Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
Orukeet streaming benchmarks
Measured September 16, 2026 through https://orukeet.gizmovoice.ai, backed by three warm NVIDIA L4 GPUs in Oregon, the Netherlands and Singapore. Each GPU is in a GCP g2-standard-8 instance. Model: orukeet-v0.1.0.
Stop to final latency
Ten seconds of mono PCM16 audio at 16 kHz, uploaded during simulated real-time capture. Timing begins immediately before commit after the last audio frame, and ends on the full transcript. Recording duration, session/token setup, cleanup and paste are excluded. The endpoint does not return interim words.
| Client | Median | p95 | Recordings |
|---|---|---|---|
| California Mac | 92.3 ms | 144.8 ms | 15 |
| Oregon cloud VM | 37.4 ms | 88.7 ms | 15 |
| Netherlands cloud VM | 28.5 ms | 31.2 ms | 15 |
| Singapore cloud VM | 23.2 ms | 26.8 ms | 15 |
Median inference was approximately 15.3 ms; median server time was approximately 17 ms. Each cloud region used three five-recording runs, while the Mac used one fifteen-recording session. Colocated cloud clients are not representative of residential users in those countries.
Three fresh token-authenticated sessions using the PR adapter separately measured 78.9, 84.9 and 83.6 ms commit-to-final, 151–359 ms for token issuance, and 211–274 ms to reach WebSocket readiness. The live adapter receipt retains each run and its transcript hash. Start connection setup during capture and buffer initial samples. Very short recordings can still pay setup time.
Concurrent recording clients
32 paced connections per region overlapped for a measured fleet peak of 96. Three runs completed 864 recordings with no failed utterances, transcript mismatches or capacity retries.
| Region | Median | p95 |
|---|---|---|
| Oregon | 34.7 ms | 89.9 ms |
| Netherlands | 28.5 ms | 30.4 ms |
| Singapore | 22.9 ms | 25.0 ms |
The configured socket ceiling is not a demonstrated user count. These results establish 96 concurrent recording connections under this workload, not the maximum fleet capacity.
Saturated streaming throughput
Audio was sent without capture pacing over 16 reused WebSockets per GPU. Rates count finals inside fixed 20-second measurement windows and report the median of three runs per region.
| Region | Ten-second utterances per second | p95 final latency |
|---|---|---|
| Oregon | 86.80 | 370.8 ms |
| Netherlands | 91.15 | 544.0 ms |
| Singapore | 90.45 | 511.3 ms |
Including responses drained after the windows, these runs completed 16,237 utterances without a failed utterance or mismatch. They generated 10,888 expected capacity responses; clients honored backoff and retried commit without reupload. The longest final was about 2.58 seconds. These are saturation rates, not a low-latency operating target or user count. Summing the medians is not an independently measured fleet throughput.
An earlier fixed-workload test used eight sockets and 100 utterances each, repeated three times per region. It measured 63.4/s in Oregon, 92.5/s in the Netherlands and 93.1/s in Singapore. Slower Cloudflare paths in Oregon left slow sockets draining after faster ones completed. Both methods are retained in the benchmark receipt.
Oregon inference stayed near 15 ms while Cloudflare Seattle connections had roughly 17–19 ms application-ping round trips and Portland connections had roughly 70–71 ms. The slower run is retained in the p95. Before/after routes were measured at different times and should not be interpreted as a causal speed comparison.
Reliability and coverage
The native latency and load campaigns completed 24,361 matching finals without failed utterances. This count includes expected saturation backpressure and is separate from the following checks:
- 189 live API/authentication checks passed, including TLS, missing/invalid auth, token expiry, cross-region single-use consumption and replay rejection.
- 11 recording-boundary checks passed at 31, 61 and 600 seconds and above the limit. Above-limit audio was rejected with WebSocket close code 1009.
- During a controlled European readiness drain, the first alternate new socket appeared at 17.3 seconds; three consecutive alternate sockets were confirmed by 25.3 seconds. An existing socket with half a recording buffered survived and completed correctly. All three backends recovered. A hardware failure can still interrupt an active socket.
- The country proxy audit reached 18 countries with 20 verified route pairs and 80 matching transcripts. The initial Australian exit failed on both routes; three fresh exits passed. These timings include California → Decodo residential exit → API → return and are not native-country latency estimates.
The fixture tests serving performance and transcript parity, not general recognition accuracy. Measurements do not establish a worldwide SLA or native hotkey-to-paste latency. Signed desktop acceptance and private account/usage integration are separate rollout requirements.
Reproduce the adapter latency check
Use Node 24 and a mono 16 kHz PCM16 WAV file. Inject ORUKEET_SERVICE_KEY through your operator secret manager; never commit it. Run:
node scripts/benchmark-orukeet.cjs /absolute/path/to/audio.wav
The optional second argument is a local JSON file with a text field for exact transcript comparison. SMOKE_RUNS defaults to three and accepts one through ten. Output includes the fixture hash, region, setup time, inference/server/queue time, commit-to-final latency and a transcript hash; it excludes transcript text and credentials. This operator tool combines backend token issuance and desktop transport for testing. It does not substitute for testing the private customer login/allowance route.