the time zone picker ,then re-time the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .
Whisper Standard Local
Noisy Cafe 0.0% 0 err / 190w
Noisy Cafe
- Noise has zero effect — identical output to clean version
alternative wrong extra missing
hey ,so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are three separate problems and we have been treating them as one .first ,the email verification step has a 48% drop-off rate ,which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .second ,the time zone selection screen is buried inside settings and people who skip it end up with notifications at 3 in the morning ,which is ,you know ,the worst possible experience .we need to bring that into the main flow .third ,the welcome email goes out two hours after sign up ,which feels kind of random .it should go out either immediately while the user is still excited ,or i mean strategically the next morning ,but not at this awkward in-between moment .so my proposal is in order of impact ,fix the email verification first ,then move the time zone picker ,then retime the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .
Ultra Cloud
Noisy Cafe 1.1% 4 err / 190w
Noisy Cafe
- Near-perfect under café noise — only 1 minor substitution
- "in-between" split once
alternative wrong extra missing
hey ,so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are three separate problems and we have been treating them as one .first ,the email verification step has a 48% drop-off rate ,which means almost half of people who sign up never actually finish .that alone is worth looking at before we touch anything else .second ,the time zone selection screen is buried inside settings and people who skip it end up with notifications at 3 in the morning which is you know the worst possible experience .we need to bring that into the main flow .third ,the welcome email goes out two hours after sign up which feels kind of random .it should go out either immediately while the user is still excited or ,or i mean ,strategically the next morning .but not at this awkward in-between moment .so my proposal is ,in order of impact ,fix the email verification first ,then move the time zone picker ,then re-time the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .thank you .
S1-Voice Cloud
Noisy Cafe 4.8% 7 err / 188w
Noisy Cafe
- Handles café noise better than casual clean — rewriting helps here
- "in-between" → "in between"
- Some filler words not stripped
alternative wrong extra missing
hey !so i wanted to follow up on what we discussed yesterday about the onboarding flow .i think there are three separate problems and we have been treating them as one .first ,the email verification step has a 48% drop-off rate which means almost half of people who sign up never actually finish that alone is worth looking at before we touch anything else .second the time zone selection screen is buried inside settings .and people who skip it end up with notifications at 3:00 in the morning which is you know the worst possible experience .we need to bring that into the main flow .third ,the welcome email goes out two hours after sign-up which feels kind of random .it should go out either immediately while the user is still excited or i mean strategically i'm not going to go early the next morning but not at this awkward in-between moment .so my proposal is :in order of impact fix the email verification first then move the timezone picker then retime the welcome email .what do you think ?let me know if you want to grab 15 minutes tomorrow to talk through it .
SuperWhisper UX & Integration
Getting started & flow
Onboarding flow
Reached first successful dictation in about a minute — nothing superfluous.
Hotkey customization
Default shortcut is comfortable and remappable, no system conflicts — but the push-to-talk option does not actually work.
Error messages
Shows a center-screen message when the trial runs out, but there is no fallback — and settings navigation is scattered across sections.
Recording experience
Recording overlay UX
Clear recording pill / overlay — recording state is obvious.
Stop / cancel UX
Easy to cancel a bad dictation; cancel hotkey included.
Text insertion reliability
Pastes reliably into every app tested.
Auto-insert vs clipboard
Auto-inserts the text and can restore your previous clipboard afterwards.
Managing your work
Recording history
Browsable history with search; you can open a recording to see its mode, duration and even the prompt used. No export.
Mode / model switching
Fast switching by hotkey and from the pill UI.
Idle resource use
~160 MB RAM · 0.3% CPU at rest (cloud).
SuperWhisper Features
Text processing
AI post-processing
Cloud AI modes rewrite text — many models, BYOK for several providers. But S1-Voice over-rewrites and drops content.
Custom vocabulary / dictionary
Per-word replacements applied at transcription.
Text snippets / expansion
Bundled into the custom-dictionary feature, not a separate snippets UI — and also doable via LLM post-processing instructions.
Output & extras
File transcription
Hidden behind the tray icon. Broken on LLM modes (returns a stale buffer); only works on the Voice mode, and the UX is so confusing it barely counts.
Export (txt / srt / json)
No txt / srt / json export, and history cannot be bulk-exported.
Music auto-mute
Pause, lower, or fully mute media while recording.
Voice commands
Local recognition
Offline / local inference
Genuine offline mode — no audio leaves device in local configuration.
Multiple model options
Parakeet, Whisper Standard (hidden), S1-Voice, Ultra — but best local model is buried.
SuperWhisper Privacy
SuperWhisper keeps audio on-device when using local models. Cloud models upload audio to modal.com only after you press Stop.
Audio uploaded in cloud mode only
Endpoints: modal.com, api.superwhisper.com
Audio sent only after you press Stop
Nothing is uploaded until you confirm by pressing Stop. Cancel before then and the audio never leaves.
Sends more than audio
In cloud mode: active app name, focused element text, clipboard contents, computer name, locale, timezone
Opt out of training on your data
Your recordings are not used to train models.
Disable analytics & tracking
You can turn off product analytics and telemetry.
Turn off history storage
You can set the app to never store your transcription history.
From the privacy policy not scored
- Privacy policy guarantees data is never used to train AI models and is not retained on SuperWhisper servers — all storage is local.
- States it collects no usage data and uses no cookies or tracking technologies.
- Note: the observed cloud mode still sends app context and clipboard to Modal.com — stronger than the policy implies, so cloud users should not assume "local-only".
Pricing
Monthly $9/mo
Lifetime $299 one-time
Trial limit 15 minutes total (all models)
Account required Optional
Methodology
Accuracy scores use WER (Word Error Rate) computed against multi-reference ground truth
with {a|b} alternates for valid transcription variants (e.g. 48% and
forty-eight percent are both accepted). Audio delivered via virtual cable from
ElevenLabs TTS. Single test session on 2026-05-30.