header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Superwhisper Open-Sources 600M-Parameter Small Model, Specifically Designed for Speech Transcription

Perceive-and-Beat AI News Flash: AI voice input tool Superwhisper has released its first open-source model, S1-mini. With only about 600 million parameters, it is based on Qwen3-0.6B fine-tuning and is specifically designed to clean up drafts generated by speech recognition, capable of running entirely on local devices.

S1-mini itself does not handle audio listening. Models such as Whisper and Parakeet first convert sound into text before handing it over to S1-mini. It removes vocal fillers like "um," corrects self-corrections, adds punctuation and capitalization, and can normalize dictated numbers, dates, times, amounts, and emails into standard written formats.

The model currently only supports English. The official quantized version is around 462MiB and can run on a regular laptop CPU. It also provides GGUF, which can be integrated with llama.cpp, Ollama, and LM Studio. The official testing on 7519 pieces of English data achieved a 94.8% token accuracy rate; however, no cross-model comparison scores were provided.

Actually, S1-mini entered the Superwhisper App for experimental testing as early as June, with the official claiming that most devices could achieve processing speeds of around 200 tokens/s. The weights are now publicly available, allowing other speech input, meeting transcription, and subtitling applications to be directly deployed locally.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish