Performance¶
This page reports measured speech-transcription performance of the bundled Windows build, to help you pick a backend and model size for your recordings.
All numbers were measured on 2026-09-02 on a clean Windows 11 Pro machine
(Azure Standard_B4s_v2: 4 vCPU, 16 GB RAM, Microsoft Defender active),
CPU only, using a 64-second spoken-audio clip. Your hardware will differ,
but the relative differences between backends and sizes carry over.
Transcription speed¶
Time to transcribe the 64-second clip with warm caches (model weights already downloaded and loaded once before):
| Model | Faster Whisper | OpenAI Whisper | WhisperX |
|---|---|---|---|
| tiny | 3.5 s | 8.2 s | 3.9 s |
| base | 6.6 s | 9.9 s | 6.0 s |
| small | 19.2 s | 23.1 s | 15.3 s |
| medium | 56.5 s | 63.3 s | 42.0 s |
| large-v2 | 97.9 s | 114.4 s | 73.4 s |
| large-v3 | 97.0 s | 114.9 s | 72.1 s |
| large-v3-turbo | 56.4 s | 200.2 s * | 51.1 s |
Avoid OpenAI Whisper with large-v3-turbo on CPU
In our test this combination fell into repeated temperature-fallback decoding and produced hallucinated extra text (152 words transcribed against ~134 actually spoken) while also being the slowest run in the entire matrix. Use Faster Whisper or WhisperX if you want the turbo model on a machine without a GPU.
How to read the table:
- The default small model runs at roughly 3-4x realtime on every backend — an hour of audio takes 15-20 minutes.
- WhisperX is the fastest backend from medium upwards thanks to batched inference, and keeps even large-v3 close to realtime.
- OpenAI Whisper is consistently the slowest — it computes in fp32 on CPU and may re-decode segments with temperature fallbacks, so its times also vary more between runs.
Model loading¶
Each processing run first loads the selected model:
| Backend | Typical warm load time |
|---|---|
| Faster Whisper | 1-10 s (size-dependent) |
| OpenAI Whisper | 2-20 s (size-dependent) |
| WhisperX | 9-20 s (also initialises a voice-activity pipeline) |
First use of a model (cold start)¶
The first time you select a model size, its weights are downloaded and caches are built. On a fast connection this takes:
| Model size | First-use overhead |
|---|---|
| tiny / base | a few seconds |
| small | ~10 s |
| medium | 30-60 s |
| large variants | 1-2.5 min (~3 GB download each) |
WhisperX additionally performs a one-off pipeline initialisation of about 100 s on its very first run.
The app is not stuck
During a cold start the progress display can sit still while weights download in the background — this is the "first use may download model weights" stage, not a hang. On slow connections the large models can take considerably longer than the times above.
Requirements on Windows¶
On a freshly installed Windows the app requires the Microsoft Visual C++ Redistributable (x64); without it the detection and speech backends fail to start. Installers built after 2026-09-01 detect and install it automatically. For older builds, install it manually from aka.ms/vs/17/release/vc_redist.x64.exe.
Methodology¶
The benchmark harness lives in the repository at
scripts/azure-wintest-bench.ps1.
It generates a ~64 s spoken clip with Windows text-to-speech (16 kHz mono),
then for each backend and model size runs one cold pass (fresh process,
weight download, first Defender scan) and one measured warm pass with word
timestamps enabled. Word counts across all runs were compared against the
known text to confirm real transcription took place.