Numbers

Sources: docs/talk-spike.md (“Speech in through CrispASR”), etalk-engine (MIN_GPU_BYTES, MAX_WARMUP_RTF), Licences (download sizes).

Speech in

Silero VAD and Parakeet TDT 0.6B v3 (Q8_0), on an RTX 4090 and an i9. Ten synthetic English clips, run through the worker’s own pipeline. Word error rate counts against the spoken form, so “New 54” for “NEW fifty four” is an error.

	GPU (CUDA)	CPU, 8 threads
Text after you let go	10–41 ms	102–450 ms
Word error rate	14.9 %	12.6 %
Words while you speak	every 300 ms	no
Model load	0.72 s	0.48 s
VRAM	1,246 MiB	none

Words while you speak need 1.6 GB of VRAM free. Our German and English built-in clips came through word for word on both. No human recordings yet.

Speech out

Qwen3-TTS 0.6B Base (F16) and its codec, on the GPU only.

VRAM, peak	3.7 GB
GPU needed	4 GB or more
Real-time factor, RTX 4090	0.13–0.27
Real-time factor, CPU	~2.0, so no speech on the CPU

A real-time factor below 1 is faster than real time. A GPU must reach 0.6 in a warm-up sentence. Time to first audio is not measured on this model yet; the earlier CustomVoice Q8_0 model took 0.3–0.5 s for a short sentence.

Downloads

	Linux	macOS	Windows
Speech out, first use	~2.97 GB	~2.19 GB	~2.28 GB
Speech in models	675 MB	675 MB	675 MB
CPU-only runtime (no GPU driver)	22.7 MB	—	58.4 MB

Linux: runtime 196 MB, cuBLAS 581 MB, model 1.83 GB, codec 358 MB. macOS: the runtime ships in the app, so only the models download.

Targets

What we aim for; these are targets, the Today column is measured.

	Target	Today
First words while you speak	< 300 ms	not measured
Final text	< 500 ms	10–41 ms GPU, 102–450 ms CPU
First spoken audio	< 400 ms at p95	not measured on this model
Local models, VRAM	≤ 4 GB	3.7 GB speech out + 1,246 MiB speech in: over, together