Efficient model included
EYRO’s Efficient English model is designed for usable local dictation on everyday CPU-only and integrated-graphics laptops.
Performance
Powerful hardware should feel powerful
EYRO chooses a speech model for recognition quality and a backend for execution speed. Supported GPUs and NPUs may sharply reduce the wait after speaking, while CPU continuity remains available. A fallback changes the compute path—not the selected model’s intended recognition target.
Performance proof
Less waiting after you speak.
On a reference i9 + RTX 5070 Ti laptop using EYRO’s Medium speech model and the same test text, GPU-integrated paths averaged 2.27 seconds versus 6.84 seconds without GPU acceleration, about 3.02× faster processing in this internal test.
Non-GPU reference6.84 s
GPU accelerated2.27 s
Results are from an internal founder benchmark. Performance varies by hardware, model, audio length, backend, power mode, and warm/cold state.
Calculated effect
3.02×
faster processing in this internal test
EYRO Medium model, same text, i9 + RTX 5070 Ti, two runs per path
Guidance
Choose the quality level. Let EYRO find the reliable compute path.
Models are explicit. Backend calibration, warm processing, and recovery happen underneath them.
Recommended model
EYRO’s Recommended English model is a verified optional download offering a stronger balance of recognition quality, storage, memory, and speed.
Maximum model
EYRO’s full-precision Maximum model remains available for powerful PCs when the highest recognition tier matters most.
Warm workers
Keeping the engine warm between recordings reduces repeated cold-start cost on supported setups.
Hardware-aware calibration
Setup compares reliable local runs and may choose CPU, broad Vulkan acceleration, optional NVIDIA CUDA, or compatible Intel NPU acceleration.
Continuity fallback
A failed accelerator does not discard the recording. EYRO continues on a verified fallback and tests the faster path again after a cooldown.
Method
Publishable methodology
Claims stay adjacent to method. The internal reference result used on this site follows these rules.
- Fix audio file, model hash, EYRO version, runtime, language, and decoding settings
- Record OS, CPU, GPU, RAM, power mode, drivers, and plugged-in state
- Separate cold first run from warm steady-state runs
- Prefer median and P95 over fastest-run marketing when sample size allows
- Compare transcript identity / WER, not speed alone, when changing backends
Results are from an internal founder benchmark. Performance varies by hardware, model, audio length, backend, power mode, and warm/cold state.