---
title: "Local Speech to Text with Whisper | Micdrop"
description: "Which Whisper checkpoint to run locally, and why a checkpoint tied to one language beats a heavier generic one."
url: "https://micdrop.dev/docs/local-models/speech-to-text"
---

# Local Speech to Text with Whisper

Whisper is the local option, running inside the Node process through [@micdrop/whisper](/docs/speech-to-text/whisper). Choose the checkpoint mainly by the language of the call, not by its size.

Checkpoint

Languages

Download

On a 3 second sentence

Word errors in French

`tiny`

Around a hundred

~45 MB

~320 ms

Not measured

`base`

Around a hundred

~85 MB

~440 ms

25%

`small`

Around a hundred

~250 MB

~1000 ms

22%

`turbo`

Around a hundred

~850 MB

~3800 ms

17%

`french`

French

~390 MB

~1100 ms

2%

Word errors were measured on five French sentences with proper nouns, numbers and homophones. Latencies were measured on a MacBook Pro M2, on CPU, with the model loaded.

## Choosing the checkpoint

Whisper always processes a 30 second window, whatever the length of the sentence, so a larger checkpoint makes every sentence slower, short ones included. A larger checkpoint is more accurate in English. For other languages, a checkpoint fine-tuned on that language is both more accurate and faster than a larger generic one.

The `french` checkpoint is as fast as `small`, and more accurate in French than `turbo`, which is twice as large and 3 times slower. It also writes numbers in words instead of digits, which the voice needs. The generic `base` checkpoint gets a quarter of the words of a French sentence wrong.

The [Whisper page](/docs/speech-to-text/whisper) gives the repository behind each shorthand, and how to pick a checkpoint fine-tuned on another language.

## Naming the language

Set `language` when you know the language of the call. It skips language detection and avoids a wrong guess on a short sentence.

```typescript
new WhisperSTT({
  model: 'french',
  language: 'fr',
})
```

## Where it runs

On a Mac, transcription runs on the CPU. Transformers.js has no GPU device there, and the CoreML provider was slower than the CPU on these models. The [latency and memory](/docs/local-models/performance) page has the measurements and the options for Windows and Linux.

## What was measured and set aside

[Voxtral Mini 3B](/docs/local-models/explorations/voxtral-stt) transcribes more accurately than Whisper `base`, but is 35 times slower per utterance.

## Alternatives to Whisper

Vosk, Moonshine, NVIDIA Parakeet and Kyutai STT skip Whisper’s 30 second window and are still to be measured on this machine. You can [run them from Node](/blog/local-speech-to-text-node) through a native addon, the `vosk` npm package or a separate server. Qwen3-ASR, Cohere Transcribe and NVIDIA’s Nemotron models join them in a [ranking of open source speech to text models](/blog/top-open-source-speech-to-text-models) for voice agents.
