superwhisper/s1-mini

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:otherArchitecture:Transformer0.3K Featherless Exclusive Cold

S1-mini by Superwhisper is a 0.6 billion parameter text normalizer fine-tuned from Qwen3-0.6B, designed to transform raw Automatic Speech Recognition (ASR) transcripts into clean, written English text. It removes fillers, resolves false starts, applies punctuation and capitalization, and formats numbers, dates, and emails. Achieving 94.8% token accuracy on a held-out English test set, this model is optimized for on-device deployment and runs efficiently on laptop CPUs.

Loading preview...

Model Overview

S1-mini, developed by Superwhisper, is a specialized 0.6 billion parameter causal language model, fine-tuned from Qwen3-0.6B. Unlike general-purpose LLMs, S1-mini performs a single, critical task: normalizing raw ASR transcripts into polished, readable text. It effectively removes disfluencies like fillers and false starts, corrects self-corrections, and applies proper capitalization and punctuation. Additionally, it intelligently formats spoken numbers, dates, times, currency, and email addresses into their written forms.

Key Capabilities

  • Text Normalization: Transforms raw, often unpunctuated and lowercase ASR output into clean, grammatically correct text.
  • Configurable Output: Users can control the output's styling (casual, semi-casual, semi-formal, formal), structure (prose, lists), and context (general, email) via a control line in the input.
  • High Accuracy: Achieves 94.8% token accuracy on a dedicated English test set of 7,519 cases.
  • Lightweight & Efficient: With 0.6B parameters, it's designed to run comfortably on laptop CPUs, making it suitable for on-device deployment.
  • English-only: Currently supports English exclusively.

Use Cases

S1-mini is ideal for integration into applications that process speech-to-text output, such as:

  • Dictation apps
  • Meeting-notes tools
  • Live captioning services
  • Voice-driven editors

It serves as a crucial post-processing step, taking raw ASR output and delivering text ready for display or further processing. The model requires a specific system prompt and control line for optimal performance, and it's recommended to decode greedily and chunk longer transcripts for best results.

Finetunes
2 models