GigaAM for macOS

About

GigaAM runs GigaAM v3, the Russian speech recognition model by Sber, on your Mac. The add-on uses the end-to-end version of the model, which writes punctuation and capital letters itself.

This is a local engine. Your audio is transcribed on your Mac. The add-on uses the network to download the model.

GigaAM recognizes only Russian. For other languages, choose another engine.

Requirements

This release requires TypeWhisper 1.8.0 or newer on macOS 14 or later and a Mac with Apple silicon. It is available for the direct download of TypeWhisper; the Mac App Store edition does not include it.

The model is a download of about 230 MB. The add-on states 2 GB of memory or more for it. You need no API key and no account.

Setup

1. In TypeWhisper, open Settings > Discover plugins and install GigaAM. The add-on then appears in the Settings sidebar, and its page opens.
2. Select Download & Load beside GigaAM v3 (Russian, with punctuation).
3. Select GigaAM (Russian) as your transcription engine in the dictation settings.
4. Set the language of the dictation to Russian or leave it on automatic detection.

Settings

Setting	What it controls
HuggingFace Token	Optional. A token raises the download rate limits of Hugging Face. Save checks the token with Hugging Face and stores a valid one in the macOS Keychain. Remove deletes the stored token.
Model	Shows GigaAM v3 with its size and memory requirement. Download & Load downloads the model and loads it. Unload releases the model from memory and deletes the downloaded files.

Transcription

The add-on transcribes the finished recording. It does not stream partial text while you speak. It splits a recording longer than 20 seconds into sections at a pause nearby and transcribes them one after the other.

The model writes punctuation, capital letters and numbers as digits. Every recording is transcribed as Russian, also when the dictation is set to another language. Translation is not supported, and TypeWhisper dictionary terms are not passed to the model.

Accuracy

In the TypeWhisper benchmark on 300 Russian recordings, GigaAM v3 made fewer errors than the other engines TypeWhisper offers for Russian:

Engine	Word error rate
GigaAM v3 with punctuation (sherpa-onnx export)	4.3%
OpenAI gpt-4o-transcribe (cloud)	4.8%
Whisper large-v3	5.2%
Whisper large-v3 turbo	5.9%
Qwen3-ASR 1.7B	8.9%
Parakeet TDT 0.6B v3	10.1%

The benchmark ran GigaAM through the sherpa-onnx export of the model. The add-on’s model files reached 4.0% when run with the add-on’s runtime on the same 300 recordings. The recordings come from FLEURS, Russian LibriSpeech and Common Voice. The benchmark repository describes the test sets and the method.

Local processing and privacy

Recognition runs on your Mac. The add-on sends no audio and no text to a cloud service.

The model files come from the Hugging Face repository istupakov/gigaam-v3-onnx: the int8 versions of the encoder, decoder and joint network and the vocabulary. TypeWhisper stores them in ~/Library/Application Support/TypeWhisper/PluginData/com.typewhisper.gigaam/models. The add-on also contacts Hugging Face when it checks a token.

GigaAM v3 is published by Sber under the MIT license. The ONNX versions of the model and the feature extractor that ships with the add-on come from the onnx-asr project by Ilya Stupakov, also under the MIT license.