How to Choose Local Speech-to-Text for an Older PC

On an older PC, start with a smaller speech recognition model that supports your language. Test it with plain transcription in the editor you use, then try a larger model only if the smaller one keeps missing words you need. Choose using the time it takes you to get usable text, including corrections.

Whisper offers models in several sizes and English-only and multilingual variants. Check the model label before downloading, especially if you switch languages. Some implementations, such as whisper.cpp, support quantized models with different resource requirements. The reference model table and a model's download size describe different things from how the desktop app will perform on your computer. Whisper model sizes and languages · whisper.cpp model quantization

Measure the four sentences you actually need

Pick a note or plain editor, a microphone, and one speech model. Turn off optional AI rewriting during the comparison so that corrections and timing reflect the speech step. After one warm-up, use the same four sentences on each candidate.

Record the model, editor, time to usable text, and number of meaning-changing corrections. Try each model at least twice and keep the microphone, room, and target app the same. If you'll dictate in two languages, include a mixed sentence. Compare behavior with the browser or editor you normally leave open.

Choose using the result

• Small model is quick and keeps the important details. Keep it for messages and notes.
• Small model misses recurring terms. Try one larger model that supports the same language. Keep the larger one if the reduction in correction work makes the added wait acceptable.
• Both models miss the same words. Check the selected microphone, speak the same line in a quieter place, and run the test again before downloading another model.
• Plain text is quick but polished output takes longer. Measure cleanup as its own step. For short notes, plain transcription may already be enough.

Write down a repeatable rule for your work. For example, you might favor the smallest model that consistently preserves your ticket numbers in short messages. That criterion is more useful than a generic recommendation to buy hardware or use the largest model.

Voconly offers multiple local speech models and optional text processing. Use its model list to find an available candidate for your language and computer. If you compare devices or model families, record the app version and model variant so you can repeat the test later. Voconly model selection guide · Local ASR comparison