Talkhawk
Parakeet v3 on a desktop CPU: speech to text at 20× real time
Measured October 29, 2026 with Talkhawk 0.2.0.
Talkhawk, a dictation app for Windows, runs NVIDIA's open Parakeet speech model on the processor. On a Ryzen 7 7800X3D desktop it turns 21 seconds of speech into text in 1.03 seconds, with the GPU idle. Here are the timings, how they were taken, and how the app uses that speed to have the text ready when you let go of the shortcut.
| Speech | Time to text | Faster than real time | Peak memory |
|---|---|---|---|
| 5.2 s | 0.21 s | 25× | 743 MB |
| 21.0 s | 1.03 s | 20× | 846 MB |
| 51.7 s | 2.49 s | 21× | 891 MB |
| 4 min 20 s | 12.0 s | 22× | 942 MB |
How it was measured
- PC: AMD Ryzen 7 7800X3D, 8 cores and 16 threads, on Windows 11. It also has an RTX 5070 Ti, which Talkhawk doesn't use.
- Model: parakeet-tdt-0.6b-v3 in its int8 ONNX export, run by the open source transcribe-rs crate on ONNX Runtime's CPU execution provider.
- Program: the talkhawk.exe from the 0.2.0 Store package. Its
--dictateflag loads the model, reads a 16 kHz mono WAV file, transcribes it and prints both times. - Audio: four clips made with the speech synthesizer built into Windows (System.Speech, default voice) at 16 kHz mono. Their texts are in dictation-on-a-cpu.txt.
- Each clip ran three times, each in a fresh process. The three runs were within 3% of each other. Memory is the process's peak working set, read every 50 ms.
What these numbers don't show
Synthetic speech is clean and evenly paced, so these clips measure speed, not accuracy. For accuracy, NVIDIA's model card reports a 6.34% average word error rate for English on the Hugging Face Open ASR Leaderboard. Across the 25 languages of the FLEURS test set the average is 11.97%, from 3.00% for Italian to 24.03% for Slovenian, scored without punctuation.
There is no laptop number yet. A laptop processor with fewer cores and lower clocks will be slower, by an amount nobody has measured on Talkhawk so far.
How the app uses the speed
Dictation needs the text right after you stop talking, so speed on one long file isn't enough. Talkhawk transcribes while you're still speaking.
While you hold Ctrl + Win, the recording is cut into phrases at pauses. A phrase ends at a pause of 550 ms once it is at least 3 seconds long and holds a second of speech. If you never pause, it ends at its quietest moment before 25 seconds. Each finished phrase goes to the model on a background thread, and its words appear in the on-screen indicator while you keep talking.
Pause detection goes by loudness alone. It follows the room's noise floor, which drops at once to any quieter moment but rises only about 1 dB a second, so a long sentence never becomes the floor.
When you let go, what's left depends on how long you talked:
- Up to 20 seconds, the whole clip goes through the model again in one pass, because one window gives the best punctuation. On this CPU that takes about a second at most.
- Longer than that, the phrases already transcribed stay, and only the audio after the last pause is transcribed. That is at most 25 seconds of audio, about 1.2 seconds at the speed above, and usually a few seconds of audio, about 0.2 seconds.
What stays on the PC
Recording and speech recognition never leave the PC, and dictation works offline. The free plan stops there. Talkhawk Pro adds Polish, which removes filler words and fixes punctuation by sending the transcribed text, never the audio, to a cloud language model. Turn Polish off and your dictations stay on your PC.
Built on
NVIDIA's Parakeet TDT 0.6B v3 (CC BY 4.0), transcribe-rs, ONNX Runtime and Tauri. Talkhawk itself is closed source. It's free in the Microsoft Store for Windows 11, with no word limit.