Performance

What performance means in a browser AI transcription tool

Performance is not only speed. In a browser setting it also includes startup cost, memory headroom, long-run stability, and whether the session survives a realistic queue.

Main factors

Model size

Larger models demand more memory and can amplify instability on constrained devices.

Quantization

Smaller representations often improve feasibility and completion rate in the browser.

Backend choice

WebGPU can accelerate throughput, but WASM may still be the safer path on some machines.

Job length

Short demos and multi-hour media expose very different runtime behavior.

Performance guidance

If the browser feels fragile, step down in model size before blaming the concept of local AI itself. A smaller model that completes cleanly is usually more valuable than a larger one that spends its time reallocating memory or failing halfway through a long job.

For folders and long recordings, queue pacing and cleanup matter as much as raw throughput. That is why batch systems need disciplined resource handling rather than only faster kernels.

Compare models on the same workload

Separate download time from transcription time. Model downloads depend on network speed and cache state; inference time depends on the device and runtime. Combining them into one number makes the result hard to reproduce.

Choose a representative 60–120 second clip with clear speech and a known transcript.
Record device, operating system, browser version, model and runtime backend.
Load the model once and exclude download and warm-up from the measurement.
Run the same clip three times, recording completion time and any fallback or browser error.
Compare the words against the reference transcript, checking names and numbers separately from speed.

Keep results attached to the tested device. They describe that setup; they do not predict performance on every computer.

When a long job stalls

Confirm that model loading completed, then try a short clip with the same model. If the short clip works and the long one fails, reduce model size, close memory-heavy tabs and split the recording. If both fail, retry the short clip after changing the backend when that option is available.

For a useful bug report, include browser and version, operating system, device memory, GPU model if known, model name, runtime status, recording duration and the last status message. Avoid attaching sensitive recordings; use a short sample you are allowed to share.