Features
Everything in the free app.
And what we measured.
Dictation and meetings share one app, one download and one set of models. Each has its own page with every feature in detail.
Dictation ›
- Push to talk or toggle, your own hotkey
- Waveform overlay and chimes
- Voice commands, English and German
- Paste the last text again
- Polish, with presets
- Auto cleanup
- Correction dictionary and learn from edits
- Review and retry
Meetings ›
- Record a call without a bot
- Numbered speakers
- Dictation filtered out of the transcript
- File upload with speaker detection
- Export as TXT, SRT and VTT
- Play back or delete, audio storage optional
More in the app
Dictate during calls
Mic guard (off by default) mutes other apps' microphones while you hold the hotkey. Browser calls go silent as soon as you hold the hotkey (checked with Chrome calls). Discord's desktop app ignores it; bind its Push to Mute to the same keys.
Speak German. Type English.
Translate before typing through a translator server that you run yourself. If it is on another computer, your text goes to it. The transcript is still saved as spoken.
Your own server
Point NexusVox at any server that speaks the OpenAI-compatible HTTP or the realtime protocol. Audio leaves the PC only if that server runs elsewhere. The same goes for the language-model and translator servers.
Your GPU.
Or none at all.
Two models, switchable in the dashboard, plus a slot for your own server if you already run one.
Fastest
NVIDIA GPU
Parakeet or Whisper Large V3 Turbo, in Docker.
- 8 s of speech
- 0.08 s
- Word error rate
- 5.8 % EN · 5.2 % DE
- Video memory
- ~2 GB
No GPU needed
CPU only
Parakeet or Whisper Large V3 Turbo, in-process. No Docker.
- 8 s of speech
- 0.34 s · 16 threads
- Word error rate
- 6.9 % EN · 6.8 % DE
- Minimum
- 4 threads, 4 GB free RAM
- 25 European languages with Parakeet, detected automatically
- 99 languages with Whisper Large V3 Turbo
Speed: the median for an 8 s dictation, the same clips on both. Error rates: GPU on the full FLEURS test split (647 English, 862 German samples), CPU on 25 English and 21 German test clips up to 20 s; Parakeet TDT 0.6B. Pinned to 4 threads, an 8 s dictation takes 2.5 s on the CPU. Whisper Large V3 Turbo covers languages Parakeet does not.
The numbers, and where they come from
Every figure on this site, with the file it comes from and how it was measured. All CPU figures come from one PC, so read them as one small test, not as a promise for your machine. Read speech in a quiet room is easier than real dictation.
| Figure | What it is | Source in the repository | Test setup | Date |
|---|---|---|---|---|
| 0.17 s | to transcribe 10 s of English speech on a GPU | README.md, Performance table (scripts/benchmark.py; raw runs in benchmarks/parakeet-tdt-0.6b_en_us_20260406_062117.json and benchmarks/parakeet-tdt-0.6b_de_de_20260406_062522.json) | FLEURS test split (647 English / 862 German samples), Parakeet TDT 0.6B in the Docker container, one PC with a 16 GB NVIDIA GPU | 2026-04-06 |
| 5.8 % | word error rate, English, on a GPU | README.md, Performance table (scripts/benchmark.py; raw runs in benchmarks/parakeet-tdt-0.6b_en_us_20260406_062117.json and benchmarks/parakeet-tdt-0.6b_de_de_20260406_062522.json) | FLEURS test split (647 English / 862 German samples), Parakeet TDT 0.6B in the Docker container, one PC with a 16 GB NVIDIA GPU | 2026-04-06 |
| 5.2 % | word error rate, German, on a GPU | README.md, Performance table (scripts/benchmark.py; raw runs in benchmarks/parakeet-tdt-0.6b_en_us_20260406_062117.json and benchmarks/parakeet-tdt-0.6b_de_de_20260406_062522.json) | FLEURS test split (647 English / 862 German samples), Parakeet TDT 0.6B in the Docker container, one PC with a 16 GB NVIDIA GPU | 2026-04-06 |
| 0.08 s | for a 3 s dictation on a GPU (median) | docs/parakeet-cpu-test.md, Latency table (GPU container column) | Same FLEURS clips as the CPU column, sent to the Parakeet Docker container over HTTP; RTX 5070 Ti | 2026-10-31 |
| 0.08 s | for an 8 s dictation on a GPU (median) | docs/parakeet-cpu-test.md, Latency table (GPU container column) | Same FLEURS clips as the CPU column, sent to the Parakeet Docker container over HTTP; RTX 5070 Ti | 2026-10-31 |
| 0.18 s | for a 3 s dictation on a CPU (median) | docs/parakeet-cpu-test.md, Latency table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; Parakeet TDT 0.6B, ONNX int8, in-process, no Docker | 2026-10-31 |
| 0.34 s | for an 8 s dictation on a CPU (median) | docs/parakeet-cpu-test.md, Latency table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; Parakeet TDT 0.6B, ONNX int8, in-process, no Docker | 2026-10-31 |
| 0.51 s | for an 8 s dictation, pinned to 8 threads | docs/parakeet-cpu-test.md, Minimum hardware table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; process pinned to the first 8 logical processors with --threads 8 | 2026-10-31 |
| 2.5 s | for an 8 s dictation, pinned to 4 threads | docs/parakeet-cpu-test.md, Minimum hardware table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; process pinned to the first 4 logical processors (2 cores + SMT); a slower CPU will be slower still | 2026-10-31 |
| 100 s | to transcribe a 30-minute recording on a CPU | docs/parakeet-cpu-test.md, Real-time factor and RAM table (RTF 0.055) | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; one 30-minute test file through the voice-activity-detection path, no speaker detection | 2026-10-31 |
| 18× | faster than the recording is long, on a CPU Calculated: 30 min ÷ 100 s | docs/parakeet-cpu-test.md, Real-time factor and RAM table | Same run as the 100 s figure: 30 minutes of audio in 100 s | 2026-10-31 |
| 2.25 GB | of memory at the peak, for that 30-minute file | docs/parakeet-cpu-test.md, Real-time factor and RAM table | Peak while transcribing the 30-minute file; about 0.2 GB of it is the benchmark script’s own buffers | 2026-10-31 |
| 777 MB | of memory with the speech model loaded | docs/parakeet-cpu-test.md, Real-time factor and RAM table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; model loaded, no audio | 2026-10-31 |
| 0.067 | real-time factor for a 5-minute recording on 8 threads | docs/parakeet-cpu-test.md, Minimum hardware table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; 5-minute file, process pinned to 8 threads. A factor of 0.067 means 100 s of audio take 6.7 s | 2026-10-31 |
| 0.296 | real-time factor for a 5-minute recording on 4 threads | docs/parakeet-cpu-test.md, Minimum hardware table | AMD Ryzen 7 5800X, 8 cores / 16 threads, 47.9 GB RAM, Windows 11; 5-minute file, process pinned to 4 threads. A factor of 0.296 means 100 s of audio take 29.6 s | 2026-10-31 |
| 6.9 % | word error rate, English, on a CPU | docs/parakeet-cpu-test.md, Accuracy table | 25 English FLEURS test clips up to 20 s; Parakeet ONNX int8 on the CPU. Small sample: differences of a point or two are noise | 2026-10-31 |
| 6.8 % | word error rate, German, on a CPU | docs/parakeet-cpu-test.md, Accuracy table | 21 German FLEURS test clips up to 20 s; Parakeet ONNX int8 on the CPU. Small sample: differences of a point or two are noise | 2026-10-31 |
| 7.3 % | word error rate, German, short clips on a CPU | docs/parakeet-cpu-test.md, Long audio table | Every single German FLEURS clip (50 to 60, each up to about 20 s), raw model output; Parakeet ONNX int8 on the CPU | 2026-10-31 |
| 7.1 % | word error rate, German, items of about 57 s on a CPU | docs/parakeet-cpu-test.md, Long audio table | 12 German items of about 57 s made of consecutive FLEURS clips with 0.6 s of silence between them, raw model output; Parakeet ONNX int8 on the CPU | 2026-10-31 |
| 2.4 % | word error rate, English, 48 to 55 s of continuous speech on a CPU | docs/parakeet-cpu-test.md, English over 20 s on continuous speech | LibriSpeech test-clean, six chapters by six speakers, 18 items of 48 to 55 s, raw output, mean of the six chapter scores. Clean audiobook reading: microphone dictation with noise and hesitation will score worse | 2026-10-31 |
| 0.53 s | for Auto cleanup of a sentence on a GPU | docs/llm-cpu-test.md, second run, Speed table (Cleanup, 1st run, median) | Second run (112 items), 1st request of each item, median with the model loaded; a sentence of 10 to 25 words in, about 20 tokens out; Ollama defaults. Ryzen 7 5800X (8 cores / 16 threads) for the CPU row, RTX 5070 Ti 16 GB for the GPU row, Windows 11. Model: gemma4:e4b, 120 requests | 2026-11-01 |
| 2.9 s | for Auto cleanup of a sentence on a CPU | docs/llm-cpu-test.md, second run, Speed table (Cleanup, 1st run, median) | Second run (112 items), 1st request of each item, median with the model loaded; a sentence of 10 to 25 words in, about 20 tokens out; Ollama defaults. Ryzen 7 5800X (8 cores / 16 threads) for the CPU row, RTX 5070 Ti 16 GB for the GPU row, Windows 11. Model: gemma4:e2b, the default without an NVIDIA GPU, 120 requests; the larger gemma4:e4b takes 5.2 s on this CPU. A slower CPU will be slower | 2026-11-01 |
| 0.54 s | to rewrite a short selection on a GPU | docs/llm-cpu-test.md, second run, Speed table (Polish, 1st run, median) | Second run (112 items), 1st request of each item, median with the model loaded; a sentence of 10 to 25 words in, about 20 tokens out; Ollama defaults. Ryzen 7 5800X (8 cores / 16 threads) for the CPU row, RTX 5070 Ti 16 GB for the GPU row, Windows 11. Model: gemma4:e4b, 52 requests | 2026-11-01 |
| 2.0 s | to rewrite a short selection on a CPU | docs/llm-cpu-test.md, second run, Speed table (Polish, 1st run, median) | Second run (112 items), 1st request of each item, median with the model loaded; a sentence of 10 to 25 words in, about 20 tokens out; Ollama defaults. Ryzen 7 5800X (8 cores / 16 threads) for the CPU row, RTX 5070 Ti 16 GB for the GPU row, Windows 11. Model: gemma4:e2b, the default without an NVIDIA GPU, 52 requests; gemma4:e4b takes 3.6 s on this CPU. A slower CPU will be slower | 2026-11-01 |
Pro
NexusVox Pro is a separate paid add-on, available from December. Everything on this page stays free in the core app.